UNIVERSIDAD POLITÉCNICA DE MADRID

ESCUELA TÉCNICA SUPERIOR DE

INGENIEROS INDUSTRIALES

Computational Modeling of Powder Bed

Fusion Manufacturing of Metals

TESIS DOCTORAL

SEYED MOHAMMAD ELAHI

Master Ingeniero Mecánico

2022
 Departamento de Ingenierı́a Mecánica
Escuela Técnica Superior de Ingenieros Industriales

Computational Modeling of Powder Bed

Fusion Manufacturing of Metals

SEYED MOHAMMAD ELAHI

Master Ingeniero Mecánico

Directores de Tesis

Ignacio Romero Olleros

Doctor Ingeniero Industrial

Damien Tourret
Doctor en Ciencia e Ingenieria de Materiales

2022
Seek knowledge from the cradle to the grave.
l-Qasim Ferdowsi Tusi
 Acknowledgements
Foremost, I would like to express my sincere gratitude to my advisors Prof.
Ignacio Romero and Dr. Damien Tourret for the continuous support of my Ph.D.
study and research, and their patience, guidance, and enthusiasm. Their invalu-
able patience and feedback helped me in all the time of research and writing of this
thesis. Besides my advisors, I would like to thank the rest of my thesis commit-
tee for their encouragement and insightful comments. Additionally, this endeavor
would not have been possible without the generous support from Spanish Min-
istry of Science, which financed my research under the Retos-Colaboracion project
ENVIDIA (Ref. RTC-2017-6150-4).

I am also grateful to all of my groupmates in Computational Solid Mechanics

group and Modeling and Simulation of Materials Processing group at IMDEA
Materials institute for their feedback sessions and moral support.

Last, but not least, my warm and heartfelt thanks go to my family, especially
my parents, for the tremendous support and hope they had given to me. Without
that hope, this thesis would not have been possible. Thank you all for the strength
you gave me.

Seyed Mohammad Elahi

28th , October, 2022
 Abstract
Fusion-based additive manufacturing of metal is particularly challenging to
simulate due to the broad range of length/time scales of multiphysics phenom-
ena to bridge. Integrated Computational Materials Engineering (ICME) offers a
promising route to accelerate the design of new materials, optimize manufacturing
processes, and thus improve material properties and performance. ICME heav-
ily relies on linking multiple modelling techniques relevant to distinct length/time
scales and/or different physics. While a wide range of physics-based and predictive
models exist, the efficient linking of these models remains challenging. In the con-
text of additive manufacturing, interdependent phenomena across scales include:
multicomponent alloys with complex phase diagrams, multiple phase transforma-
tions and temperature-dependent properties; complex thermal exchanges in the
vicinity of the melt pool; and the development of nonequilibrium grain microstruc-
tures within this melt pool.

A multiscale modeling strategy (i.e. micro and macro) has been developed to
simulate powder-bed fusion of metallic alloys, which combines: (1) Temperature-
dependent alloy properties and phase diagrams calculated by CalPhaD, (2) macroscale
thermal and thermo-mechanical simulations of the material addition and fusion
using finite elements (FE), (3) Phase-field (PF) simulations of the melt pool’s
solidification at the microscopic scale, and (4) Simulations of the microscopic so-
lidification of the melt pool using cellular automaton (CA). Realistic processing
parameters are used in the methodology to simulate selective laser melting (SLM)
of an Inconel 718 alloy.

In the first step, three-dimensional thermal and thermo-mechanical frameworks

were developed – using an in-house finite element code (IRIS) and a material
library (MUESLI) implemented in C++ programming language – by considering
these four factors: (1) Representing the thermal input from the scanning laser, (2)
A material model to specify the temperature-dependent material property in the
cyclic heating and cooling environment, (3) A physics-based method to explain the
layer build-up process, and (4) A laser material interaction model to account for
transient thermo-mechanical phenomena.

Then, the thermal model was coupled with CalPhaD and PF methods to predict
the formation of microstructures at the scale of the entire melt pool. In this part,
the effect of temperature-dependent properties and the necessity of taking into
account differences in properties between the powder bed and the dense material
have been discussed. We perform an appropriately-converged PF solidification
simulation at the scale of the entire melt pool using a two-dimensional longitudinal
slice of the thermal field calculated through FE simulations. This calculation has
over one billion grid points, yet is performed on a single cluster node with eight
graphics processing units (GPUs). These microscale simulations, with a level of
detail down to individual dendrites, offer new insight into the selection of grain
texture via polycrystalline growth competition under realistic SLM conditions.

Finally, comparisons between phase-field (PF) and cellular automaton (CA)

simulations of polycrystalline growth in a two-dimensional melt pool under addi-
tive manufacturing (powder-bed fusion) relevant conditions have been performed.
The generated grain structures from local (point-by-point) measurements and grain
orientation distributions that were averaged over various simulations have been
compared. We investigated how the melt pool aspect ratio and the CA spatial dis-
cretization level affected the selected grain texture. Our simulations demonstrate
that only PF simulations are capable of capturing the fine-grained microscopic
features related to transient growth conditions and solid-liquid interface stability,
such as the initial planar growth stage before its cellular/dendritic destabiliza-
tion or the early elimination of unfavorably oriented grains due to neighbor grain
sidebranching. CA grid refinement can only partially address the resulting dis-
agreement between PF and CA predictions. In addition, with some variability on
the CA grid and melt pool shape, overall grain distributions averaged throughout
the entire melt pools of several simulations appear to lead to a noticeably better
agreement between PF and CA. This research offers a helpful step in that direction
by quantitatively comparing both approaches at process-relevant length and time
scales, even though more effort is still needed, especially to identify the appropriate
selection of CA spatial discretization and its link to characteristic microstructural
length scales.
Contents

Acknowledgements III

Abstract V

List of Figures XI

List of Tables XIII

1 Introduction 1

1.1 Additive manufacturing (AM) . . . . . . . . . . . . . . . . . . . . . 1

1.2 Selective laser melting (SLM) . . . . . . . . . . . . . . . . . . . . . 4

1.2.1 Finite element modeling of SLM . . . . . . . . . . . . . . . . 6

1.2.2 Modeling of multi-layer parts . . . . . . . . . . . . . . . . . 6

1.3 Modeling categories . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

1.3.1 Temperature field . . . . . . . . . . . . . . . . . . . . . . . . 8

1.3.2 Residual stresses and distortion . . . . . . . . . . . . . . . . 12

1.3.3 Melt pool characteristics . . . . . . . . . . . . . . . . . . . . 15

1.4 Multiscale simulation of SLM . . . . . . . . . . . . . . . . . . . . . 20

I
Contents

2 Thermal Finite Element Framework 25

2.1 Modeling of laser thermal input . . . . . . . . . . . . . . . . . . . . 27

2.1.1 Laser scan path . . . . . . . . . . . . . . . . . . . . . . . . . 29

2.2 Temperature-dependent alloy properties . . . . . . . . . . . . . . . 30

2.2.1 Properties of the powder bed . . . . . . . . . . . . . . . . . 31

2.3 Boundary conditions . . . . . . . . . . . . . . . . . . . . . . . . . . 32

2.4 Modeling of layer build-up and state change . . . . . . . . . . . . . 34

2.5 Thermal solver . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

2.6 Verification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

2.7 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40

3 Thermo-Mechanical Finite Element Framework 45

3.1 Finite element method for fully coupled thermomechanical analysis 46

3.1.1 Model description . . . . . . . . . . . . . . . . . . . . . . . . 46

3.1.2 The first law of thermodynamics . . . . . . . . . . . . . . . 47

3.1.3 Materials with internal variables . . . . . . . . . . . . . . . . 47

3.2 Coupled thermomechanics . . . . . . . . . . . . . . . . . . . . . . . 49

3.2.1 Small strain thermomechanics . . . . . . . . . . . . . . . . . 49

3.2.2 Small strain thermomechanical materials . . . . . . . . . . . 50

3.2.3 Thermoelasticity . . . . . . . . . . . . . . . . . . . . . . . . 51

3.2.4 General thermomechanical materials . . . . . . . . . . . . . 51

3.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54

4 Multiscale Modeling of SLM 55

4.1 Computational thermodynamics of multicomponent alloy . . . . . . 56

4.1.1 Temperature-dependent alloy properties . . . . . . . . . . . 56

II
 Contents

4.1.2 Phase diagram and pseudo-binary approximation . . . . . . 57

4.2 Macroscopic thermal simulations . . . . . . . . . . . . . . . . . . . 59

4.2.1 Thermal problem . . . . . . . . . . . . . . . . . . . . . . . . 59

4.2.2 Numerical implementation . . . . . . . . . . . . . . . . . . . 60

4.3 Macroscale to microscale coupling . . . . . . . . . . . . . . . . . . . 61

4.4 Microscale modeling of microstructure growth in the melt pool . . . 63

4.4.1 Phase-field model . . . . . . . . . . . . . . . . . . . . . . . . 63

4.4.2 Polycrystalline solidification . . . . . . . . . . . . . . . . . . 65

4.4.3 Numerical implementation . . . . . . . . . . . . . . . . . . . 66

4.5 Application to selective laser melting of inconel 718 alloy . . . . . . 68

4.5.1 Processing conditions . . . . . . . . . . . . . . . . . . . . . . 68

4.5.2 Thermal simulations and parameters (FE) . . . . . . . . . . 68

4.5.3 Microstructure simulations and parameters (PF) . . . . . . . 70

4.6 Results and discussion . . . . . . . . . . . . . . . . . . . . . . . . . 71

4.6.1 Temperature-dependent thermophysical properties . . . . . . 71

4.6.2 Macroscopic thermal field . . . . . . . . . . . . . . . . . . . 72

4.6.3 Microstructure growth in the melt pool . . . . . . . . . . . . 76

4.7 Conclusion and perspectives . . . . . . . . . . . . . . . . . . . . . . 81

5 Comparing Phase-field and Cellular Automaton Models for Simulation of

Microstructural Evolution During Melt Pool Solidification 85

5.1 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85

5.1.1 Thermal field . . . . . . . . . . . . . . . . . . . . . . . . . . 86

5.1.2 Cellular automaton . . . . . . . . . . . . . . . . . . . . . . . 91

5.1.3 Post-processing . . . . . . . . . . . . . . . . . . . . . . . . . 94

III
Contents

5.2 Results and discussion . . . . . . . . . . . . . . . . . . . . . . . . . 98

5.2.1 Longitudinal section . . . . . . . . . . . . . . . . . . . . . . 98

5.2.2 Cross-sections . . . . . . . . . . . . . . . . . . . . . . . . . . 104

5.3 Conclusions and perspectives . . . . . . . . . . . . . . . . . . . . . . 110

6 Summary and Perspectives 117

6.1 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117

6.2 Perspectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118

Bibliography 119

IV
List of Figures

1.1 The classification of metallic powder additive manufacturing pro-

cesses [1]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

1.2 Concept of SLM process. (i) High-power laser melts selective areas
of the powder bed. (ii) Process is repeated for successive layers. (iii)
Loose powder removed and finished part revealed [3]. . . . . . . . . 5

1.3 (a) 3D finite element model and (b) Gaussian laser energy density [6]. 7

1.4 Temperature profiles of consecutive spots [8]. . . . . . . . . . . . . . 9

1.5 Heat flux magnitude for four different laser power [7]. . . . . . . . . 10

1.6 Temperature field at different times during 3 passes with (#A) uni-
directional scanning and (#B) alternate scanning [11]. . . . . . . . 11

1.7 Temperature distribution at the end of the construction process

after dwell time [5]. . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

1.8 Four scanning strategies: (a) horizontal, (b) vertical, (c) successive,
(d) LHI [28]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

1.9 The residual stress distributions when the powder bed was cooled
down at 100 s: (a) X-component of stress along the scan direction;
(b) Y-component of stress transversal to the scan direction; (c) Z-
component of stress through the thickness of the layer; (d) Von
Mises stress [31]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

V
LIST OF FIGURES

1.10 The residual stress (1) layer 1; (2) layer 2; (3) layer 3. a, d, g
Residual stress along x-direction. b, e, h Residual stress along y-
direction. c, f, i Residual stress along z-direction [23]. . . . . . . . . 16

1.11 Displacement of the melt pool center with respect to the laser beam
[1]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

1.12 Melt pool length evolution for different process parameters [37]. . . 18

1.13 The side view of the propagation of the melt pool [40]. . . . . . . . 19

1.14 Schematic of scan paths: (a) Z-shaped and (b) S-shaped[44]. . . . . 21

1.15 Evolution of the predicted microstructure in different time steps[53]. 22

1.16 3D view of (a) analytical and (b) CA prediction of a dendritic grain

envelope. In (1) the temperature is uniform (G = 0 K/m) whereas
in (2) G = 250 K/m [67]. . . . . . . . . . . . . . . . . . . . . . . . . 23

2.1 Selective laser melting process. . . . . . . . . . . . . . . . . . . . . . 26

2.2 Laser beam irradiance with a Gaussian distribution [34]. . . . . . . 28

2.3 Different laser scanning paths. . . . . . . . . . . . . . . . . . . . . . 28

2.4 Different scan strategies. . . . . . . . . . . . . . . . . . . . . . . . . 29

2.5 Using G-code file to define the laser path. . . . . . . . . . . . . . . 30

2.6 Heat transfer during SLM additive manufacturing. . . . . . . . . . . 33

2.7 Schematic of layer build-up process. . . . . . . . . . . . . . . . . . . 36

2.8 Comparing the result of a) IRIS and b) [74]. . . . . . . . . . . . . . 39

2.9 Comparing the result of a) IRIS and b) [93]. . . . . . . . . . . . . . 40

2.10 Temperature field. . . . . . . . . . . . . . . . . . . . . . . . . . . . 41

2.11 Three different states containing: Powder, Fluid, and Solid. . . . . . 42

2.12 Created solid state of layers during scanning. . . . . . . . . . . . . . 42

2.13 Temperature-time diagram for three points in different layers. . . . 43

VI
 LIST OF FIGURES

2.14 State-time diagram for three points in different layers. . . . . . . . . 44

3.1 Results of thermomechanical simulation. (a) temperature field (b)

States (including powder, solid, and fluid) (c) Von Mises stress ten-
sor (d) norm of strain tensor . . . . . . . . . . . . . . . . . . . . . . 54

4.1 Principal elements of the multiscale modeling approach, namely:

CalPhaD thermodynamics, Finite Elements thermal simulation of
SLM processing, and Phase-Field solidification simulation at melt
pool scale. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56

4.2 Schematics of the elliptic temperature field approximation. . . . . 62

4.3 Temperature-dependent material properties for alloy IN718: (a) En-

thalpy, (b) Heat capacity, (c) Density, (d) Conductivity [83]. Cal-
PhaD calculated data (database TCNI8) appear as symbols while
piecewise linear fits used in FE simulation appear as lines (thin
purple for bulk, thick green for powder bed). . . . . . . . . . . . . 72

4.4 Temperature distribution during heating stage of tenth layer as pre-

dicted by finite element simulation. . . . . . . . . . . . . . . . . . . 73

4.5 Melt pool shape within the longitudinal section of Figure 4.4b, com-
paring FE results and the elliptical approximation (Eq. (4.8)): (a)
FE-predicted temperature field (color background) and isotherms
for T = TS = 1554 K (solid blue line and symbols), T = TL =
1625 K (solid red line and symbols) and T = 1569, 1583, 1597, and
1611 K (black solid lines), compared to elliptic approximation along
the same temperatures (white dashed lines); (b) temperature gradi-
ent as a function of the polar angle θ (see Figure 4.2) as predicted
by FE along the solidus (blue circle symbols) and liquidus (red di-
amond symbols) isotherms compared to the elliptic approximation
(solid green line). . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74

VII
LIST OF FIGURES

4.6 Top view of the liquidus and solidus isotherms (solid lines) and
computed boundary between powder bed and bulk (dense) states
(dashed line) when considering different thermal properties in dense
and powder bed states (a), and when the powder bed has the same
properties as the bulk material (b). . . . . . . . . . . . . . . . . . . 75

4.7 Liquidus (red) and solidus (blue) isotherms in the central longitu-
dinal section for the reference simulation of Figure 4.4 (solid lines)
compared to equivalent simulations (dashed lines) with a constant
conductivity (a), constant density (b), or equal properties in the
powder bed and bulk material (c). . . . . . . . . . . . . . . . . . . . 76

4.8 Phase-field simulation results showing grain structure formation (left)

and solute (Nb) concentration field (right). Iso-temperature lines
show T = TL = 1625 K (red), T = 1550 K ≈ TS (blue), and in-
termediate temperature with steps of 5 K (black). The simulation
domain, moving at a velocity V , is delimited with dashed black lines.
Zoomed-in regions at the top right are marked with a dashed rect-
angle in the resulting full-scale maps at t = 1.0 ms and t = 2.0 ms.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79

4.9 Solute (Nb) segregation profiles along different line scans (A–E) in
the solidified region as predicted by phase-field simulations. The
dashed line in the bottom plots marks the solute concentration of
the eutectic point (L→Nifcc +Ni3 Nb) in the Ni-Nb phase diagram. . 80

5.1 Schematics of analytical temperature approximation T (r, θ): (left)

location of elliptical solidus isotherm from the center (x0 , y0 ) of the
melt pool, (right) radial interpolation between center temperature
Tc and solidus temperature TS . . . . . . . . . . . . . . . . . . . . . 88

VIII
 LIST OF FIGURES

5.2 Schematics of analytical temperature approximation T (rθ): (a) lo-

cation of elliptical isotherm from the center (x0 , y0 ) of the melt pool,
(b) radial interpolation between center temperature Tc and solidus
temperature TS ). . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88

5.3 Time evolution (snapshots) of the locations of the solidus and liq-
uidus isotherms in a cross section, comparing FE results of Ref. [79]
(symbols) to Eq. (5.1) (lines). . . . . . . . . . . . . . . . . . . . . . 90

5.4 The most important neighborhood types [141] . . . . . . . . . . . . 92

5.5 Reproducing the results in Ref. [67] . . . . . . . . . . . . . . . . . . 92

5.6 Selected grain orientations (color map) predicted by phase-field (top

row) and cellular automaton (bottom rows) with different grid coars-
ening levels hCA /hPF = 9, 48, and 288. The brighter (washed out)
area is outside of the region Φ considered for the quantitative mea-
surements and comparisons of grain distributions. The dashed black
line represents the limit of the melted region. . . . . . . . . . . . . 99

5.7 Comparison of selected grain orientations (color map) predicted by

phase-field (top row) and cellular automaton (bottom rows) with
different grid coarsening levels hCA /hPF = 9, 48, and 288, within
the regions A, B, and C highlighted in Fig. 5.6. The PF grain map
is overlaid with the inner grain dendritic structure. The dashed
black lines show the limit of the melted region. . . . . . . . . . . . 100

5.8 Difference between orientation maps in CA and PF results, as a

fraction area of the measured area Φ, considering CA results us-
ing either Moore (eight-neighbor) or von Neumann (four-neighbor)
neighborhood, as well as artificially higher (V = 10×VKGT ) or lower
(V = VKGT /10) growth velocities, i.e. respectively multiplying or
dividing the right-hand side of Eq. (5.12) by a factor of 10. Vertical
lines show the measured values of primary (λ1 ) and secondary (λ2 )
dendrite arm spacings. . . . . . . . . . . . . . . . . . . . . . . . . . 102

IX
LIST OF FIGURES

5.9 Difference between orientation maps in CA and PF results, as a

fraction area of the measured area Φ, considering CA results in the
cross-section simulations: (a) averaged over all 15 simulations, (b)
averaged over five simulations for each melt pool shape (reference,
wider/shallower, and deeper/narrower). The shaded areas associ-
ated with each curve represents the distribution between the five
different runs (between minimum and maximum CA-PF difference). 105

5.10 Distribution of grain orientations, averaged over five simulations for

three different melt pool shapes, namely the reference case fitted to
FE results (top row), a wider and shallower melt pool (central row),
and a deeper and narrower melt pool (bottom row), comparing PF
and CA results for a fine (hCA /hPF = 3, left column) and a coarse
(hCA /hPF = 144, right column) CA grid. (PF results are the same
in left and right columns, only duplicated for readability). . . . . . 106

5.11 Grain orientation maps for the reference melt pool dimension given
by FE thermal simulation (w0 = 128 µm, d0 = 96 µm), predicted
by phase field (top row) compared to CA with fine (central row)
and coarse (bottom row) grid. The white dashed lines identify the
melted region. Arrows at the top illustrate the predominant selected
grain orientations within the melt pool. . . . . . . . . . . . . . . . 108

5.12 Grain orientation maps for a wide and shallow melt pool (w0 =
256 µm, d0 = 48 µm), predicted by phase field (top row) compared
to CA with fine (central row) and coarse (bottom row) grid. The
white dashed lines identify the melted region. Arrows at the top il-
lustrate the predominant selected grain orientations within the melt
pool. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109

X
 LIST OF FIGURES

5.13 Grain orientation maps for a deep and narrow melt pool (w0 =
64 µm, d0 = 192 µm), predicted by phase field (left column) com-
pared to CA with fine (central column) and coarse (right column)
grid. The white dashed lines identify the melted region. The bot-
tom row highlights grains with orientations 25◦ ≤ α ≤ 35◦ (red)
and 65◦ ≤ α ≤ 75◦ (green). . . . . . . . . . . . . . . . . . . . . . . 110

5.14 Simulations of cross-section melt pool solidification for the reference

case (i.e. thermal field as extracted from FE results) for five different
initial grain distributions (rows) as predicted by phase-field (left col-
umn), and cellular automaton with fine (central column) and coarse
(right column) spatial discretization. Domain size ≈ 257 µm × 97 µm.114

5.15 Simulations of cross-section melt pool solidification for the refer-

ence case (i.e. thermal field as extracted from FE results) for five
different initial grain distributions (rows) as predicted by phase-
field (left column), and cellular automaton with fine (central col-
umn) and coarse (right column) spatial discretization. Domain size
≈ 513 µm × 49 µm. . . . . . . . . . . . . . . . . . . . . . . . . . . . 115

5.16 Simulations of cross-section melt pool solidification of the deeper/narrower

melt pool for five different initial grain distributions (rows) as pre-
dicted by phase-field (left column), and cellular automaton with fine
(central column) and coarse (right column) spatial discretization.
Domain size ≈ 129 µm × 193 µm. . . . . . . . . . . . . . . . . . . . 116

XI
XII
List of Tables

4.1 Chemical composition of superalloy Inconel 718 . . . . . . . . . . . 57

4.2 Material properties and process parameters for FE simulations. . . . 69

4.3 Material properties and process parameters for PF simulations. . . . 71

5.1 Considered thermal configurations for cross-section simulations. . . 91

5.2 Computational performance for longitudinal section simulations. . . 96

5.3 Computational performance for cross-section simulations for refer-

ence melt pool size. . . . . . . . . . . . . . . . . . . . . . . . . . . . 97

5.4 Computational performance for cross-section simulations for wide

& shallow melt pool size. . . . . . . . . . . . . . . . . . . . . . . . 97

5.5 Computational performance for cross-section simulations for deep

& narrow melt pool size. . . . . . . . . . . . . . . . . . . . . . . . . 98

XIII
XIV
Chapter 1
Introduction

1.1 Additive manufacturing (AM)

It is considered that the use of additive manufacturing (AM) represents a new

industrial revolution after steam engines, computers, and the internet. Unlike sub-
tractive manufacturing, such as conventional machining, casting, and forging pro-
cesses, AM constructs a three-dimensional (3D) structure by continuously adding
the material layer by layer with the guidance of a computer-aided design (CAD)
model. Various types of promising materials, including polymers, metals, ceram-
ics, glasses, bio-materials, and composite materials, have emerged in various types
of AM methods, including stereolithograph (SLA), selective laser sintering (SLS),
fused deposition modeling (FDM), laminated object manufacturing (LOM), binder
jetting (BJ), selective laser melting (SLM).

AM refers to the manufacturing of parts by adding material in layers. Key fo-

cal points of AM, rather than the conventional subtractive technologies for metals
processing, have prompted an expanding consideration by a few core manufactur-
ing industries, such as aerospace, automotive, medical, and consumer products.
Many AM processes for creating metallic parts, using various blends of stock ma-
terial form, material delivery, and heat source have been developed throughout

1
1.1 Additive manufacturing (AM)

Figure 1.1: The classification of metallic powder additive manufacturing pro-

cesses [1].

the years. The first step in the powder-based procedures is spreading the metallic
powder on the bed and then it will be scanned by the beam or fed directly to
the heat source-affected area. Depending on whether the stock material gets fully
melted, partially melted, or consolidated using a polymer binder, the powder bed
processes can be categorized. Over the past few decades, a large number of various
processes, containing all of the previously mentioned categories, have been devel-
oped. In Figure 1.1, the classification schematic of those processes is shown. The
most profit-oriented powder bed processes among many types are selective laser
melting (SLM), electron beam melting (EBM), and selective laser sintering (SLS)
[1].

AM technology is disruptive, and it has the potential to become one of the most
widely used manufacturing technologies. Despite the high cost of AM parts, this
technology has some important advantages compared to the other technologies. In
the following, four main advantages of AM are listed, which result from either the
digital nature of the process chain or the design of complex structures [2].

• Producing a part 100 times using traditional techniques, costs almost the
same as producing 100 single different parts in AM, which is important for
the medical sector and is called mass customization. Since each patient

2
 1.1 Additive manufacturing (AM)

needs unique implants, AM produced implants increase the comfort level of

the patient as well as the durability of the implant.

• Part design plans can be improved to have a similar functionality with

less material, since the material is added rather than removed. For spe-
cific boundary conditions and load cases, we can minimize the weight using
topological optimization which normally causes 30% or more weight saving.
Topological optimization is of great interest in the aerospace industry.

• Minimizing the weight can usually be combined with improving the perfor-
mance. The currently used parts are designed considering the limitations of
common production techniques. There may be other optimal designs for a
certain application but it can be very hard or impossible to produce using the
conventional methods. For instance, cooling channels in conformal cooling
can be located much closer to the regions that are being cooled, at the same
time, the heat exchanger surface can be maximized.

• Finally, this manufacturing method is considered as a fast technique. The

production process is quite slow, but on the other hand, the process chain,
leading up to the production, is fully digital. This means, comparing to the
weeks that are needed for the molds to be prepared, it is a matter of days
before a 3D file is ready to be printed. In addition, making a mold for each
part is expensive, as opposed to AM where the digital design can be adjusted
quickly.

Although AM gives us the ability to build parts made of multiple materials,

freedom in design, and nearly needs no computer-aided design (CAD)-to-part time,
it still needs further developments to gain wider acceptance and establishment into
the modern industrial practice. This process is complicated and contains physical
mechanisms which should be studied more. Consequently, the produced parts by
AM processes do not meet the mechanical performance requirements declared by
manufacturers, prediction of their distortion behavior is difficult, and repeatabil-
ity/reproducibility is an area of underdeveloped research. The mentioned facts,

3
1.2 Selective laser melting (SLM)

along with long lead times, are considered as the main reasons which prevent AM
to be widely used in producing metallic parts. Low mechanical performance of
the produced part is the result of consequence of the presence of porosity, thermal
residual stresses, and distortion. Instabilities while the material is getting consoli-
dated generate the porosity. Also, the thermal gradient during the building of the
part and the thermomechanical nature of the process itself causes the generation
of distortion and residual stresses. A final part with dimensional inaccuracies and
low mechanical properties is the result of residual stresses and distortion [1].

1.2 Selective laser melting (SLM)

One of the most interesting AM techniques is Selective Laser Melting (SLM),

because it gives the possibility to construct the almost full density metallic parts
layer by layer without any post-processing. Dr. M. Fockele and Dr. D. Schwarze
from F & S Stereolithographietechnik GmbH, along with Dr. W. Meiners, Dr. K.
Wissenbach, and Dr. G. Andres from Fraunhofer ILT, developed this process in
order to build metal parts from metallic powders. This technique is a powder bed
fusion process, in which a high-intensity laser is used as an energy source to melt
and fuse the selected regions of powder in layers in accordance to the computer
aided design (CAD) data. The first applied patent for this technology was in 1997
to the German Patent and Trade Mark Office, and it was published in 1998. Also
Das and Beaman registered a patent in 2001 on the basis of their developments in
direct selective laser sintering (SLS) [3], [4].

The SLM procedure contains a series of stages starting from CAD data prepa-
ration to removing the constructed part from the platform. Prior to uploading the
CAD data to the SLM machine to produce the components, the software has to
process the STereoLithography (STL) files, such as Magics, to support structures
for all the hanging features and to create slice data for laser scanning of each layer.
The first step in the building process is putting a thin layer of metal powder on
the plate in the building chamber. In the next step, the selected areas are melted

4
 1.2 Selective laser melting (SLM)

Figure 1.2: Concept of SLM process. (i) High-power laser melts selective areas
of the powder bed. (ii) Process is repeated for successive layers. (iii) Loose
powder removed and finished part revealed [3].

and fused in accordance with the processed data using a high-energy density laser.
After completing the scanning process, the next layer of powder will be deposited
and scanned by the laser. Successive layers of powder are added and scanned until
the part is completely fabricated. In order to allow a single melt vector to com-
pletely fuse with the preceding layer and the neighboring melt vectors, the effective
parameters of the process, such as scanning speed, layer thickness, hatch spacing,
and laser power are adjusted. After completing the laser scanning process, unused
powders are removed from the chamber and the fabricated part will be separated
from the substrate plate manually or using electrical discharge machining (EDM).
The whole process, except the data preparation and removing the scanned com-
ponent from the platform, is automated. Figure 1.2 illustrates the concept of the
SLM building process [3].

SLM processes have attractive aspects for both research and industry. However,
large thermal gradients may induce a high level of residual stress which causes
cracks to appear in the component or the supporting structures during fabrication
or removal from the build plate. Researchers have conducted different studies to
avoid these kinds of problems in various metallic alloys [5].

5
1.2 Selective laser melting (SLM)

1.2.1 Finite element modeling of SLM

The finite element method (FEM) works on the basis of discretizing a contin-
uum domain into a finite number of elements, so that the problem is reduced to
a finite number of unknowns. Having a simpler geometry compared to the actual
structure, it is easier to analyze each element. It is proven that FEM is an efficient
tool to solve eigenvalue, boundary, and initial value problems in most engineering
disciplines. This method is also very popular in studying the physical phenomena
in AM processes. Also, it works well in parametric optimization of AM, which can
prevent conducting several physical experiments. Therefore, the most regularly
used numerical method to predict the temperature and stress fields in SLM is the
finite element method. Many researches have been carried out to simulate a SLM
process and predict the stress formation, part distortion, and the temperature gra-
dient. A finite element model of SLM and Gaussian laser energy density is shown
in Figure 1.3. In order to have a successful FE model of SLM, four consistent as-
pects needs to be addressed : (1) a physics-based method which describes the layer
build-up process, (2) a thermal model to represent scanning laser thermal input, (3)
a laser material interaction model accounting for the transient thermo-mechanical
phenomena, and (4) a material model which defines the temperature-dependent
material property in the cyclic heating and cooling ambiance [1], [6], [7].

1.2.2 Modeling of multi-layer parts

Adding to or removing material from a system causes to appear or disappear

certain elements in the model. In this case, the element birth and death technique
[7]–[9] can be used to activate or deactivate arbitrary elements. In this process,
the elements are activated at the required time points. While using the mentioned
modeling technique in a top-bottom fashion, all the elements, containing those
which will be activated at later steps of the analysis, would be created initially
in the FE model. The term “death” in this method does not mean removing the
elements to attain element death, but means deactivating them using a severe re-
duction factor - generally 1E-9 (ANSYS10, 2007) and multiplying it to the stiffness

6
 1.2 Selective laser melting (SLM)

Figure 1.3: (a) 3D finite element model and (b) Gaussian laser energy density
[6].

matrices (or conductivity, or other analogous quantity) of the corresponding ele-

ments. In the same way, returning their stiffness coefficients to the original values
activates the elements (i.e. ‘born’)[9].

While the element birth method is employed, in each layer the elements are
activated using the chamber temperature as the initial condition of the thermal
analysis. Also for the structural analysis, the elements are activated strain-free in
accordance with the time they are scanned. The dead weights for the deactivated
elements of the powder layer are not included in the weight of the system. An
important point to emphasize is that the bottom elements of the newly added
layer have joint nodes with the top elements of the previous layer at the boundary.
All nodes of the newly added layer are activated at the chamber temperature
except the joint nodes which remain at the attained temperature.

7
1.3 Modeling categories

1.3 Modeling categories

The modeling studies of selective laser melting using finite element considered
here can be classified in three categories, namely aiming at accurate predictions
of:

• Temperature field

• Residual stresses and distortions

• Melt pool characteristics

1.3.1 Temperature field

A key issue to the product quality in 3D printing is the thermal behavior during
the selective laser melting of metal powders. Therefore, the development of heat
transfer analysis during the selective laser melting process in 3D printing indus-
tries is of benefit. Useful information to determine the distribution of the thermal
stresses and to predict the residual stresses and distortion can be obtained from
the temperature field history of the part during the process [1], [10]. The temper-
ature distribution rate in the powder bed and consolidated layers is fast in SLM,
which has drawn many researcher’s attention to investigate the temperature field
in different stages of the process [6], [8], [11]–[18]. Also, since the thermal interac-
tions of successive layers affect the temperature gradients, modeling the problem
involving multiple layers is of great importance [4], [8], [18], [19]. Using processing
parameters, such as laser power and scan speed, the transient thermal behavior
during the SLM process can be controlled. Thus, investigating the relationship
between processing parameters and thermal behavior is of importance to obtain
the desired SLM-fabricated components [7], [19]–[21].

Matsumoto et al. [22] conducted one of the earliest studies on the finite ele-
ment modeling of SLM to obtain the temperature distribution within the powder,
molten, and solidified parts of single metallic layers during the selective laser pro-
cessing, they developed a 2D finite element model. In order to study the effects

8
 1.3 Modeling categories

Figure 1.4: Temperature profiles of consecutive spots [8].

of different parameters on temperature, some other researchers have developed fi-

nite element models of SLM as well [12], [13], [15]. Their investigations showed
that decreasing the scan speed, increasing the laser power, and a low scan interval
causes a higher maximum temperature in the powder bed and wider scan track
width. The temperature field is also significantly influenced by the scan modes. In
transverse scan mode, the temperature in the powder bed has fluctuations while in
lengthwise scan mode it has moderate variation. In addition, researchers [15] have
investigated the influence of the process parameters and material properties on the
predicted temperature profile along the center of the laser beam path. They con-
cluded that the packing density (or porosity) significantly affects the temperature
profile. The distribution at various times that the laser scans across the surface
in the first layer was studied by Roberts et al. [8]. They showed that successive
heating and cooling to the chamber temperature happens within a few tenths of a
millisecond. According to Figure 1.4, they suggested that the irradiated spots are
subjected to rapid thermal cycles. The mentioned thermal cycles are associated
with proportional thermal stress variations. Contuzzi et al. [4] implemented a
finite element simulation of SLM, which can evaluate the evolution and the distri-
bution of the temperature during the process. Their simulations showed that the
maximum temperature in the first layer is more than in the second layer because of

9
1.3 Modeling categories

Figure 1.5: Heat flux magnitude for four different laser power [7].

the higher solid material thickness that raises the heat loss in the bulk. The results
of the performed investigations by Hussein et al. [6] illustrated that as the laser
source moves along the track and the melt pool moves along with the laser source,
the thermal field changes. Also, the temperature gradient in the front side of the
moving laser beam is much sharper than that in the rear side. The reason can be
the molten material which is cooled rapidly has higher conductive properties than
untreated powder in front of the laser. Fu and Guo [7] showed that, according to
Figure 1.5, increasing the laser power results in the enhancement of the heat flux
magnitude. As reported by Li and Gu [19], increasing the distance from the top
of the melt pool, the gradient of the temperature decreases. The reason is that,
compared to the molten material, the solidified material in the previous layers has
higher conductivity. In another research work [20], they showed that because of
the heat accumulation phenomenon in SLM, the average temperature of the pow-
der bed gradually increases during laser scanning. And increasing the scan speed
causes a slight increment of the maximum temperature gradient in the molten
pool, while the maximum temperature gradient increases significantly by increas-
ing the laser power. Ridlbauer et al. [14] presented the temperature distributions
of a polymeric material in various time steps and showed that the simulated and
experimentally measured temperatures are in good agreement with each other.

10
 1.3 Modeling categories

Figure 1.6: Temperature field at different times during 3 passes with (#A)
unidirectional scanning and (#B) alternate scanning [11].

A 3D thermomechanical FE model for ceramic was developed by Chen et al.

[11]. Figure 1.6 compares the temperature fields at different stages during 3 passes
with unidirectional scanning and alternate scanning. Since the whole system is
progressively heated with time, the comparison results show that, in both cases,
the melt pool becomes larger from the first pass to the last one.

The SLM process was simulated using phase transformation models by Bartel et
al. [17]. They developed a constitutive model in the form of a phase transformation
approach. In contrast with common approaches of SLM modling, in this approach
different states of the material — powder, molten, and re-solidified — are captured
as distinct states with respective volume fractions. Li et al. [18] studied the
temperature field modeling of selective laser melting. Using their model, they
have studied the temperature-dependent material properties consisting of enthalpy,
thermal conductivity, thermal expansion coefficient, density, Young’s modulus,
and yield stress. The developed model can estimate the melt pool shape and the

11
1.3 Modeling categories

Figure 1.7: Temperature distribution at the end of the construction process

after dwell time [5].

temperature evolution, which can help to optimize the process parameters in order
to avoid inadequate penetration or over vaporization.

Heat transfer analysis of the SLM process was the subject of some other research
works [5], [16], [21]. Two methods were validated by Zhang et al. [5] regarding
heat transfer resolution in order to reduce the computational time: asynchronous
resolution and time scaling. Both techniques showed good agreement between
their computation times and accuracy levels. Figure 1.7 shows the temperature
distribution at the end of the process as the outcome of their research.

1.3.2 Residual stresses and distortion

A thin metal powder layer is selectively melted by a controlled laser beam,

during the SLM process, and the material experiences many physical transforma-
tions (from powder to liquid and then to solid). Significant distortions, residual
stresses, and, in some cases, cracks and delamination appear due to severe tem-
perature fluctuations. To achieve high-quality specimens in selective laser melting
manufacturing, an accurate estimation of distortion and residual stresses is needed.
The layer which is being scanned during the process undergoes rapid thermal cy-
cles. Residual stresses are generated by sharp temperature gradients, which are
detrimental to the structural integrity and the proper functioning of the built
components. Hence, building a method that can control the quality of the model

12
 1.3 Modeling categories

is particularly important. Experimental measurements of distortion and residual

stresses are expensive and time consuming since many process variables affect SLM.
The quality of calculations of the numerical thermo-mechanical models that can be
used for their estimations depends critically on the accurate transient temperature
field which affects both the distortion and the residual stresses [23]–[25].

High tensile stresses can be seen between the solidified tracks at the side end
of the solid part during the solidification of the neighboring track, which may
cause the appearance of cracks in the layer. As the track length increases, the
amount of distortion of the solid layer increases. Shortening the scanning track
can prevent the distortion of the solid layer on the powder bed. To scan a large
layer, it is suggested to divide the area of the layer into some small segments
with strong frames, so that each segment can be scanned with short tracks. The
residual stresses have their maximum value at the interface between the part and
the substrate [22], [26]. It is shown by Hussein et al. [6] that the elastic modulus
decreases which leads to stress reduction under high temperatures. Hence, closer
to the melt pool, high thermal stresses that appear in the solidified regions tend
to decrease. Also, due to the track overlap, some previously scanned elements are
reheated when the neighboring tracks are being scanned, which causes the residual
stresses to be released. Papadakis et al. [27] presented a model that decreases the
computation time and the modeling effort but still provides reliable structural
results. Using this model, they computed the final shape distortion prior to and
after support cutting, the final stress relaxation after support cutting, and the
stress distribution after thermo-mechanical analysis.

Some FE models for fast prediction of component residual stresses and distor-
tion using different scan strategies in SLM were developed by Li et al. [28]–[30].
The four investigated scan strategies in their study are shown in Figure 1.8. Es-
timation of the component distortions illustrated that the biggest deflection along
the crosswise direction and the smallest one along the longitudinal direction are
produced by the horizontal scanning pattern. The vertical scanning pattern, on
the contrary, causes the smallest deflection along the crosswise direction and the
biggest deflection along the longitudinal direction. In order to reduce the part

13
1.3 Modeling categories

Figure 1.8: Four scanning strategies: (a) horizontal, (b) vertical, (c) successive,
(d) LHI [28].

distortion, the successive scanning strategy is preferred to the LHI pattern. While
in the LHI and the successive patterns the tensile residual stress in the component
was rotated accordingly, both the vertical and the horizontal sequential patterns
produce an almost uniform tensile residual stress. In addition, by increasing the
depth in the part for the mentioned four scanning strategies, the residual stresses
became more tensile.

Some researchers have focused on the magnitude of residual stresses in different

directions (X, Y, and Z) [24], [31], [32]. The Von Mises stress and three components
of the residual stresses after cooling the powder bed are shown in Figure 1.9. This
figure illustrates the stress distributions in the powder bed for different oriented
components are different and the area near the edge of the part, mainly at the end
of the scanning track, have larger stresses.

Great attention has been paid to the effect of the process parameters, such as
overlap between adjacent paths, scan speed, different layers, scan strategy, and
laser power on distortion and residual stresses [23], [25], [33]. According to the
results, by increasing the scanning speed and decreasing the laser power, the va-
porization gradually weakens. As the scanning speed decreases and the laser power
increases, within the optimum forming window, the residual stress rises [33]. Fig-
ure 1.10 shows the residual stresses in different directions in three layers. The

14
 1.3 Modeling categories

Figure 1.9: The residual stress distributions when the powder bed was cooled
down at 100 s: (a) X-component of stress along the scan direction; (b) Y-
component of stress transversal to the scan direction; (c) Z-component of stress
through the thickness of the layer; (d) Von Mises stress [31].

optimization of SLM process parameters was investigated by Wang et al. [23].

The laser scanning strategy, the scanning speed, and the laser power affect the
heat dissipation, the cooling rate, and the energy input, respectively. The process
parameters in various aspects have an influence on the temperature gradient dis-
tribution of the melting layer affects the final distribution of the residual stress
and causes different thermal stresses. Hence, by optimizing the process parame-
ters combination, one can effectively improve the distribution of thermal stress,
enhance the quality of the part, and reduce the residual stress.

1.3.3 Melt pool characteristics

Melt pool dimensions have a significant effect to ensure that a strong bonding
exists between the powder particles. The porosity and therefore the quality of the
specimen are prominently influenced by the bonding of the particles. Moreover,
the melt pool length and width which have effect on the already built areas should

15
1.3 Modeling categories

Figure 1.10: The residual stress (1) layer 1; (2) layer 2; (3) layer 3. a, d, g
Residual stress along x-direction. b, e, h Residual stress along y-direction. c, f,
i Residual stress along z-direction [23].

be thoroughly considered. Loss of surface quality and dimensional inaccuracies

may be caused by remelting of such areas. The point which possesses the highest
temperature in the melt pool is called “center of the melt pool”. Contrary to
what normally is expected, various researchers have reported that this point is
not located at the center of the beam spot. It is observed that the location of
the center of the melt pool is a little displaced backward from the laser scanning
direction. By increasing the beam power, this displacement can be increased [1].
Figure 1.11 shows this phenomenon.

The process parameters, such as laser power and scanning speed, mainly control
the average applied energy per unit volume (which affects the dimension of the
molten pool). Researchers have reported that increasing the scanning speed results
in a larger length to diameter ratio in a molten pool. If this ratio exceeds π in
a molten pool, balling effect will appear. Breaking up the molten pool into balls
specifies the balling effect which causes the degradation of the surface integrity
of the final specimens. Hence, the balling effect limits the range of the optimal

16
 1.3 Modeling categories

Figure 1.11: Displacement of the melt pool center with respect to the laser beam
[1].

scanning speed to get the optimal surface integrity. Some reports suggest that
the range of the optimal speed can be expanded by increasing the laser power
[7]. The size of the melt pool is the subject of many studies [14], [34]–[36]. Also,
the influence of the process parameters on the melt pool characteristics has been
investigated by several researchers [10], [19], [20], [37]–[39]. As reported by Li and
Gu [20], increasing the laser power increases the width and depth of the molten
pool, while increasing the scan speed decreases them. The influence of different
parameters on the melt pool geometry was studied by Cheng and Chou [37]. They
showed that decreasing the beam diameter may cause reduce the scanning length
that is required for melt pool fully development. Slower beam speed and higher
beam power may lead to longer scanning length to reach the steady state of the melt
pool. The simulation results of the melt pool length evolution process with different
process parameters are shown in Figure 1.12. The volume shrinkage should be
considered in the numerical simulation because it influences the dimensions of the
molten pool. While the effect of the volume shrinkage is being considered, the
length and the width of the molten pool are smaller and the depth is slightly
bigger [10]. Loh et al. [40] discussed the progression of the melt pool in detail.

17
1.3 Modeling categories

Figure 1.12: Melt pool length evolution for different process parameters [37].

The side view of the melt pool at six consecutive time points is presented in Figure
1.13, which shows the propagation of its front with the laser beam movement.

The laser beam center, at a steady state, is at a radius from the melt front.
After volume shrinkage, heat will be conducted to the adjacent bottom powder
and the molten powder, while the laser beam hits on the molten powder. The
bottom powder is melted and conducts heat to the powder which is above it. In
addition, the heat which is transferred from the edge of the moving laser beam to
the powder on top will melt it, causing the melt pool front to progress laterally
but not from top to bottom.

18
 1.3 Modeling categories

Figure 1.13: The side view of the propagation of the melt pool [40].

19
1.4 Multiscale simulation of SLM

1.4 Multiscale simulation of SLM

Multiple modeling techniques relevant to different length/time scales and/or

different physical phenomena are widely used in ICME. Despite the fact that a
wide range of models have been developed, effective bridging across models and
scales remains a significant challenge. Many modeling approaches for powder-bed
fusion processes have emerged as a result of the growing use of metal additive
manufacturing in the last decade [41]–[43] . Multi-physics models have been devel-
oped at the macroscopic/process scale to combine fluid and solid mechanics with
thermal transport (conduction, convection, and radiation) during melting and so-
lidification phenomena caused by a moving heat source (e.g. laser or electron
beam) above a powder bed [44]–[49]. These studies have demonstrated the impor-
tance of fluid flow within the melt pool — in particular, Marangoni convection and
recoil pressure — on the formation of critical defects, such as porosity, spattering,
denudation, and balling [44], [46]–[49]. The scanning method, such as the hatch
pattern [44], [48] like Figure 1.14, was also shown to have a significant impact on
the build quality. Residual stresses were shown to be strongly influenced by the
scanning path, the cross-section thickness, and the presence of additional lasers,
with considerable variations between vertical (build) and horizontal directions [48].
Meso-scale phase-field (PF) models of powder particle melting and solidification
were also shown to capture the effect of heat source power and speed on densi-
fication, defects, and surface morphology [50]–[52]. In the case of high growth
velocities reaching absolute interfacial stability, three-dimensional PF simulations
of polycrystalline grain structures at the melt pool scale were recently reported
[53]. Figure 1.15 shows the evolution of the predicted microstructure using a
three-dimensional PF simulation. However, it is restricted to alloys and growth
velocities that result in full solute trapping and solid-liquid interface planar re-
stabilization. Grain growth competition often includes complicated mechanisms,
such as dendritic side-branching and impingement of individual dendrites or cells,
since most metallic alloys develop dendritic or cellular microstructures during AM
[54], [55].

20
 1.4 Multiscale simulation of SLM

Figure 1.14: Schematic of scan paths: (a) Z-shaped and (b) S-shaped[44].

The choice of the model at the grain structure scale typically leads to the
classic trade-off between efficiency (and hence scale) and physics-based accuracy.
Among the coarse-grained efficient techniques, those based on the Kinetic Monte
Carlo (KMC) method makes it easy to simulate three-dimensionally at the entire
melt pool size, including multiple layers [56], [57]. However, they usually do not
integrate some essential physics behind polycrystalline microstructure selection
(e.g. the anisotropic growth kinetics and preferred growth directions of crystalline
grains). Models that combine mesoscale cellular automata (CA) with macro-scale
thermomechanics (e.g. employing finite elements or differences) offer a good com-
promise of efficiency and physics-based considerations [58]–[64]. In such models,
grains can be constructed from polyhedral building blocks, whose vertices mark
the preferred crystallographic growth directions. Figure 1.16 compares analytical
and numerical (CA) prediction of a 3D dendritic grain envelope, with and without
temperature gradient. Growth velocities in these directions follow simplified, yet
physics-based, kinetic laws for crystal growth — e.g. using Ivantsov-based rela-
tions or power laws relying on local supersaturation or undercooling. As such,
these models still include some adjustable phenomenological parameters. It was
recently shown that polycrystalline growth can be predicted by CA-based mod-
els with an accuracy comparable to that of phase-field, as long as the cell size is
adjusted to the length scale relevant to the grain growth competition — e.g. the
“height” difference between two competing grains [65] or the spacing among active
secondary branches [66].

21
1.4 Multiscale simulation of SLM

Figure 1.15: Evolution of the predicted microstructure in different time

steps[53].

22
 1.4 Multiscale simulation of SLM

Figure 1.16: 3D view of (a) analytical and (b) CA prediction of a dendritic grain
envelope. In (1) the temperature is uniform (G = 0 K/m) whereas in (2) G =
250 K/m [67].

Most accurate physics-based models at the scale of dendritic/cellular arrays

require a numerical discretization corresponding to the microstructural scale of
interest, resulting in computationally costly simulations. Phase-field (PF) models
are undoubtedly the most realistic method for simulating the development of mor-
phologically complex interfaces using just thermodynamics and kinetics[68]–[73].
The PF method has been proven especially effective in the field of solidification,
where “mesoscale” interface formulations have been used to stay true to the well-
known sharp-interface problem, even for diffuse interfaces far broader than the real
interface width or capillary length [68], [72]. However, the usual microstructural
length scale, i.e. the local interface curvature, limits spatial discretization upwards
[71].

Some computational techniques that combine macroscopic thermal models with

lower-scale microstructure models have been presented[41]–[43]. Recent investiga-
tions have employed macroscopic AM processing thermal simulations to provide
thermal conditions for smaller scale PF solidification simulations[74]–[77]. Yet,

23
1.4 Multiscale simulation of SLM

while full melt pool simulations appear within reach using computationally-efficient
parallelized implementations (see, e.g. [78]), PF simulations have been mostly re-
stricted to the growth of a handful of dendrites with thermal conditions relevant to
a subset of the melt pool region [74]–[77]. In the scope of our project, we performed
a PF solidification simulation at the scale of the entire melt pool, which gives new
insights into grain texture selection through polycrystalline growth competition
under realistic SLM circumstances, down to individual dendrites [79].

24
Chapter 2
Thermal Finite Element Framework

The SLM procedure contains a series of stages, from CAD data preparation to
removing the constructed part from the platform. Before uploading the CAD data
to the SLM machine to produce the components, the software has to process the
STereoLithography (STL) files, created, for example, by the Magics software, to
support structures for all the hanging features and to create slice data for laser-
scanning of each layer. The first step in the manufacturing process consists in
placing a thin layer of metal powder on the substrate (usually metallic), which
is located on a built platform. In the next step, the selected areas are melted
and fused in accordance with the processed data using a high-energy density laser.
After the scanning process, the next layer of powder will be deposited and scanned
by the laser. Successive layers of powder are added and scanned until the part is
completely fabricated. In order to allow a single melt vector to completely fuse
with the preceding layer and the neighboring melt vectors, effective parameters
of the process, such as scanning speed, layer thickness, hatch spacing, and laser
power, are adjusted. After completing the laser scanning process, unused powder is
removed from the chamber, and the fabricated part will be separated manually or
using electrical discharge machining (EDM) from the substrate plate. The whole
process, except the data preparation and removing the scanned component from

25
 Figure 2.1: Selective laser melting process.

the platform, is automated. Figure 2.1 illustrates the concept of the SLM building
process.

A thermal model with spatial accuracy is critical to simulating the laser heating
in SLM. Besides an accurate thermal model, an effective material model that
considers the property changes in the cyclic heating and cooling environment is
also essential. It has been proved that thermophysical properties, such as thermal
conductivity and specific heat change significantly when a material is heated up
to a liquid state from a solid state. Moreover, the thermal conductivity for solid
substrate and powder bed is quite different. So it is essential to have a model that
considers all of these issues [7].

Four coherent argument factors must be taken into account for a finite element
model of SLM to be successful: (1) a thermal model to represent the thermal input
from the scanning laser, (2) a material model to specify the temperature-dependent
material properties in the cyclic heat and cooling environment, (3) a physics-based
method to explain the layer build-up process, and (4) a laser material interaction
model to take into account transient thermo-mechanical phenomena.

26
 2.1 Modeling of laser thermal input

This chapter describes the temperature field model using a moving heat source
approach. The numerical solution in this research was carried out using the in-
house finite element program IRIS, and the material library MUESLI [80], imple-
mented in the C++ programming language, to obtain a thorough understanding of
the thermal behavior. A Galerkin method employing hexahedral elements was used
to discretize the initial boundary value problem in space. The resulting semidis-
crete equations were integrated in time with the implicit Backward-Euler method
[81].

2.1 Modeling of laser thermal input

In the simulation, the laser is defined as an energy source term. Both surface or
volumetric thermal loading can be used to model the heat source. In order to keep
the computational cost as low as possible, the heat source is considered as a 2D
heat flux applied on the surface of the powder bed. The thermal load transferred
by the laser is called laser irradiance. The top-hat and Gaussian models, which
are the most widely adopted models, are used to model the distribution of laser
irradiance across the beam. In the Gaussian laser beam model the laser irradiance
is symmetric with respect to its propagation direction, and in most cases, the
maximum irradiance is at the center of the beam pattern. The thermal heat flux
is described as
2r2
 
2AP
I= exp − 2 , (2.1)
πw2 w
where A is the laser absorptivity of powder material, P is the laser power, w is the
laser spot diameter, and r is the radial distance from the center of the laser beam.
Figure 2.2 shows the irradiance values of the laser heat source with a laser radius
equal to 35 µm.

It is important to be able to model different laser scanning paths that can

investigate different shapes and the effect of parameters on them (especially for
thermo-mechanical investigations). So paths have been defined for the laser. Fig-
ure 2.3 presents some of these paths, including filled and empty circles, filled and

27
2.1 Modeling of laser thermal input

Figure 2.2: Laser beam irradiance with a Gaussian distribution [34].

Figure 2.3: Different laser scanning paths.

empty rectangles, and a double loop shape. Usually, two scan strategies can be
used (as shown in Figure 2.4): unidirectional and zig-zag patterns. The zig-zag
strategy is a frequently used strategy, but using the unidirectional strategy is not
common because it leads to lower density.

28
 2.1 Modeling of laser thermal input

Figure 2.4: Different scan strategies.

2.1.1 Laser scan path

Typically, a “slicer” software creates the laser scan path based on CAD data as
a G-code or other specific machine language file. To define a collection of cross-
sections, the slicer software computes the intersection of the CAD model with
several planes. The present study uses G-code files to recover the discretized laser
beam paths. The FE solver is then used to analyze and transfer each trajectory.

The procedures described above were implemented in C++. During the prepa-
ration stage, the CAD model of the analyzed geometry is created, and the G-code
of the construction process (laser trajectory, dwell time, etc. ) is generated by
the “Slicer”. Thanks to the “Interpreter”, relevant information for the numerical
simulation is obtained, and the determination of the fractions of each layer can
be done. The developed numerical model is then launched to perform the simu-
lation of the whole construction process by chaining the treatment of each layer,
fractions of each layer, inter-layer dwell time, deposition of a new layer, and re-
meshing. Figure 2.5 shows an example in which a G-code file has been used to
create the laser scan path.

29
2.2 Temperature-dependent alloy properties

Figure 2.5: Using G-code file to define the laser path.

2.2 Temperature-dependent alloy properties

We calculated alloy properties as a function of the temperature assuming com-

plete equilibrium, i.e., a lever rule solidification path, but the methodology would
be just as readily applicable considering alternative assumptions, e.g., Gulliver-
Scheil [82]. In particular, within a temperature range spanning all phase trans-
formations, here from 250 K to 5000 K, we estimated the temperature-dependent
density, ρ, as well as the enthalpy per unit mass, h. From the latter, we calculated
an effective heat capacity cp (T ) = ∂h(T )/∂T that includes not only the actual
heat capacity but also the enthalpies (latent heat) of transformations in the rel-
evant temperature ranges. The heat capacity is an extensive property of matter
that depends on the number of species in the system and is sensitive to phase
changes. From the tabulation of the fraction of liquid versus temperature, one can
also extract the key transformation temperatures, namely liquidus (TL ), solidus
(TS ), and boiling (TV ) temperatures.

In particular, we apply the approach to a multicomponent Inconel 718 (IN718)

superalloy using the software ThermoCalc with Ni-alloys database TCNI8. In the
interest of computational efficiency, the resulting h(T ) and ρ(T ) used in the FE cal-

30
 2.2 Temperature-dependent alloy properties

culations are approximated by piecewise linear functions (see Fig. 4.3). CalPhaD-
calculated properties can also be complemented by any further temperature-dependent
property from other sources, such as illustrated here using a thermal conductiv-
ity, κ(T ), from the literature [83]. Due to the high thermal conductivity of metals,
powder-bed technology’s main heat transfer processes are thermal loss via the pow-
der that a component is immersed in and heat conduction through the building
platform.

2.2.1 Properties of the powder bed

Researchers have reported different models to calculate the effective thermal

conductivity and density of the powder bed. Some of them considered constant
amount for thermal conductivity of the powder or even neglected it [8], [38], [84].
Also, some authors [85] used experimental data to obtain the properties of the
powder. Thummler and Oberacker [86] presented a simple model that estimates
the thermal conductivity and density of the powder, which was adopted by some
researchers. In this model, they considered the powder properties as a fraction of
bulk material properties, which overestimates the thermal conductivity of powder:

ρp = ρb (1 − ξ), (2.2)

κp = κb (1 − ξ), (2.3)

where ρp and κp are density and thermal conductivity of powder, ρb , respectively,

and κb are density and thermal conductivity of bulk material, respectively. And ξ
is the porosity of the material.

Another model is a semi-empirical one [87]. These authors account for the
powder packing void and gas atmosphere. Compared with the previous model,
this last one is more accurate and the conductivity, for example, is given by the
expression
κb (1 − ξ)
κp = κb , (2.4)
1 + ϑ κg

31
2.3 Boundary conditions

where kg is the thermal conductivity of the gas atmosphere, ξ is the volume fraction
of voids, and ϑ is an exponential factor as follows

ϑ = 0.02 × 102(ϵ−0.3) . (2.5)

Also, these same authors estimated the density of the powder bed using a classical
rule of mixture
ρp = (1 − ξ)ρb + ξρg , (2.6)

where ρg is the density of the gas atmosphere.

Sih and Barlow [88] presented a theoretical model derived from the Zehner-
Schlunder-Damkohler (ZSD) equation [89], [90]. By assuming spherical particles,
they proposed an expression for the conductivity of the form

κp p κr
= (1 − 1 − ξ)(1 + ξ )
κg κg
(2.7)
" ! #
p 2 1 κs κr
+ 1−ξ ln( ) − 1 + ,
1 − κκgs 1 − κκgs κg κg

where κr is the thermal conductivity due to radiation among particles

kr = 4F σT 3 d. (2.8)

1
Here F is a view factor approximately equal to 3
, σ is the Stefan-Boltzmann
constant, T is temperature, and d is the average diameter of powder particles.

In this thesis, we use equations (2.6) and (2.7) to calculate the density and
thermal conductivity of powder, respectively. We consider Argon and the gas atmo-
sphere, which according to Ref. [87], thermal conductivity is equal to 0.0176 W/(m.K).

2.3 Boundary conditions

Heat convection and heat radiation across the surfaces in contact with the
environment must be accounted for as boundary conditions. Figure 2.6 shows

32
 2.3 Boundary conditions

different types of heat transfer including conduction, convection, and radiation

that take place during the selective laser melting process. Also, at the bottom of
the substrate, a Dirichlet boundary condition is considered, and heat conduction
exists between the substrate and the layers.

Figure 2.6: Heat transfer during SLM additive manufacturing.

Heat convection

According to Newton’s law, the convective heat transfer through an external

surface in contact with the environment can be modeled as

qconv (T ) = hconv (T − T0 ). (2.9)

In Eq. (2.9), hconv is the heat transfer coefficient, T is the object’s surface temper-
ature, and T0 is the temperature of the environment which is in contact with the
surface, and it can be assumed constant.

33
2.4 Modeling of layer build-up and state change

Heat radiation

Because of the high-temperature field during the laser melting additive manu-
facturing, radiation is another important heat loss mechanism. By using Stefan-
Boltzmann’s law, the radiation heat flux can be expressed as

qrad (T ) = σϵ(T 4 − T04 ), (2.10)

where σ and ϵ are the Stefan-Boltzmann constant and the emissivity of the radia-
tion, respectively.

2.4 Modeling of layer build-up and state change

The model in this study employs the element birth and death method to sim-
ulate the addition of layers with time. The procedure involves the activation of
elements at the time instants when they are required. To use this technique, all
elements would be created initially in the finite element model, including those to
be activated at later stages of the analysis and the elements in the substrate. Then,
all of the elements within the powder layers are deactivated at the beginning of
the analysis and only the substrate remains. Last, the elements in the first powder
layer are activated on the top of the substrate followed by the first laser scan.
This process is repeated for the successive layers. A latency elapses between the
activation of layers in the simulation, which mimics the new powder application
by the roller in the process. Figure 2.7 represents the layer build-up schematic.

The material above the substrate can exist at three different states: powder,
solid, and fluid. The “powder” state is used for the powder bed, which is a com-
bination of powder particles and gas, with homogeneous properties reflecting that
of the solid, gas, and morphological descriptors of the powder bed (e.g. average
particle size, packing factor, etc.), as described in the following paragraphs. After
activating a new layer, the initial state of the added material is powder. At each
time step, when the temperature of an element exceeds the alloy liquidus tem-

34
 2.4 Modeling of layer build-up and state change

perature, TL , a transition occurs from powder to fluid state. Subsequently, the

state of a fluid material point whose temperature becomes lower than the solidus
temperature, TS , is switched from fluid to solid.

The state of a point can change several times between fluid and solid, depend-
ing on the local thermal history. However, once an element has switched from
its initial powder state to a dense state, its properties remain those of dense ma-
terial for the rest of the simulation. The subsequent changes between solid and
fluid states are intrinsically represented by the temperature-dependent properties
via their variations upon phase transformation (see Section 4.6.1 and Figure 4.3).
The latent heat of transformations is not explicitly introduced in Eq. (2.11). In-
stead, the model uses an effective heat capacity, calculated as cp (T ) = ∂h(T )/∂T
for the entire temperature range, which therefore incorporates the effect of these
transformations. In practice, we use piecewise linear fits of the CalPhaD calcu-
lated h(T ) and ρ(T ), and we use the high-slope region of h(T ) just above TV for
any temperature T ≥ TV (see later Figure 4.3). This way, the critical effect of
evaporative cooling [91] is incorporated phenomenologically and the temperature
saturates naturally when T exceeds TV , without an artificial increase of thermal
conductivity [40] or reduction of the heat source term [92] in the vicinity of the
evaporation temperature.

Three different states containing “powder”, “fluid”, and “solid” are defined for
each material point in the model. At every evaluation, the local temperature is
used to update the state of the particle. As explained before, when the local
temperature of a powder point exceeds the liquidus temperature of the material,
its local state is changed to fluid. Only if this local temperature is later reduced
below the solidus threshold, the local state of this point will be changed again to
solid and it will remain like this indefinitely unless the temperature exceeds the
liquidus value in which case the state will be again that of liquidus and the process
will start over. This process can take place many times during the simulation.

35
2.5 Thermal solver

Figure 2.7: Schematic of layer build-up process.

2.5 Thermal solver

We consider the thermal problem of a moving heat source above a powder bed.
Neglecting fluid flow is expected to have a strong influence. However, for printing
in conduction mode, a thermal model is expected to provide a reasonable estimate
of the temperature profile in and around the melt pool.

If we consider Ω as a bounded domain in R3 , the governing equation of the

heat transfer problem can be written as

ρ(T )cp (T )Ṫ (x) + ∇ · q(x) = q̇(x), x ∈ Ω, (2.11)

where ρ denotes the density, cp the specific heat at the constant pressure, T the
temperature, q the heat flux vector, and q̇ the heat supplied per heat volume.
The density, the specific heat, and the heat source might be a function of the
point x, although this is not indicated in these equations. The considered thermal
constitutive model is nonlinear, and it relates to the heat flux and the temperature
gradient, thus
q(x) = −κ(T, x)∇T. (2.12)

36
 2.5 Thermal solver

In the above expression, κ is a scalar and denotes the conductivity of the ma-
terial, which can be non-homogeneous and temperature-dependent. However, it
is assumed to be isotropic. The partial differential equation (2.11) must be sup-
plemented with boundary and initial conditions to complete the initial boundary
value problem. A starting temperature for all the points can be used to impose the
initial conditions, which are the first step temperatures in the transient solutions.
So at the initial instant the temperature field must be of the form

T (x, 0) = T0 (x). (2.13)

To solve the above initial-boundary value problem, it has to be discretized in

time and space. The Galerkin method has been used in space. A variational
statement of the problem is required for this method, which can be obtained by
transforming the strong-form of the partial differential equation into its weak form.
To define the latter, let V h denote a finite dimensional subspace of the (infinite
dimensional) solution space of the initial boundary value problem (2.11)-(2.13).
Then, the finite element semidiscrete solution is the temperature T h ∈ V h such
that, for all W h ∈ V h , the following holds:
Z Z Z
h h h h h
ρcp Ṫ W dΩ + ∇T k(T )∇W dΩ = q̇W h dΩ
Ω Ω Z Z Ω
+ hsup W h dA + hconv (T h − T0 )W h dA (2.14)
Γ1 Γ2
Z
+ σ((T h )4 − T04 )W h dA.
Γ3

Recognising the arbitrariness of the test functions, a standard manipulation leads

to the matrix form of the problem, namely,
 
dT
[C(T)] + [K(T)]{T} − {f} = 0, (2.15)
dt

37
2.5 Thermal solver

where [C] is heat capacity matrix, [K] is the thermal stiffness matrix, {T} is the
vector of nodal unknown temperature values and {f} is internal and external flux
vector.

A description of the matrices in Eq.(2.15) is given below:

• The thermal stiffness matrix [K] can be expressed as follows

Z Z
T
[K(T)] = [B] [k(T)][B]dV + h[N]T [N]dS, (2.16)
V S

where [B] is the matrix containing the first derivatives of shape functions, dV
denotes the volume of the element and h is the convective heat transfer coef-
ficient, dS denotes the surface area of the element, and Tf is the temperature
of the environment into the working chamber.

• The thermal specific heat matrix [C] can be expressed as follows

Z
[C(T)] = cp (T )[N]T [N]dV, (2.17)
V

where [N] is the three-dimensional nodal shape function.

• The thermal flux vector [F] can be expressed as follows

Z Z Z
{f} = T
q̇[N] dV − T
hsup [N] dS + hTf [N]T dS, (2.18)
V S S

where q is the input flux depending on the boundary conditions.

An explicit or implicit time integration scheme can be used to discretize time.

In this research, the backward Euler implicit method has been used. In this scheme,
the time derivative has been replaced by a forward difference approximation. And
then, it evaluates the remaining terms at the unknown time instant. Given the
temperature Tnh at time tn , its value at time tn+1 = tn + ∆t is the solution of the
problem which can be calculated using the following equation

h
= inv(∆tKab )(Tnh ) −∆tK(Tnh )(Tnh ) + ∆tF (Tnh ) .
 
∆Tn+1 (2.19)

38
 2.6 Verification

h
This problem is nonlinear, having the temperature field Tn+1 as the unknown.
h,(k)
Newton-Raphson method has been used to solve this problem. Having Tn+1 as
h,(k+1) h,(k)
the solution after k iterations, an improved solution Tn+1 = Tn+1 + ∆T h can
been found. The problem has been linearized to obtain the unknown increment
∆T h .

2.6 Verification

For verification of the obtained results, two samples similar to the ones em-
ployed in references [74], [93] were modeled. Figures 2.8 and 2.9 show the com-
parison between obtained results and [74], [93], respectively. As it can be seen,
the magnitude and gradient of temperatures in both figures have an acceptable
agreement.

Figure 2.8: Comparing the result of a) IRIS and b) [74].

39
2.7 Results

Figure 2.9: Comparing the result of a) IRIS and b) [93].

2.7 Results

In this section, the material properties and process parameters have been chosen
the same as in Zhang et al. [5]. To prevent losing any data during the simulation,
two criteria were considered for the mesh size. The first one is that the size of
each element should be smaller than the laser spot area, and the second one is
that the path which crosses each time step should be shorter than the size of an
element. The modeling of each layer is divided into two parts: the heating stage
and the cooling stage. After the cooling stage, a new layer of powder is added,
and the heating stage of the next layer will start. This procedure continues until
the cooling stage of the last layer. Figure 2.10 presents the temperature gradient
in the scanning path of the laser during the heating stage.

As mentioned in section 2.4, three states with different material properties have
been implemented. At first, the elements of the active layer are powder (shown
with blue color). Then, the laser starts to scan the layer, and meanwhile, the
properties of each element whose temperature reaches the liquidus temperature of
the material will be switched from powder to fluid. The fluid state is shown in
green color. The temperature of the fluid elements will start to decrease through
conduction and convection after the laser point is passed over the corresponding
elements. And finally, the temperature of each fluid element is decreased until
it reaches the solidus temperature of the material. In this stage, the material

40
 2.7 Results

Figure 2.10: Temperature field.

properties of the element are switched to solid properties and are shown with
red color. Also, if the temperature of a solid element increases to the liquidus
temperature, its properties will change to fluid again. Figure 2.11 shows all of
these three states during the cooling stage of the layer.

Also, Figure 2.12 presents created solid state of specimen after some layers
during the scanning process.

Figure 2.13 illustrates the temperature vs. time diagram of three points for a
three-layer model in which all the points located on the same vertical line. The
diagram corresponding to the point in the bottom layer has three peaks. The first
one is related to the moment that the laser passes through this point in the first
layer, and the other two peaks are related to the moment that the laser passes
above it while on the second and third layers, respectively. As can be seen in the
figure, by adding these layers, the peak temperatures for the point in the first layer
decrease because these layers exist between the point and the laser. The heating
and cooling times for each layer were 11.45 seconds. When a new layer is added,
due to the conduction between the layers, the temperature of the points in the

41
2.7 Results

Figure 2.11: Three different states containing: Powder, Fluid, and Solid.

Figure 2.12: Created solid state of layers during scanning.

previous layers is decreased suddenly, and the temperature of the point in the new
layer is increased. Also, the peak temperature corresponds to the time in which
the laser spot reaches the mentioned coordinate. The first peak temperature of
the point in the first layer, due to different boundary conditions, is higher than
the first peak load of the other layers.

42
 2.7 Results

Figure 2.13: Temperature-time diagram for three points in different layers.

A diagram of state vs. time for these three points is shown in Figure 2.14. The
values of 0, 1, and 2 in the contours correspond to the powder, fluid, and solid
states, respectively. The result for a state is smoothed between Gauss points, so
the states are not exactly equal to one or two. For the points in the first and
second layers, when the laser passes through every point, their states are changed
to fluid, and when the new layer is added, they are changed to solid due to the
decreased temperature caused by the heat transfer between layers. But for the
point in the third layer, the state is changed from powder to solid when the laser
passes through the point. This shows that the temperature becomes lower than
solidus temperature rapidly.

43
2.7 Results

Figure 2.14: State-time diagram for three points in different layers.

44
Chapter 3
Thermo-Mechanical Finite Element
Framework

The general equations that describe all initial boundary value problems in small
strain thermomechanics are introduced in this chapter. As a result, the concepts
proposed to apply to all potential material models and provide general constraints
that must be followed by all in order to be thermodynamically sound. This chapter
is not meant to provide a full discussion of these types of mechanical problems,
but rather a summary of the most important findings that will be later recalled for
material model development. We assume the reader is familiar with mechanical
terms such as stress and strain, as well as thermodynamic concepts such as energy,
entropy, temperature, and heat. Many complete monographs on this subject are
available to the reader [94], [95].

We limit our discussion to materials with internal variables, which are among
the many types of material models that may be utilized in small strain thermome-
chanics. The defining property of these materials is that each material point has a
state that is totally characterized by its strain tensor, temperature, and a collection
of scalars, vectors, or tensors known as internal variables. In a sense, these vari-

45
3.1 Finite element method for fully coupled thermomechanical analysis

ables encapsulate all of the information imprinted in the material microstructure

by the preceding history of deformation and temperature.

3.1 Finite element method for fully coupled thermo-

mechanical analysis

Both the mechanical and thermal fields must be solved simultaneously for any
time of interest throughout the loading process for fully coupled thermo-mechanical
analyses of the response of a deformable body to be investigated. A precise math-
ematical description of the corresponding initial boundary value problem for the
thermo-mechanical process which is established in this subsection based on the
governing equations and the initial and boundary conditions of the thermal and
mechanical fields as introduced below are required for an accurate solution of the
coupled fields.

3.1.1 Model description

The mechanical response of deformable bodies is investigated in small strain

thermomechanics. A body is defined to be an open subset Ω of three-dimensional
Euclidean space, whose boundary is denoted Γ. The position of material points
in the body x ∈ Ω, may be determined, and their deformation can be described
analytically using the displacement u : Ω → R3 , a smooth vector field in which
the deformation of the body maps points x to new positions x + u(x ).

Generally, mechanical and thermal loading can be applied to a deformable

body. Throughout the whole domain of the body, a volume force b : Ω → R3
and a heat supply h : Ω → R are defined. These have force and heat per unit
mass dimensions, respectively. Furthermore, the body may be loaded across its
boundaries. For that, let the boundary split into two disjoint subsets Γu and Γt .
Then, the displacement field would have a known value on Γu and there would be
an imposed surface vector field of tractions t on Γt .

46
 3.1 Finite element method for fully coupled thermomechanical analysis

3.1.2 The first law of thermodynamics

The first law of thermodynamics states that in a closed system, the total energy
is conserved. In the case of a continuum body, this is equal to saying that all
mechanical and thermal energy exerted on it is either used to change the kinetic
energy or stored elastically:
Z Z Z Z Z Z 
d 2
ρb · v dV + t · v dA + ρe
hdV + hdA = ρ|v | dV + ρedV .
Ω Γt Ω Γq dt Ω Ω
(3.1)
A standard localization argument provides the differential expression of the first
law, that is,
ρė = σ · ε̇ + ρe
h − ∇ · h. (3.2)

3.1.3 Materials with internal variables

The thermomechanical behavior of a material point is determined by its present

and previous strain and temperature values. As a result, in order to model ma-
terials, it is necessary to maintain track of the history of strain and temperature
across time and for all points of interest. Except in very simple situations, this is
impractical or perhaps impossible in terms of analysis and computations. Instead,
we will only investigate a class of materials whose response is univocally described
at any given time by the strain, temperature (or any other thermodynamic vari-
able), and a (limited) collection of scalars, vectors, or tensors referred to as the
internal variables, which we will label ξ. There is no unique approach to define
these variables, thus it is up to the modeler to choose them properly so that there
are not too many, but enough to construct a model that is accurate sufficiently.
Internal variables should be chosen to express the microstructural changes that are
assumed to be relevant for the response as a general, though sometimes ambiguous,
guideline.

A material’s internal energy density, e, is a scalar function with internal vari-

ables e = e(ε, ξ, η). Where ε is the strain tensor, ξ the internal variables, and η

47
3.1 Finite element method for fully coupled thermomechanical analysis

the entropy density. The first law of thermodynamics, as expressed in Eq.(3.2),

can be rewritten as
ρT η̇ = σ v .ε̇ + β.ξ̇ + ρe
h − ∇.h, (3.3)

where we have introduced the definitions

∂e ∂e ∂e
σv = σ − σe, σe = ρ , β = −ρ , T = . (3.4)
∂ε ∂ξ ∂η

σ e are viscous stresses, the variables β are forced conjugate to the internal vari-
ables, and T is the temperature. Similarly, the free energy Ψ may be calculated
using the partial legendre transform of the internal energy with respect to the
entropy, which is specified as a function Ψ = Ψ(ε, ξ, T ), i.e.,

Ψ(ε, ξ, T ) = inf [e(ε, ξ, η) − ηT ]. (3.5)

η

Instead of relations (3.4), we get the following when we use this definition:

∂Ψ ∂Ψ ∂Ψ
σe = ρ , β = −ρ , T = . (3.6)
∂ε ∂ξ ∂η

It is usually easier to remove the entropy from Eq. (3.3) by computing its rate as
a function of the free energy. Then, using the chain rule, we arrive at

∂ 2Ψ ∂ 2Ψ ∂ 2Ψ
ρT η̇ = −ρT ε̇ − ρT ξ̇ − ρT 2 Ṫ . (3.7)
∂T ∂ε ∂T ∂ξ ∂ T

We define these terms for future reference,

∂ 2Ψ ∂ 2Ψ ∂ 2Ψ
m = ρT , µ = ρT , cv = −ρT , (3.8)
∂T ∂ε ∂T ∂ξ ∂T 2

as a result, we can express the energy balance as

cv Ṫ = (σ v + m).ε̇ + (β + µ).ξ̇ + ρe
h − ∇.h. (3.9)

48
 3.2 Coupled thermomechanics

The power density is used to define structural heating as

H = −m.ε̇ − µ.ξ̇. (3.10)

The structural heating is occasionally replaced by a simpler equation that is merely

proportional to the power expended, say, by plastic deformations.

3.2 Coupled thermomechanics

3.2.1 Small strain thermomechanics

Thermomechanics is a theory that predicts the mechanical response of de-

formable solids when it is coupled with thermal effects such as thermal expansion,
etc., but also when dissipative mechanisms that may occur during deformation
generate heat that affects the body’s thermal field. We can conclude from this
description that the displacement and temperature increment fields, u and T , are
the unknowns in a thermomechanics problem. These two fields can be derived
using the linear momentum and energy balance equations, i.e.

∇ · σ + ρb = ρa, (3.11a)
cṪ = ∇ · h + ρȟ, (3.11b)

where σ is the stress tensor, ρ the density, b the body force, a = ü the acceleration,
c the heat capacity, h the heat flux, and ȟ the supplied heat per unit volume. The
heat supply h is determined by thermomechanical phenomena occurring at the
material point level.

ρȟ = (σ v + m) · ε̇ + (β + µ) · ξ̇. (3.12)

In Eqs. (3.11), the coupling between the two unknown fields u and T are not
obvious. Let ε = ∇s u represent the infinitesimal strain tensor and ξ an array of
the material’s internal variables. The stress tensor will thus have a constitutive

49
3.2 Coupled thermomechanics

relationship with these two objects as well as the temperature increment. Similarly,
let β represent the dissipative forces conjugate to the internal variables. Then the
product
∆ ≡ σ v · ε̇ + β · ξ̇, (3.13)

is dissipated power in the deformation and it must be non-negative.

3.2.2 Small strain thermomechanical materials

When a material’s mechanical response and thermal state are coupled, the
material is said to be thermomechanical. This coupling is often only taken into
account in one direction: temperature increment from a reference state creates
strains, which may then lead to stresses in deformable bodies. In this chapter, we
focus on strongly coupled thermomechanical materials that exhibit the reciprocal
phenomena, which is that increases in strain cause increase in temperature.

As a result, the state of a thermomechanical material point is uniquely deter-

mined by its present strain ε, absolute temperature T , and, possibly, a collection
of internal variables defined collectively as ξ. Based on the data provided by any
of its thermodynamic potentials, all the pertinent information from its state may
be gathered at a material point. For instance, the free energy is the functional

Ψ = Ψ(ε, T, ξ). (3.14)

According to a standard thermodynamic analysis, the stress σ, the entropy η, the

tensor of elasticities C, the stress-temperature tensor m, and the heat capacity cv
at constant volume are defined, respectively, as

∂Ψ ∂Ψ ∂ 2Ψ ∂ 2Ψ ∂ 2Ψ
σ= , η=− , C= , m= , cv = −T .
∂ε ∂T ∂ε2 ∂ε∂T ∂T 2
(3.15)

50
 3.2 Coupled thermomechanics

3.2.3 Thermoelasticity

The simplest type of thermomechanical materials is the thermoelastic one. In

this material model, a point’s state is solely described in terms of its strain ε and
absolute temperature T . Thus, the free energy has a simple expression:

1 cv
Ψ = ε.Cε + (T − T0 )m.ε − (T − T0 )2 − η0 (T − T0 ), (3.16)
2 2T0

here, T0 is the reference temperature, η0 is the reference entropy, and C is a fourth-

order tensor of elasticities. The coupling of the thermal and mechanical effects is
expressed by the second order tensor m, which must be symmetric. The tensors
C and m for isotropic thermoelasticity have the following form:

C = κI ⊗ I + 2µIdev , m = −3καI, (3.17)

where κ, µ, and α are scalars and correspond, respectively, to the bulk modulus, the
shear modulus, and the coefficient of thermal expansion. Applying the equations
(3.15) to thermoelastic materials, it provides a simple form of the free energy in
such materials as

σ = Cε + (T − T0 )m, η = −m.ε, (3.18)

and we recover the definitions of C, m, and cv . When it comes to isotropic, linear

thermoelasticity, we may easily obtain

σ = κtr[ε]I + 2µdev[ε] − 3ακ(T − T0 )I, η = 3ακtr[ε]. (3.19)

3.2.4 General thermomechanical materials

The free energy provides all the data necessary to determine the stress, elas-
ticities, entropy, etc. at a material point of a thermoelastic material, as discussed
in section 3.2.3. Kinetic equations are necessary to explain the evolution of the

51
3.2 Coupled thermomechanics

internal variables, as in the isothermal case. This suggests that for each type of
material, a distinct formulation and solution procedure should be used.

However, it turns out that we can reuse a lot of the results from previous
sections to derive the response of fully coupled inelastic material points with a
certain ease. In order to do this, let’s look at materials whose free energy function
is represented as

Ψ(ε, T, ξ) = Ψmec (ε, ξ) + Ψcou (ε, T ) + Ψth (T ). (3.20)

The energy terms Ψmec , Ψcou , and Ψth correspond to the purely mechanical contri-
butions to the free energy, the coupling energy, and purely thermal energy, respec-
tively. Materials in which the constants on the inelastic part of the mechanical
response depending on temperature are disqualified by this additive decomposition
(excluding, for example, materials with thermal plastic softening).

The state of a thermomechanical material point consists of the variables {ε,

ξ, T }. Given its value at a given instant, the rate ε̇ and the gradient ∇T , the
material point needs to evaluate its equilibrium properties, the rate of the internal
variables, and the heat flux.

Equilibrium properties

The properties that are evaluated using the free energy or its derivatives are
those that are in equilibrium, namely,

∂Ψ ∂ 2Ψ ∂ 2Ψ ∂ 2Ψ
Ψ = Ψ(ε, ξ, T ), σ= , C= , m= , cv = −T .
∂ε ∂ε2 ∂ε∂T ∂T 2
(3.21)
To evaluate all of these, only the current state is required.

52
 3.2 Coupled thermomechanics

Kinetic equations

We can calculate the time-changing fields in the body using kinetic equations.
The heat flux, for example, is of the form

h = −κ∇T, (3.22)

and can be calculated form the tensor of conductivities and the temperature gra-
dient. In the case that the material is thermally isotropic, this last tensor is a
multiple of the identity. In general, we can define a thermal kinetic potential

1
X th (∇T ) = ∇T.κ∇T, (3.23)
2

so that
∂X th
h=− . (3.24)
∂∇T
The dissipation ∆ can be expressed also in terms of dissipation potentials. Let
X v (σ v ) and X i (β) be a viscous potential and an internal dissipation potential,
respectively and X v∗ , X i∗ their Legendre transforms. We may then postulate the
kinetic relations
∂X v∗ ∂X i∗
σv = , β= , (3.25)
∂ ε̇ ∂ ξ̇
in such a way that the dissipation is

d
∆ = σ v .ε̇ + β.ξ̇ = ∂ε̇ X v∗ .ε̇ + ∂ξ̇ X i∗ .ξ̇ = (X v∗ + X i∗ ). (3.26)
dt

The evolution of each class of constitutive models (thermoelastic, thermoplas-

tic, thermoviscoelastic, etc.) is furnished with a different set of internal variables
ξ and kinetic potential X i (ξ̇).

53
3.3 Results

3.3 Results

Figure 3.1 illustrates the results of a thermomechanical simulation during the

heating stage in which G-code file has been used to define the laser path. Fig-
ure 3.1(a) and (b) show the temperature field and different states of material,
respectively. Figure 3.1(c) shows the Von Mises stress which is only defined for
solid state of the material. Also, comparing Figure 3.1(a) and (c) shows that the
stress in the areas with lower temperature is higher than the hotter zones. This is
due to the fact that the residual stress in the fluid is zero and by solidifying the
fluid, the residual stress begin to grow and by decreasing the temperature it will
increase in the solid part of the material. Figure 3.1(d) shows the norm of the
strain tensor in the different states of material and we can see that the strain in
the powders around the solid part is higher than other areas. It is important to
note that in this study the strain resulting from the volumetric change during the
phase transformation is considered to be zero [24], [96].

Figure 3.1: Results of thermomechanical simulation. (a) temperature field (b)

States (including powder, solid, and fluid) (c) Von Mises stress tensor (d) norm
of strain tensor

54
Chapter 4
Multiscale Modeling of SLM

• S.M. Elahi, R. Tavakoli, A.K. Boukellal, T. Isensee, I. Romero, D. Tour-

ret, Multiscale simulation of powder-bed fusion processing of metallic alloys,
Computational Materials Science 209 (2022) 111383.

The multiscale simulation strategy described here is based on three main ele-
ments, namely: (1) computational thermodynamics (CalPhaD) to calculate temperature-
dependent properties of complex alloys and their phase diagram, (2) macroscale
(FE) thermal simulation of the powder-bed fusion, and (3) microscale (PF) simula-
tion of microstructure development by crystal growth within the melt pool. Macro-
and microscale simulations are coupled through the temperature field, estimated
within a two-dimensional slice in the vicinity of the melt pool, and used as input to
the phase-field simulations. The latter are upscaled using massive parallelization
on Graphics Processing Units (GPUs) in order to make them applicable at the full
melt pool scale, without compromising their accuracy, i.e. retaining the required
grid size of a few nanometers. Figure 4.1 schematizes the principal elements of
the resulting computational framework and details are presented in the following
subsections.

55
4.1 Computational thermodynamics of multicomponent alloy

Figure 4.1: Principal elements of the multiscale modeling approach, namely:

CalPhaD thermodynamics, Finite Elements thermal simulation of SLM pro-
cessing, and Phase-Field solidification simulation at melt pool scale.

4.1 Computational thermodynamics of multicomponent

alloy

The CalPhaD (CALculation of PHAse Diagrams) approach is used to compute

the thermodynamic parameters used in the FE and PF simulations. The method
is based on databases that describe the free energies of various phases, and it
minimizes the free energy globally. In order to calculate thermodynamic equilibria,
or phase diagrams, for complicated multicomponent alloys and then examine the
effects of individual alloying element variation, extensive datasets are required.

4.1.1 Temperature-dependent alloy properties

Lever rule solidification path was used to compute alloy properties as a function
of temperature, although the process would also work with other assumptions, such
as Gulliver-Scheil [82]. In particular, we calculated the temperature-dependent
density, ρ(T ), as well as the enthalpy per unit mass, h(T ), across a temperature

56
 4.1 Computational thermodynamics of multicomponent alloy

range spanning all phase transformation, here from 250 K to 5000 K. Using the lat-
ter, we were able to determine an effective heat capacity cp (T ) = ∂h(T )/∂T that
took into account both the actual heat capacity and enthalpies (latent heat) of
transformations occurring within the required temperature ranges. The key trans-
formation temperatures, namely liquidus (TL ), solidus (TS ), and boiling (TV ), may
also be derived from the tabulation of the fraction of liquid versus temperature.

We specifically use ThermoCalc and the TCNI8 Ni-alloys database to apply

the approach to the multicomponent Inconel 718 (IN718) superalloy, whose com-
position is shown in Table 4.1. The resulting and used ρ(T ) and h(T ) in the FE
calculations are approximated by piecewise linear functions in the interest of com-
putational efficiency (see Section 4.6.1). Any additional temperature-dependent
property from other sources, such as the thermal conductivity used here as an
instance from the literature [83], may be added to CalPhaD-calculated values.

Table 4.1: Chemical composition of superalloy Inconel 718

Element Ni Cr Nb Mo Ti Al
Weight % 50.0 17.0 4.75 2.8 0.65 0.2

4.1.2 Phase diagram and pseudo-binary approximation

For the PF simulations, phase diagram features are also computed using Cal-
PhaD. We employ a reference quantitative phase-field model for binary alloy so-
lidification, as detailed in Section 4.4. Thus, a pseudo-binary Ni-5wt%Nb is used
to represent the alloy, but the coupling technique may be easily expanded without
losing generality to other models, such as employing models for multicomponent
[97]–[99], and/or integrating rapid solidification effects [100]–[105]. Solute trapping
is intentionally left out of the current scope, such that the solid-liquid interface
is assumed to remain close to thermodynamic equilibrium, with local deviation
from the phase diagram solely due to solute and curvature effects. Therefore, ap-
plications of the methodology are limited to a scan velocity at which the solute
trapping is negligible in order to keep simulations quantitative (Section 4.5). We

57
4.1 Computational thermodynamics of multicomponent alloy

pay particular attention to Niobium because it is one of the solute elements in

IN718 that exhibit the highest segregation during solidification, which is crucial
for the formation of secondary intermetallic phases as well as in the hot cracking
susceptibility of highly segregated grain boundaries.

A multicomponent alloy can be approximated as a pseudo-binary system in

a variety of ways. The simplest approach is to directly take into account the
binary Ni-Nb system at the relevant Nb concentration (see, e.g., [74]). However,
this may cause a significant mismatch between FE (full alloy) and PF (pseudo-
binary) estimates in terms of transformation temperatures (e.g. TL , TS ) when
employing thermophysical properties of the full alloy. Instead, we are attempting
to match the liquidus temperature TL , as well as the solute partitioning of Nb
at the solid-liquid interface for a temperature that is close to TL . To achieve
this, we compute the thermodynamic equilibrium of the whole alloy at T = TL
using CalPhaD (ThermoCalc, TCNI8) (Table 4.1). At this point, we calculate the
liquidus slope with respect to the Nb concentration, m = ∂T / ∂c, and the solute
partition coefficient of Niobium, k = cs /cl , where cs and cl are the concentrations
(weight) of Nb in the solid and liquid phases, respectively. Then, assuming a
linearized phase diagram, we compute the appropriate (fictitious) pure solvent
′
melting temperature, TM , and the solidus temperature, TS′ , using TL = TM
′
+mc∞
′
and TS′ =TM +(m/k)c∞ , with c∞ =5wt%Nb the nominal alloy solute concentration
′
taken into account in the pseudo-binary approximation. (Prime symbols on TM and
TS′ here indicate this approximate pseudo-binary representation.) As a result, this
′
approach may result in some difference in terms of the linearized TM and TS′ (see
Section 4.6.1), but it has the large advantage of retaining an accurate description
of solute partitioning and interface equilibrium in the vicinity of T =TL as well
as a matching of temperature TL . Other parameters, such as the liquid solute
diffusivity or the solid-liquid interface Gibbs-Thomson coefficient, are extracted or
calculated from literature data (see Section 4.5.3 and Table 4.3).

58
 4.2 Macroscopic thermal simulations

4.2 Macroscopic thermal simulations

We consider the thermal problem of a moving heat source above a layer of

powder. It is anticipated that ignoring mechanics, particularly fluid flow, would
have a significant impact. The temperature profile in and around the melt pool is
predicted to be reasonably estimated by a thermal model for printing in conduction
mode, nevertheless. Chapter 2 provides detailed information regarding the laser
thermal input, Eq. (2.1), boundary conditions, Eq. (2.3), material addition and
state change, Eq. (2.4), and thermal solver, Eq. (2.5).

4.2.1 Thermal problem

As explained in detail in section 2.5, the governing equation for the heat transfer
problem can be stated as assuming a bounded domain Ω in R3 .

ρ(T ) cp (T ) Ṫ (x) = ∇ · κ(T )∇T , x ∈ Ω, (4.1)

where T stands for temperature, ρ for density, cp for specific heat at constant
pressure, and κ for the material’s scalar (i.e. isotropic) conductivity. ρ, cp , and
κ are material properties that depend on temperature and are hence a function
of location x and time t. The laser beam is represented as a 2D heat flux that
is applied to the top surface of the powder bed and Eq. (4.1) excludes a volumic
heat source term. The heat input is assumed to be symmetric with respect to the
laser beam axis and the total irradiance follows a standard Gaussian model.

2r2
 
2AP
I= exp − 2 , (4.2)
πw2 w

where A is the absorptivity of the powder bed, P is the laser power, w is the laser
beam diameter, and r is the radial distance from the center of the laser beam.

59
4.2 Macroscopic thermal simulations

Convection and radiation are taken into account along external boundaries.
The model for the convective heat transfer is

qc (T ) = hconv (T − Text ), (4.3)

where hconv is the convective heat transfer coefficient and Text is the temperature
of the surrounding environment. The Stefan-Boltzmann law is used to simulate
radiation as a significant heat loss mechanism because of the high temperatures
reached during laser melting. The radiative heat flux is expressed as

qr (T ) = σϵ(T 4 − Text
4
), (4.4)

where σ is the Stefan-Boltzmann constant and ϵ is the material emissivity. Finally,

a Dirichlet boundary condition of the form T = Text is set at the bottom of the
substrate, where a substrate material is assumed to exchange heat via conduction
with the powder bed, and the solidified material.

4.2.2 Numerical implementation

The above-described thermal model was numerically solved in three dimensions

using an in-house finite element code (IRIS) and a material library (MUESLI
[80]) implemented in C++ programming language. A Galerkin method employing
hexahedral elements was used to discretize the initial boundary value problem
in space. The resulting semidiscrete equations were integrated in time with an
implicit Backward-Euler method [81].

All layers’ initial elements, including those that will be activated later, are
created. Each time new layers are deposited, they gradually become active. A
latency time in between each successive layer activation is considered that either
indicates the period for applying new powder or allowing enough time for the
substance to cool to room temperature.

60
 4.3 Macroscale to microscale coupling

4.3 Macroscale to microscale coupling

The kinetics of microstructural development within the melt pool is typically

limited by species diffusion since the thermal diffusivity of metals is several orders
of magnitude higher than the solute diffusivities. So that we can decouple the com-
putation of the temperature field from that of the solidification within the melt
pool, we make the commonly held assumption that the phase change has little
impact on the temperature field. This results in a one-way coupling via the tem-
perature field, which is calculated by FE and then imposed in the PF simulation.
The resulting PF simulations (see Section 4.4) are rigorously analogous to those
using the classical one-dimensional frozen temperature approximation [106], but
imposing a different temperature field T (x, t). Along the laser path, we extract
this 2D temperature field from the sample’s longitudinal section. We approximate
the temperature field analytically for simplicity and computational efficiency. The
selected expression seeks to provide an approximate but appropriate description
of the temperature field in the area between liquidus and solidus temperatures,
which is important for the evolution of the microstructure. Hence, we approxi-
mate solidus and liquidus isotherms as two ellipses
s
(lL dL )2
rL (θ) =
2
2 , (4.5)
dL cos(θ) + lL sin(θ)
s
(lS dS )2
rS (θ) =
2
2 , (4.6)
dS cos(θ) + lS sin(θ)

where rL and rS are the respective radii of the T = TL and T = TS ellipses as a

function of the angle from the top surface

y − y0
θ = tan−1 , (4.7)
x − x0

with (x0 , y0 ) the center of the ellipses (see Figure 4.2a). The deepest point of the
melt pool is indicated here by x0 , which may be slightly shifted from the center
of the heat source. Melt pool dimensions appear explicitly in Eqs (4.5) and (4.6)

61
4.3 Macroscale to microscale coupling

in the form of the length (lL , lS ) and depth (dL , dS ) of the corresponding solidus
(subscript S) and liquidus (L) isotherms. The resulting temperature field is then
interpolated linearly between TL and TS , i.e. between rL (θ) and rS (θ) at a given
θ, as (Figure 4.2b)

r − rL (θ)
T (r, θ) = TL + (T0 − TL ) . (4.8)
rS (θ) − rL (θ)

Figure 4.2: Schematics of the elliptic temperature field approximation.

This equation has the advantage of only requiring four variables that may
be changed (lL , lS , dL , and dS ), which can be measured directly and therefore
monitored automatically from the FE results. The deepest points of the TS and
TL isotherms are considered to be aligned on the same x0 . However, this is often
the case since the two isotherms tend to be close to each other at the bottom of
the melt pool (see Figs 4.4, 4.5).

Although we only use a steady temperature profile moving at a constant speed

to show the method for the longitudinal cross section, it may be applied to any
cross section or temperature field, including time-dependent temperature fields.
Furthermore, it is simple to substitute the proposed analytical equation with an
effective interpolating method that directly estimates T (x, t) from the FE results.

62
 4.4 Microscale modeling of microstructure growth in the melt pool

4.4 Microscale modeling of microstructure growth in

the melt pool

4.4.1 Phase-field model

We consider a classical quantitative PF model for the solidification of dilute

binary alloys [106]. We apply non-linear preconditioning of the PF equation to
reduce the sensitivity of the results to the spatial grid size [107]. Final PF equations
in two spatial dimensions take the following form [54]:

√
   
T − T0 2 ∂ψ
 2
 2 2 2
1− as (α) = ∇ as (α) · ∇ψ + as (α) ∇ ψ − ϕ 2 |∇ψ| −
mc0l ∂t
   
∂ ′ ∂ψ ∂ ′ ∂ψ
− as (α) as (α) + as (α) as (α)
∂x ∂y ∂y ∂x
√ √
 
T − T0
+ 2 ϕ − 2 λ(1 − ϕ2 ) U + , (4.9)
mc0l (1 − k)

(1 − ϕ2 ) ∂ψ ∇ψ
   
1+k 1−k ∂U 1−ϕ
− ϕ = ∇ · D̃ ∇U + [1 + (1 − k) U ]
2 2 ∂t 2 4 ∂t |∇ψ|
2
(1 − ϕ ) ∂ψ
+ [1 + (1 − k) U ] √ , (4.10)
2 2 ∂t

where T is the temperature field, ϕ is the classical phase-field variable (+1 in

the solid and −1 in the liquid), ψ is the preconditioned phase-field variable with
√


ϕ(x, y, t) = tanh ψ(x, y, t)/ 2 , α = arctan ∂y ψ/∂x ψ is the angle between the
1 2 c/c0
solid-liquid interface normal and a fixed reference direction, U = 1−k l
1 − ϕ + k (1 + ϕ )
−


1 is the dimensionless solute supersaturation, with c the solute concentration field,
c0l = c∞ /k the solute concentration of a flat interface at the reference (solidus) tem-
perature T0 for an alloy of nominal solute concentration c∞ , k is the interface solute
partition coefficient and m is the slope of the liquidus line. In Eqs (4.9)-(4.10),
space is scaled in units of the diffuse interface width, W , and time is in units of the
relaxation time, τ0 , at the temperature T0 [106]. Considering that interpolation

63
4.4 Microscale modeling of microstructure growth in the melt pool

functions used in Eqs (4.9) and (4.10) are determined based on the thin-interface
asymptotic analysis [106], [108], their solutions will remain quantitative while using
W much larger than the capillarity length. The capillarity length, p0 , is expressed


at T0 as p0 = Γ/ |m|c∞ (1/k − 1) , where Γ denotes the Gibbs-Thomson coefficient
of the solid-liquid interface. The non-dimensional value for the liquid diffusion
coefficient, D̃, and the coupling factor, λ, are computed according to

Dτ0 W
D̃ = 2
= a1 a2 , (4.11)
W p0

W
λ = a1 , (4.12)
p0
where D is the liquid diffusion coefficient (Eq. (4.10) neglects diffusion in the solid
√
phase), a1 = 5 2/8, and a2 = 47/75. The standard form of the fourfold anisotropy
of the surface tension γ(ᾱ) = γ̄as (ᾱ) is considered with

as (ᾱ) = 1 + ε4 cos (4ᾱ), (4.13)

where γ̄ is the average surface tension in a (100) plane, ε4 is the strength of the
surface tension anisotropy and ᾱ is the angle between the normal to the interface
and a fixed crystalline axis. For a crystal misorientation α0 with respect to the
coordinate axes, the anisotropy as a function of the angle α between the interface
normal, and the axis x is

as (α) = 1 + ε4 cos 4 (α − α0 ) . (4.14)

Kinetic undercooling is also ignored, such that τ0 can be computed as

W2
τ 0 = a2 λ . (4.15)
D

The only model parameter that should be properly selected for the purpose of
making quantitative predictions is W . The solid-liquid interface is assumed to
be at local equilibrium under the current PF model, which is noteworthy. It is

64
 4.4 Microscale modeling of microstructure growth in the melt pool

therefore restricted to a modest velocity range toward the lower velocity range
relevant to SLM and is rigorously valid in a regime for which the solute trapping
effect can be neglected (see Section 4.5.1).

4.4.2 Polycrystalline solidification

Our goal is to simulate the epitaxial growth and grain growth competition of
columnar grains with different crystal orientations in the melt pool, relevant to
the process conditions studied here (see Section 4.6.2). A straightforward method
for modeling bi-crystal grain growth competition [54] is directly extended to poly-
crystals. The index of grains, which has a value of −1 in the liquid and a positive
or zero integer value in each grain, is stored in an integer field. Each index maps
to a given orientation. Here, we only consider 90 grain orientations for the sake of
simplicity, such that the solid grain index is taken within the [0, 89] range. This
index is directly used as the grain orientation value in degrees. When 1 − ϕ2 ex-
ceeds a specified limit, here fixed at 0.01, the grain index in the liquid is changed
to the index value that is most frequently present in the immediate grid point
neighborhood. This technique ensures that the correct equation is solved close to
the interface by creating a narrow halo of orientation field in the liquid surrounding
a grain. The solid-solid grain boundary will remain “frozen” when a grain index is
assigned to a grid point, as the crystal index field no longer evolves. As a result,
the evolution of solid-state grain boundaries is not taken into consideration by
this technique. Yet, it remains adequate to study grain growth competition for
well-developed dendrites, since triple points and grain boundaries are relatively
deep, and the region of interest remains close to the primary tips. The approach
has the key advantage of reducing directly to a reference, thoroughly validated,
quantitative PF model at the solid-liquid interface, while being computationally
efficient compared to a model using multiple order parameters (e.g. [109]).

65
4.4 Microscale modeling of microstructure growth in the melt pool

4.4.3 Numerical implementation

Equations (4.9) and (4.10) are solved in 2D using an Euler explicit time method
with a constant time step of ∆t on a finite-difference grid of square elements
with a grid spacing of ∆x. Based on the stability of the Laplacian operators in
Eqs (4.9) and (4.10), the time step size is chosen to be 0.3 times the maximum
time step. Laplacian operators are discretized using the typical second-order five-
point stencil. Other terms in equations (4.9) and (4.10) are discretized using
central difference schemes; for further information, see appendices A and B of the
reference [54]. The anisotropy terms and anti-trapping current are only computed
where |1 − ϕ2 | ≥ 10−6 , i.e. close to the solid-liquid interface, in order to minimize
the computing cost in the bulk phases away from the interface. If not, these terms
are set to zero.

For both PDEs, homogeneous Neumann (no-flux) boundary conditions are ap-
plied to every domain boundary. The signed distance function to the liquidus
isotherm with positive values in the solid phase is the initialization for the phase-
field ψ. The dimensionless supersaturation field is initialized based on the equi-
librium concentration over the entire domain, i.e. U = −1. The grain index field
is initialized to −1 in the liquid region (ψ < 0) and it is initialized to a Voronoi-
based distribution of grain indices in the solid. To do so, we randomly generate
N Voronoi cell centers in the entire domain using fast Poisson disk sampling [110]
with a random grain index within [0 − 89]. Grain indices are then allocated to
each finite difference grid point using a classical Voronoi tessellation algorithm.

To include the whole melt pool tail, the simulation domain is slightly bigger
than lS × dS . We use a standard moving frame algorithm to calculate a grain map
for a solidified length longer than the simulation domain. New grid points at the
alloy nominal concentration are added on the right side of the domain, and values
of the fields at grid points leaving the simulation domain are stored to be used
later to reconstruct the grain map for the entire solidified length (see Figure 4.8).

We employ ∆x = 0.8W , where ∆x is calculated using a convergence study

that monitors the steady-state tip undercooling as a function of the grid size in

66
 4.4 Microscale modeling of microstructure growth in the melt pool

unidirectional solidification while taking into account the conditions that are the
most computationally demanding, namely the (lowest) temperature gradient and
the growth velocity at the tail end of the melt pool. Due to the small dendrite tip
radius and diffusion length under additive manufacturing conditions, this conver-
gence study can be quite limiting (see section 4.5.3), but it is still essential if the
objective is to quantitatively predict the growth kinetics of dendrite/cells and the
resulting grain structures. The grid size requirement for quantitative predictions
makes even 2D PF simulations at the scale of the melt pool very computationally
demanding. As a result, advanced acceleration schemes are necessary, which are
briefly discussed below.

Using the computer unified device architecture (CUDA) programming lan-

guage, we implemented the model for parallel computing on multi-graphic pro-
cessing units (Multi-GPU). Even though simulations are massively parallelized,
we also aim at providing a solution that can be implemented on medium-size com-
puting hardware accessible in-house to most research laboratories or companies.
For this reason, we limit the current study to simulations performed on a single
cluster node with eight Nvidia GPUs (RTX 2080Ti). The computational domain
is partitioned into 8 almost equal layers along the y-direction with a similar (total)
number of grid points in the x-direction using a simple layer-wise domain decom-
position. We consider one extra halo layer of points at the top and bottom rows
of each domain to simplify the imposition of boundary conditions and inter-GPUs
data exchange. The ψ-field calculation at the next time step and the U -field cal-
culation at the next time step are the two main kernel calls that compose the time
loop. Then, the time stepping is achieved by swapping pointer addresses between
arrays containing values at the current time step and arrays containing values at
the next time step for ψ and U . Direct GPU-GPU communication is used to up-
date the halo grid data following the execution of each kernel (memory copy from
one GPU to its neighbor GPUs). We avoid using expensive GPU-to-CPU and
CPU-to-GPU data transmission in this way.

67
4.5 Application to selective laser melting of inconel 718 alloy

4.5 Application to selective laser melting of inconel 718

alloy

4.5.1 Processing conditions

We illustrate the methods described above and their coupling with the sim-
ulation of Selective Laser Melting of IN718 alloy (Table 4.1). It would not be
suitable to take into consideration a laser scan velocity comparable to or higher
than the onset velocity for solute trapping because we chose a PF model that does
not take solute trapping into account. We use a scan velocity of V = 0.1m/s,
which should stay sufficiently below the onset of significant solute trapping, with a
typical onset of solute trapping at solidification velocities on the order of ≈ 1m/s.
In Section 4.6.3, this assumption is further examined. We considered a laser power
of P = 100W, which corresponds to a linear energy density of P/V = 1.0 J/mm
under real SLM conditions for Inconel 718 [76].

4.5.2 Thermal simulations and parameters (FE)

Table 4.2 provides a summary of the variables used in macroscopic FE simu-

lations. CalPhaD is used to determine the most of material properties, including
TL , TS , ρ(T ), and cp (T ). The thermal conductivity κ(T ) is extracted from [83].
The laser absorption, convection, and emissivity coefficients have been selected
according to [111]. The view factor is determined in accordance with [88], and the
average particle diameter is chosen to be equal to the layer thickness. By monitor-
ing the dimensions of the melt pool (lS , lL , dS , and dL ), we were able to ensure the
numerical convergence of the simulation with respect to the grid element size, the
number of deposited layers, and the domain size. We discovered that in order to
obtain a steady melt pool size and shape, we needed to consider (1) a minimum of
two grid elements within the powder layer thickness hp = 30 µm, and (2) at least
five or six powder layers on top of the substrate, taking into account successive
heating and cooling stages, using a cooling time of 0.166 s between layers, which

68
 4.5 Application to selective laser melting of inconel 718 alloy

Table 4.2: Material properties and process parameters for FE simulations.

Properties Symbol Value Unit
Solidus temp TS 1554 K
Liquidus temp TL 1625 K
Bulk (dense) Boiling temp TV 3038 K
material Heat capacity cp Fig 4.3b J.kg−1 .K−1
Density ρ Fig 4.3c kg.m−3
Thermal conductivity κ Fig 4.3d W.m−1 .K−1
Average diameter d 30 µm
Powder bed Porosity ξ 0.3 -
View factor F 0.33 -
Gas (Argon) Density ρg 1.66 kg.m3
Heat capacity cp 677 J.kg−1 .K−1
Substrate
Density ρ 7900 kg.m−3
(Stainless steel)
Thermal conductivity κ 24.9 W.m−1 .K−1
Laser Power P 100 W
Scan speed V 0.1 m.s−1
Process
Beam diameter w 70 µm
Thickness of layer hp 30 µm
Absorption coefficient A 0.55 -
Boundary Convection coefficient hconv 15 W.m−2 .K−1
conditions Emissivity ϵ 0.3 -
External temperature Text 273 K

was found to be sufficient for the part to cool down close to the external tempera-
ture. Additionally, we determined that a domain size of 0.21 × 1.2 × 0.6 mm3 was
sufficient to stabilize a steady melt pool. In order to model 10 powder layers, we
employed a grid element size of 15 µm, and used the steady temperature profile
within the tenth layer as the thermal field for the PF simulations.

Furthermore, we performed additional simulations with either constant conduc-

tivity or constant density, keeping all other parameters the same, in order to eval-
uate the effects of the temperature-dependent properties. For these simulations,
the conductivity or density of the dense material was estimated as 29 W.m−1 .K−1
or 7400 kg.m−3 at the liquidus temperature. To estimate the effect of the different
properties between the powder bed and dense material, we also performed an ad-

69
4.5 Application to selective laser melting of inconel 718 alloy

ditional simulation in which the thermal properties of the powder bed were set to
be equal to those of the bulk (dense) material.

4.5.3 Microstructure simulations and parameters (PF)

Table 4.3 lists the computational and physical parameters used in PF simu-
lations. According to Section 4.1, the liquidus slope and solute (Nb) partition
coefficient for the full IN718 alloy were calculated (Table 4.1), and the correspond-
′
ing pseudo-binary TM and T0 ≡ TS′ values for the PF model were calculated to
match the liquidus temperature TL of the full alloy with that of the pseudo-binary
approximation. Notably, the solidus temperature is taken into account in the
pseudo-binary PF simulation (TS′ = 1549K) is still rather close to the solidus of
the full alloy accounted for in the FE simulation (TS = 1554K) and calculated
using CalPhaD. Using ab initio molecular dynamics simulations, Walbrühl and
collaborators have estimated the diffusion coefficient of Nb in Ni (Ni-10at.%Nb)
between 1903 and 2303 K, and assessed Arrhenius prefactor D0 ≈ 1.22×10−7 m2 /s
and activation energy E ≈ 55.3 kJ.mol−1 .K−1 [112]. We use this expression to es-
timate the diffusion coefficient in the vicinity of 1700 K, which we use as constant
in the PF simulation with D ≈ 2.44 × 10−9 m2 /s. The Gibbs-Thomson coefficient
of the solid-liquid interface is calculated as Γ = γ0 TM /L ≈ 2.49 × 10−7 K.m, using
the melting temperature of pure Ni, TM = 1728 K, and the latent heat of fusion,
L = 2.08 × 109 J.m−3 , calculated from the ThermoCalc (TCNI8) calculation of
h(T ) for pure Ni, and an interface excess free energy γ0 ≈ 0.3 J.m−2 consistent
with those calculated for pure Ni with molecular dynamics (capillary fluctuation
method) in several references (from 0.27 to 0.36 J.m−2 in Refs [113]–[115]). We
use an anisotropy strength for the interface excess free energy of ε4 = 0.02, which
is of the same order as identified by these atomistic simulations (e.g. ε4 ≈ 0.018
in [113]), considering that here we only use the fourfold anisotropy component ac-
cording to Eq. (4.14). A detailed investigation of convergence for these parameters
(see Sections 4.4.3 and 4.6.3) showed that well-converged simulations required a
grid spacing of ∆x = 5 nm. This value is consistent with the relevant length scales

70
 4.6 Results and discussion

under these conditions, as discussed later in Section 4.6.3. Hence, considering a

simulation domain size of 250 µm × 100 µm, a simulated time of 5 ms, and a sta-
ble time step ∆t = 0.76 ns, this resulted in 50 000 × 20 000 grid points (i.e. over
2 billion degrees of freedom) and about 6.6 million iterations. Even though it is
arguably a large simulation, it was nevertheless possible to complete it in under
ten days with reasonable computing hardware (one compute node equipped with
eight Nvidia RTX 2080Ti GPUs).

Table 4.3: Material properties and process parameters for PF simulations.

Properties Symbol Value Unit
Nominal alloy concentration c∞ 5.0 wt% Nb
Solute partition coefficient k 0.37 -
Liquidus slope m 9.0 K.wt% Nb−1
Liquid diffusion coefficient Dl 2.44 × 10−9 m2 .s−1
′
Pseudo-binary solvent melting temperature TM 1670 K
Pseudo-binary reference temperature (Solidus) T0 1549 K
Gibbs-Thomson coefficient Γ 2.49 × 10−7 K.m
Interface anisotropy ε4 0.02 -
Grid element size ∆x 5 nm

4.6 Results and discussion

4.6.1 Temperature-dependent thermophysical properties

The temperature-dependent parameters of IN718 namely: enthalpy, heat ca-

pacity, and density are calculated in the first step of our methodology using the
CalPhaD method. Figure 4.3 illustrates the CalPhaD results (represented by or-
ange symbols) as well as the piecewise linear approximations for h(T ) (i.e., piece-
wise constant cp (T )) and ρ(T ) used in the FE simulations (purple solid lines). The
temperature-dependent conductivity (Figure 4.3d) was taken from [83]. The pow-
der bed density and conductivity (thick green lines) are estimated using Eqs (2.6)-
(2.7).

71
4.6 Results and discussion

Figure 4.3: Temperature-dependent material properties for alloy IN718: (a)

Enthalpy, (b) Heat capacity, (c) Density, (d) Conductivity [83]. CalPhaD calcu-
lated data (database TCNI8) appear as symbols while piecewise linear fits used
in FE simulation appear as lines (thin purple for bulk, thick green for powder
bed).

4.6.2 Macroscopic thermal field

According to the FE thermal simulation, Figure 4.4 shows the temperature

field during the tenth layer’s heating stage. The blue and red lines, respectively,
represent the solidus and liquidus isotherms. The resulting melt pool dimensions
are lL ≈ 185 µm, lS ≈ 249 µm, dL ≈ 88 µm, and dS ≈ 96 µm. Note that in
the PF simulations, we used approximate dimensions lL = 190 µm, lS = 245 µm,
dL = 90 µm, and dS = 95 µm. Figure 4.5a shows a comparison between the elliptic
approximation using Eqs (4.5)-(4.6) and used as input in the PF simulations (white
dashed lines) and the tail half of the melt pool within the central longitudinal
section (solid lines and symbols). Figure 4.5b compares the temperature gradient
that results as a function of the angle (θ) from the domain’s top as calculated by
FE simulations along the TS and TL isotherms (symbols) and using the elliptic
approximation (solid green line). Using the elliptic approximation, the only term

72
 4.6 Results and discussion

Figure 4.4: Temperature distribution during heating stage of tenth layer as

predicted by finite element simulation.

that distinguishes the gradient measured along the solidus or along the liquidus
line is the polar component of the temperature gradient, (1/r)∂T /∂θ. However, its
magnitude remains below 104 K/m, which is negligible in comparison to the radial
component, ∂T /∂r, which is of the order 106 to 107 K/m. The curve in Figure
4.5b employs all terms along the TL isotherm, but the plots along TS and those
that just take the ∂T /∂r term into account are virtually identical. Consequently,
for an elliptically shaped melt pool, a reasonable estimation of the temperature
gradient can be conveniently obtained directly for the polar angle as ∂T /∂r =
(TL −TS )/(rL (θ)−rS (θ)), using rL (θ) and rS (θ) from Eqs (4.5)-(4.6). The analytical
function provides a reasonable approximation of the temperature field between TL
and TS , as illustrated in Figure 4.5, despite a little deviation.

Given that the temperature gradient is between 106 and 107 K/m and the
growth velocity Vgr (i.e. the solidification front velocity) is at most V = 0.1 m/s,
the ratio G2 /Vgr should be higher than 1013 K2 s/m3 , which should lead to the
epitaxial growth of columnar microstructure, consistently with the assumption

73
4.6 Results and discussion

Figure 4.5: Melt pool shape within the longitudinal section of Figure 4.4b,
comparing FE results and the elliptical approximation (Eq. (4.8)): (a) FE-
predicted temperature field (color background) and isotherms for T = TS =
1554 K (solid blue line and symbols), T = TL = 1625 K (solid red line and
symbols) and T = 1569, 1583, 1597, and 1611 K (black solid lines), compared
to elliptic approximation along the same temperatures (white dashed lines);
(b) temperature gradient as a function of the polar angle θ (see Figure 4.2)
as predicted by FE along the solidus (blue circle symbols) and liquidus (red
diamond symbols) isotherms compared to the elliptic approximation (solid green
line).

made in the PF simulation. Comparatively, Knapp et al. [84] computed that a

fully columnar structure should grow for any G2 /Vgr above 1.52 × 1011 K2 s/m3 for
Inconel 718 alloy, however considering electron beam melting.

On the top surface of the thermal field, Figure 4.4a, both isotherms show a
kink, most prominently for the solidus, in the tail end of the melt pool. This
feature results from the straight boundary of the bulk (dense) areas directly below
and behind the laser path from the material on the side of the path, which is
still in the powder state and, more importantly, has a significantly lower density
and conductivity (see Figure 4.3). Figure 4.6 shows the top surface isotherms
(solid lines) and the calculated boundary between powder bed and dense states

74
 4.6 Results and discussion

(dashed line) for the current simulation (a) as well as a simulation in which the
whole domain has the dense material’s thermophysical properties (b). Not only
does the kink in both isotherms disappear, but the size of the melt pool is also
decreased greatly in the latter case due to the easier transfer of heat down the walls
composed of dense, more conductive material. Finally, we examine the influence of

Figure 4.6: Top view of the liquidus and solidus isotherms (solid lines) and
computed boundary between powder bed and bulk (dense) states (dashed line)
when considering different thermal properties in dense and powder bed states
(a), and when the powder bed has the same properties as the bulk material (b).

temperature-dependent properties on the longitudinal melt pool shape (liquidus

and solidus isotherms). Figure 4.7 compares the reference results of Figure 4.4
(solid lines) to equivalent simulations (dashed lines) with constant conductivity (a),
constant density (b), and equal properties in the powder bed as in the bulk material
(c). A constant conductivity tends to shrink the melt pool, but a constant density
tends to slightly expand it. The most significant influence, as shown in Figure 4.7c,
appears to be the consideration of different powder bed and bulk properties. As
shown in Figure 4.6, this emphasizes the importance of accounting for the thermal
properties of the powder bed in order to get valid thermal simulations of powder-
bed fusion processes.

75
4.6 Results and discussion

Figure 4.7: Liquidus (red) and solidus (blue) isotherms in the central longitu-
dinal section for the reference simulation of Figure 4.4 (solid lines) compared to
equivalent simulations (dashed lines) with a constant conductivity (a), constant
density (b), or equal properties in the powder bed and bulk material (c).

4.6.3 Microstructure growth in the melt pool

Convergence analysis

The numerical convergence study of our PF simulations was carried out on

a reduced relevant problem containing a one-dimensional thermal field, i.e. uti-
lizing the classical frozen temperature approximation with a pulling velocity of
V = 0.1 m/s and a temperature gradient of G = 107 K/m. Measuring the steady
state dendrite tip undercooling obtained for several grid element sizes, we discov-
ered that results began to deviate significantly at ∆x ≈ 5 nm or above, establishing
the grid spacing required for quantitative simulations. Increasing the grid spac-
ing further promotes dendritic interaction and causes the formation of pockets of
highly segregated liquid between them, which resemble patterns reported in rapid
solidification experiments [74], [116] but tend to disappear when the discretization
is refined.

76
 4.6 Results and discussion

Relevant length scales

The value of ∆x ≈ 5 nm matches the most important physical length scales

in the melt pool solidification problem. Indeed, the two-dimensional one-sided tip
selection parameter is expected to be σ ∗ = 2Dp∗0 /(R2 V ) ≈ 0.15 for ϵ4 = 0.02
(see Figure 1 and Eq. (4.3) in [117]). With the alloy parameters in Table 4.3, an
approximate estimation of the steady-state growth leads to a steady dendrite tip
radius R ≈ 39.3 nm, with a capillary length p∗0 ≈ 4.75 nm at a dimensionless tip
undercooling ∆ ≈ 0.497, i.e. a Péclet number P = RV /(2D) ≈ 0.806 (whereas p0
at the solidus temperature T0 is close to 3.25 nm), and a diffusion length D/V =
24.4 nm. As a result, the grid element size ∆x = 5 nm is only eight times smaller
than the steady tip radius, five times smaller than the steady diffusion length, and
of the same order as the tip capillarity length. While only an approximate order-
of-magnitude analysis (none of the actual melt growth ever reaches steady state,
and the laser velocity V is only relevant to the melt pool’s tail region), this still
provides a reasonable explanation for why the grid element size cannot be taken
much coarser without compromising accuracy.

Solute trapping

The assumption of solid-liquid interface equilibrium with regard to solute trap-

ping is discussed here, as well as whether a laser velocity of V = 0.1 m/s is
low enough to maintain this assumption. The solute partitioning at the inter-
face changes with the interface velocity Vi like k(Vi ) = [ke + Vi /VD ]/[1 + Vi /VD ],
according to the continuous growth model (CGM) [118], [119], where ke is the
equilibrium partition coefficient and VD is the solute diffusion velocity through the
interface. One may approximate the diffusion velocity as VD ≈ 1.44 m/s using
the parameters of the pseudo-binary alloy (Table 4.3) and an order of magnitude
for the physical interface width lA ≈ 1 nm (see Eq.,(63) in [120]). Therefore, the
considered laser velocity V = 0.1 m/s, which is the highest growth velocity expe-
rienced in the melt pool at its tail end, seems to be sufficiently lower than VD for
solute trapping to remain negligible. Even if we consider V ≪ VD , the resulting

77
4.6 Results and discussion

change in partition coefficient from the CGM is only around 10%, which is still
not very significant but might be significant for higher V .

The diffuse interface width used here is W = 6.25 nm, which is sensibly higher
than the actual width of the solid-liquid interface. Should the required value of
W be further reduced for convergence, e.g. for higher V , it is worth noting that
using physically realistic diffuse interface width can lead to the prediction of solute
trapping effect in good agreement with the CGM [120].

Full melt pool simulation

Figure 4.8 shows the results of the 2D PF simulation of solidification at the

full melt pool size, illustrating the time-evolution (top to bottom) of the grain
structure (left) and solute (Nb) concentration field (right). The solute map pro-
vides a more in-depth understanding of the dendritic structures within the grains,
while the grain map displays the growth competition at the melt pool scale. While
this two-dimensional simulation alone is insufficient to draw statistically signifi-
cant conclusions on grain growth competition under conditions relevant to AM,
trends appear that highlight interesting similarities and differences with tradi-
tional, Bridgman-like, directional solidification (DS).

Similar to DS, dendritic elimination (impingement) and sidebranching events

determine the orientation of converging and diverging grain boundaries, respec-
tively [54], [55]. For example, the top right zoomed-in regions in Figure 4.8 il-
lustrate the competition between three grains for grain growth. The temperature
gradient’s amplitude and direction, as well as the local cooling rate, are continually
changing, in contrast to DS, in the region of each emerging grain boundary (GB).
This makes it challenging to identify with absolute certainty the “favorably” or
“unfavorably” oriented grains forming the GB, as these roles may switch during
the process.

These findings lead to two interesting conclusions. First, most of the grains
selected for the growth competition had a small forward tilt with respect to the
vertical direction. For most of grains, this corresponds to a principal dendritic

78
 4.6 Results and discussion

growth direction, i.e. a main crystalline orientation. However, some larger grains
persist via consecutive sidebranching despite a substantial misorientation of their
crystalline dendritic axes with this “mesoscopic” direction of maximum grain elon-
gation (see bottom right panel in Figure 4.8). The prevalence of tilted columnar
grains, regardless of their inner crystalline orientation, emphasizes the necessity of
the simulation at the full melt pool scale because these would not naturally emerge
from grain growth competition in a reduced subset of the melt pool.

Figure 4.8: Phase-field simulation results showing grain structure formation

(left) and solute (Nb) concentration field (right). Iso-temperature lines show
T = TL = 1625 K (red), T = 1550 K ≈ TS (blue), and intermediate temperature
with steps of 5 K (black). The simulation domain, moving at a velocity V ,
is delimited with dashed black lines. Zoomed-in regions at the top right are
marked with a dashed rectangle in the resulting full-scale maps at t = 1.0 ms
and t = 2.0 ms.

79
4.6 Results and discussion

Second, an apparent range of various primary dendritic spacing (PDAS) ranges

emerges (see, e.g., final, bottom-right, dendritic structure in Figure 4.8). This
PDAS heterogeneity can exist both within the same grain and among several
grains. Such examples may be seen in the zoomed-in regions (top right) of Fig-
ure 4.8, particularly in grains A and E. Grain A exhibits heterogeneity among
spacings that emerged directly from the initial almost-planar (slightly curved)
interface destabilization. It is possible to explain this heterogeneity by the quasi-
steady, but not quite steady, growth conditions, undercooling, and the fact that
spacing homogeneity within a grain can take a lot longer than it does to establish
a steady quasi-steady growth velocity [121]. On the other hand, grain E’s locally
smaller microstructural length scale is caused by the fact that the lower spacing
area emerges from sidebranching, thus forming a region with locally secondary
dendrite arm spacings (SDAS) along the diverging GB.

In Figure 4.9, we illustrate how these findings may be used for digital mi-
crostructure characterization. There, we present the grain structure (color back-

Figure 4.9: Solute (Nb) segregation profiles along different line scans (A–E)
in the solidified region as predicted by phase-field simulations. The dashed
line in the bottom plots marks the solute concentration of the eutectic point
(L→Nifcc +Ni3 Nb) in the Ni-Nb phase diagram.

80
 4.7 Conclusion and perspectives

ground) overlaid with the Nb concentration map (gray level), as well as line scans
of the Nb concentration in various melt pool regions (bottom plots). Such sig-
nals show two main remarks. First, the periodicity of the signals provides the
average primary spacing inside the grain, which may be easily extracted in a sys-
tematic manner using adapted spectral filtering techniques. Secondly, the extent
of interdendritic Nb segregation enables the identification of the region where the
formation of secondary phases is most likely to occur. Since this composition field
is reconstructed from still partially liquid regions (due to the moving frame algo-
rithm), it is appropriate to compare these segregation peaks to the eutectic triple
point (L→Nifcc +Ni3 Nb) at T ≈ 1295 ◦ C and c ≈ 21 wt%Nb in the Ni-Nb phase
diagram (according to ThermoCalc TCNI8 calculations), marked with a dashed
line in the bottom plots of Figure 4.9. As a result, this study suggests that the
intermetallic Ni3 Nb phase is most likely to occur in locations near the bottom of
the melt pool, such as region B.

4.7 Conclusion and perspectives

We built a multiscale modeling framework for the simulation of metallic al-

loy powder-bed fusion. The following methods are combined and coupled in the
framework:

• CalPhaD calculation of temperature-dependent properties and phase dia-

gram, enabling the study of alloy chemistry;

• 3D finite element thermal simulation of laser melting that considers different

properties in different locations and CalPhaD-based temperature-dependent
properties;

• 2D phase-field simulations of microstructure formation by polycrystalline

solidification in the melt pool.

The methodology was used to simulate selective laser melting of Inconel 718
superalloy. We highlighted the influence of temperature-dependent properties and

81
4.7 Conclusion and perspectives

the relevance of using distinct properties in the powder bed and dense areas to
forecast the size and shape of the melt pool. Finally, we used 2D quantitative
simulations at the scale of the entire melt pool to simulate the dynamical selection
of grain structure via polycrystalline growth competition, highlighting some key
similarities but also differences with equivalent simulations typically performed on
a reduced subset of the melt pool.

This research is arguably important in the context of Integrated Computational

Materials Engineering (ICME) for powder-bed fusion processes. It does, however,
have several restrictions, most of which are related to ongoing development work
and future directions.

The next stage in macroscopic simulations is a coupling with thermomechan-

ics, which includes fluid dynamics and plasticity. Fluid flow simulation would
allow for the prediction of defect generation [44], [46]–[49], but it would also al-
low for the extension of the approach to powder-bed melting in keyhole mode.
If a comparable level of accuracy is desired in the modeling of dendritic growth,
a complete two-way coupling between solidification and fluid flow in the liquid
would probably require a multiscale approach (e.g. concurrent grids or methods),
a computationally effective and scalable method for modeling the flow (e.g. the
Lattice Boltzmann Method [122]–[124]), and/or sophisticated algorithms for code
acceleration through parallelization [122], [125] and/or adaptive meshing [5], [126],
[127]. Another possibility is to assume a one-way coupling strategy by imposing
temperature and solute fields calculated by macroscale simulations ahead of the
solidification front in microscale simulations at a distance (larger than the diffusive
boundary layer yet smaller than the typical hydrodynamic length). It will also be
possible to forecast significant aspects of print quality, such as residual stresses
and part distortion, thanks to ongoing extensions to thermomechanics simulations
in the solid state [48]. The amount of information needed to correctly predict the
geometry of the melt pool may limit the modeling of all components, except lever-
aging advanced numerical techniques (e.g. adaptive meshing [5], [128]). On the
other hand, we believe that the existing physics-based approach should be scalable
for the simulation of “mesoscopic” representative volume elements.

82
 4.7 Conclusion and perspectives

The lack of solute trapping, the pseudo-binary alloy approach, and the lack of
solid-state microstructure development are the key limitations of PF simulations of
microstructure. It will be necessary to employ dedicated models for the extension
to multicomponent alloys or solute trapping (e.g. [97]–[99], [101], [102]). Notably,
approximate yet practical extensions of the current model have been developed
recently that provide some solute trapping matching CGM theory at growth veloc-
ities around VD [103], [104]. Using dedicated phase-field models (e.g. [129]–[133]),
solid-state microstructure evolution during heat treatments might be added asyn-
chronously. This evolution could be intrinsic (e.g. in the heat-affected zone) or
extrinsic (e.g. through aging). The rate of nucleation, which controls the size of
columnar/equiaxed grain structures, is another important aspect. Its introduction
usually makes use of phenomenological approaches (e.g. randomly seeding nuclei)
[134]. However, in order to provide accurate predictions, this would also introduce
new parameters (such as nuclei density and activation undercooling) that would
need to be properly calibrated.

Ultimately, due to the two-dimensional simulations, the pseudo-binary alloy ap-

proximation, and the fact that a statistical (high-throughput) exploration would
be necessary to extract statistically meaningful trends and conclusions, the result-
ing microstructure analysis proposed here remains semi-quantitative. However,
we believe that in terms of alloy design and process optimization in fusion-based
AM of metals, this kind of methodology provides ICME a viable way forward.
High-throughput statistical studies will help to address the pervasive issues of un-
certainty and repeatability in metal AM because of the level of microstructural
details, the few calibration factors, and the fact that such simulation is achievable
with reasonable computing resources.

83
84
Chapter 5
Comparing Phase-field and Cellular
Automaton Models for Simulation of
Microstructural Evolution During
Melt Pool Solidification

• S.M. Elahi, R. Tavakoli, I. Romero, D. Tourret, Grain growth competition

during melt pool solidification – Comparing phase-field and cellular automa-
ton models, Computational Materials Science, Accepted.

5.1 Methods

We consider material and laser parameters from the prior work (chapter 4, [79])
on the powder-bed fusion of Inconel 718 alloy with the aim of comparing PF and
CA simulations of melt pool solidification. Here, ThermoCalc’s TCNI8 database’s
CalPhaD algorithm is used to compute temperature-dependent alloy parameters,
such as heat capacity, density, and thermal conductivity, or given in experimental
literature (see section 4.6.1). Finite elements (FE) method is used to calculate

85
5.1 Methods

the steady-state thermal field in the melt pool region, as explained in detail in
section 4.2. The two-dimensional (2D) sections, longitudinal and transversal, are
extracted from the results of three-dimensional (3D) thermal field and used as the
temperature field in either PF or CA simulations. It’s important to note that we
neglect nucleation and assume that solidification only results from epitaxial growth
from the melt pool’s bottom. By doing this, we can focus on the growth competi-
tion and avoid adding any further adjustable parameters (e.g. nuclei density and
nucleation undercooling).

5.1.1 Thermal field

The temperature field in the melt pool region is extracted longitudinally (i.e.,
within a plane containing the position and route of the heat source) and cross-
sectionally (i.e., within a plane normal to to the scanning direction) based on the
results of the 3D FE. The temperature field, T (x, y, t), is then approximated using
simple analytical formulae that take an elliptical shape of the liquidus and solidus
isotherms into account.

Between the solidus (TS = 1554 K) and liquidus (TL = 1625 K) temperatures
is the most important region, where the temperature field needs to be sufficiently
defined, in terms of grain growth competition during solidification. Therefore,
regardless of the approximation accuracy below TS or above TL , the analytical
fitted functions are selected to provide a fair description of the time-dependent
location of (T = TL ) and (T = TS ), and consequently of the temperature gradient
in the freezing range.

Longitudinal section

The analytical approximation of the thermal field longitudinal section is similar

to that used and presented in section 4.3.

86
 5.1 Methods

Cross-section

In order to offer a decent match to the FE results, the cross-section tempera-

ture is also roughly estimated using elliptical solidus and liquidus isotherms, which
shrink non-linearly with time (schematized in Fig. 5.1). In this case, the temper-
ature is interpolated linearly between the temperature at the center of the melt
pool Tc (t) and the solidus temperature TS as

r
T (r, θ) = Tc (θ, t) + TS − Tc (θ, t) , (5.1)
rS (θ, t)

using the equation of an ellipsis indicating where the solidus temperature isotherm
is located
s
(wS (t) dS (t))2
rS (θ, t) = . (5.2)
(dS (t) cos(θ))2 + (wS (t) sin(θ))2

with
p
r= (x − x0 )2 + (y − y0 )2 , (5.3)
y − y0
θ = tan−1 . (5.4)
x − x0

Given below is a description of this isotherm’s time-dependent half-width wS

and depth dS


nw 
wS (t) = w0 1 − t/tf , (5.5)

nd 
dS (t) = d0 1 − t/tf , (5.6)

where the initial (i.e. maximum) half-width and depth of the melt pool are w0
and d0 , respectively, and the time tf is the point at which both wS and dS are
zero. The time evolution of wS and dS between (wS , wL ) = (w0 , d0 ) at t = 0 and
wS = wL = 0 at t = tf is parametrized using exponents nw and nd (linear evolution
if exponent equals 1, or late and steep evolution for higher exponents). Parameter

87
5.1 Methods

Figure 5.1: Schematics of analytical temperature approximation T (r, θ): (left)

location of elliptical solidus isotherm from the center (x0 , y0 ) of the melt pool,
(right) radial interpolation between center temperature Tc and solidus temper-
ature TS .

values w0 = 128 µm, d0 = 96 µm, tf = 0.00255 s, nw = 20, nd = 3 are adjusted

for a good fit to the FEM results (see Figure 5.2.a).

140
1630
120 1620
100 1610
1600 T(x0,y0) [FE]
80
T (K)

Eq.(5.8)
(μm)

1590
60 wS [FE] 1580 TL
dS [FE] 1570 TS
40
Eq.(5.5) 1560
20
Eq.(5.6) 1550
0 1540
0 0.0005 0.001 0.0015 0.002 0.0025 0 0.0005 0.001 0.0015 0.002 0.0025
(a) t (s) (b) t (s)

Figure 5.2: Schematics of analytical temperature approximation T (rθ): (a) lo-

cation of elliptical isotherm from the center (x0 , y0 ) of the melt pool, (b) radial
interpolation between center temperature Tc and solidus temperature TS ).

In order to determine a reasonable location for the liquidus temperature using

radial interpolation between Tc (θ, t) at r = 0 and TS at r = rS , the time evolution
of the melt pool’s center temperature Tc (θ, t) is described. For our specific FE
findings, Tc (θ, t) is well represented by a function that depends on the value of θ
at the beginning of melt pool shrinkage.

T1 (θ) = Tw − (Tw − Td ) sin(θ), (5.7)

88
 5.1 Methods

where Tw and Td are values of Tc extrapolated radially from the location of TS and
TL along the width and depth, respectively, and using a simple sine interpolation
between Tw and Td along θ. At the later stage of the melt pool shrinking, Tc (θ, t)
is well represented by the following time-dependent function

T2 (t) = TL − Υ (t − tL ), (5.8)

where Υ represents a constant cooling rate and tL is the approximate time at

which Tc approaches TS (see Figure 5.2.b). Using a sigmoid (hyperbolic tangent)
function, Tc (θ, t) is interpolated in time between T1 (θ) and T2 (t) between early and
late stages

T1 (θ) + T2 (t)
Tc (θ, t) =
2  
T1 (θ) − T2 (t) t − τ (θ)
− tanh , (5.9)
2 σ(θ)

in which the time and duration of the sigmoid transition from T1 to T2 are parametrized
using

τ (θ) = τw − (τw − τd ) sin(θ), (5.10)

σ(θ) = σw − (σw − σd ) sin(θ), (5.11)

here, τw and σw correspond to optimal values along the width (θ = 0), whereas τd
and σd are optimal values along the depth (θ = π/2), interpolated along θ using a
simple sine function. In terms of liquidus (TL = 1625 K) and solidus (TS = 155 K)
locations and their time evolution, parameter values Tw = 2700 K, Td = 2450 K,
Υ = 1.2 × 105 K/s, tL = 0.00195 s, τw = 0.00115 s, τd = 0.0007 s, σw = 0.00036 s,
and σd = 0.0005 s were found to give an excellent match to our FE results (see
Figure 5.3).

The temperature field provided by Eq. (5.1) is arguably approximate and has
a large number of adjustable parameters, some of which, if not chosen correctly,
may even lead to singularities. As a result, it is not intended to provide a general

89
5.1 Methods

description of a melt pool temperature field. A more accurate option might have
been to directly use a tabular FE-calculated field with an effective interpolation
method. For our investigation of grain growth competition, this analytical approx-
imation nevertheless offers a more than sufficient analytical approximation of the
temperature field calculated by FE.
t=0 t = 0.0002 s t = 0.001 s t = 0.0015 s
0 0 0 0
TS [FE]
20 TL [FE] 20 20 20
TS [Eq. (5.1)]
40 40 40 40
y (μm)

y (μm)

y (μm)

y (μm)
TL [Eq. (5.1)]

60 60 60 60

80 80 80 80

100 100 100 100

0 20 40 60 80 100 120 140 0 20 40 60 80 100 120 140 0 20 40 60 80 100 120 140 0 20 40 60 80 100 120 140
x (μm) x (μm) x (μm) x (μm)

t = 0.0018 s t = 0.0023 s t = 0.0025 s

0 0 0

20 20 20

40 40 40
y (μm)

y (μm)

y (μm)
60 60 60

80 80 80

100 100 100

0 20 40 60 80 100 120 140 0 20 40 60 80 100 120 140 0 20 40 60 80 100 120 140
x (μm) x (μm) x (μm)

Figure 5.3: Time evolution (snapshots) of the locations of the solidus and liq-
uidus isotherms in a cross section, comparing FE results of Ref. [79] (symbols)
to Eq. (5.1) (lines).

Explored thermal configurations

We simulate 15 different cross-section configurations according Table 5.1 in

addition to one longitudinal section configuration that is similar to that presented
in section 4.3. For the cross-section configurations, we consider the three different
melt pool shapes: (1) one corresponding to the original fit to the FE results (i.e.
using equations and parameters listed above), (2) one with a twice wider and twice
shallower melt pool (i.e. with w0 = 256 µm, d0 = 48 µm), and (3) one with a
twice deeper and twice narrower melt pool (i.e. with w0 = 64 µm, d0 = 192 µm).
We consider five initial configurations with different randomly generated initial
grain maps for each of these three melt pool shapes. In order to do this, a Poisson
disk sampling algorithm is used to produce the grain centers (a total of 1500
for each simulation) at random locations. Every grain center is given a random

90
 5.1 Methods

orientation, which is an integer in the range [1 − 90]. After the Voronoi tesselation
step, which propagates each center’s orientation to its corresponding polygonal
grain, the grid points with temperatures over TL , which are located in the melted
region, are reinitialized to the liquid state and their grain orientation is “erased.”
For each of the sixteen configurations, solidification calculations are run once with
PF simulations (Section 4.4) and many times with CA simulations using various
grid element sizes (Section 5.1.2).

Table 5.1: Considered thermal configurations for cross-section simulations.

Configuration Domain Size w0 d0
(µm × µm) (µm) (µm)
Reference 257 × 97 128 96
Wide & Shallow 513 × 49 256 48
Deep & Narrow 129 × 193 64 192

5.1.2 Cellular automaton

In order to track the grain envelope, we implemented a standard 2D cellular

automata model [58], [59], [67]. The grid used by the model has a uniform cell size
of hCA . Each cell is associated with a growing polygon whose vertices are oriented
toward the ⟨100⟩ (⟨10⟩ in 2D) crystalline directions of the grain and grow at a
velocity calculated from the local undercooling. The captured cell is activated and
its own associated polygon is initialized and begins to develop after the polygon
has grown enough to capture the centers of neighboring cells.

Moore neighborhood (eight-neighbor), von Neumann (four-neighbor), and Hexag-

onal are the most important types of neighborhoods (see Figure 5.4). An eight-
neighbor (Moore) capture and a standard decentered square technique were em-
ployed [61], [67], [135], [136]. There are more advanced algorithms that could
be more accurate when there is a high temperature gradient, such as those that
considering irregular (kite-shaped) quadrangles that account for the various lo-
cal velocities of the individual vertices [65], [137]–[139]. However, we choose the
decentered square method since it is one of the most extensively used due to its

91
5.1 Methods

simplicity, particularly in recent applications to additive manufacturing [61], [63],

[140], while the methodology reduces grid-induced anisotropy and correctly cap-
tures non-uniform temperature fields [67].

Model and algorithm used in CA are essentially equal to those in Ref. [61], [67].
Our algorithm was tested against the analytical result of a tilted grain growing in a
temperature gradient, specifically by reproducing the results in Ref. [67] as shown
in Figure 5.5.

Figure 5.4: The most important neighborhood types [141]

Figure 5.5: Reproducing the results in Ref. [67]

92
 5.1 Methods

An estimation of the square vertices’ growth velocity, V, was made using

VKGT = 6.45 × 10−8 × ∆T 3.83 + 5.71 × 10−6 × ∆T 1.98 , (5.12)

where the undercooling ∆T = TL − T is measured at the cell’s center. The coeffi-

cients and exponents of Eq. (5.12) were adjusted to approximate the results of the
classical Kurz-Giovanola-Trivedi model [142] at V < 0.1 m/s using parameters of
the Ni-5wt%Nb considered in the PF simulations.

For the sake of simplicity, all CA simulations consider a frame that is fixed to
the material and an evolving temperature field that either follows Eq. (4.8) with
x0 (t) = x0 (0) + VL × t, where VL = 0.1m/s is the velocity of the laser beam, or a
shrinking melt pool that follows Eq. (5.1). The boundary conditions, processing
parameters, and other alloy properties are similar to those employed in the PF
simulations.

Two significant computational advantages of the chosen CA algorithm over the

PF method are: (1) the CA model does not need to track the solute concentration
field because it just depends on the temperature field, and (2) individual cells
size hCA used in the CA can be selected to be considerably coarser than what is
necessary for accurate PF simulations (usually is set to be an order of magnitude
smaller than the dendrite tip radius [106], [143]). However, the best method for
selecting the numerical parameter hCA is still unclear. So, here we want to answer
two important questions:

• How well can a CA approximate PF predictions of grain growth competition

in a polycrystalline unsteady 2D temperature field?

• What is a suitable element size for the CA grid, or, more specifically, at
what value of hCA do CA predictions start diverging considerably from the
reference findings at low hCA ?

As a consequence, we run CA simulations for each configuration (one longitudi-

nal and fifteen cross-sections) using various hCA , and then quantitatively compare

93
5.1 Methods

the generated grain maps with one another and with the PF results (see Sec-
tion 5.1.3). We explore values of hCA with a ratio hCA /hPF between 6 and 1152 for
the longitudinal section (hPF = 5 nm) and between 3 and 144 for the cross-section
(hPF = 10 nm) simulations.

As expected, the CA simulations provide a significant computational advan-

tage over PF simulations due to the coarser grid. Each CA run was carried out
on a single Intel Xeon 6130 CPU core (2.10 GHz) rather than many parallel GPUs
as in PF simulations. The CA simulations were finished for the longitudinal sec-
tion simulation in 30 seconds or less (for hCA /hPF ≥ 192) and up to 10.5 days
(for hCA /hPF = 6), compared to the 10 days using eight GPUs (PF). An aver-
age simulation using an intermediate grid, for instance, one with hCA /hPF = 24,
lasts for roughly 4 hours. Regarding the cross-section simulation, the CA simu-
lations with the reference melt pool size were finished in between a few seconds
(hCA /hPF = 144) and 4 days (hCA /hPF = 3) compared to 44h using two GPUs
(PF). An intermediate grid, with hCA /hPF = 24, simulation takes roughly 11 min-
utes.

5.1.3 Post-processing

On the basis of the generated grain maps and overall grain textures, we seek
to compare CA and PF findings quantitatively. We also want to investigate if the
melt pool’s favored grain texture may change as a result of the shape of the melt
pool.

Since we employ a cell size hCA that is systematically a multiple of the PF grid
element size hPF , we may artificially refine the CA results to a grid that matches
that of the PF in order to compare PF and CA grain maps. In order to achieve
this, we assign each point of the refined grid included within a CA cell as the grain
orientation of the given cell. Counting the number of points inside the melt pool
that have a similar or different grain orientation will then allow us to compare the
two grids of similar size. In order to ignore the unmelted Voronoi grain distribution
at the bottom or sides of the domain, we define a region of measurement that only

94
 5.1 Methods

covers the melted region. In the longitudinal simulation, we also exclude the upper
region of the domain, which may be affected by the boundary conditions, namely
excluding the upper 10 µm at the top of the simulation from the measurement
region. Inside the resulting region, denoted ∅, we count the number number of
grid grid points (i, j) ∈ ∅ where the grain orientation differs between CA and PF,
and we normalize this number by the total number of grid points in ∅. By doing
this, we can derive a difference metric (D) that is equal to 0 if the grain maps are
precisely the same and 1 if all grid points in the region under consideration (∅)
differ. In the same melted region ∅, we also extract the overall grain orientation
distribution inside each melt pool. We only perform this analysis on the cross-
section simulations, which present a higher degree of statistical relevance with five
different simulations performed for each configuration. The grain distributions are
simply extracted as the total number of grid points in a given orientation, averaged
across the five runs for each configuration, and then binned into orientation ranges
of width 9◦ and centered on 0◦ , 9◦ , and 18◦ , . . . 90◦ (the last bin centered on 90◦
being the same as the bin centered on 0◦ ).

We also extract values of primary and secondary arm spacings from the PF-
simulated microstructures in order to compare the grid element sizes to actual
length scales of the microstructure. To achieve this, we perform line scans on the
solute composition field in large grains presenting the clearest primary or secondary
spacings (similar to those shown in Fig. 4.9, where scans A, B, C, and E corre-
spond to primary spacings and scan D corresponds to secondary spacings). The
estimation of the spacing is thus straightforward to do by identifying the first and
last concentration peaks (which correspond to interdendritic regions) and counting
the number of spacings between them. The average spacings that are produced are
corresponding to a total of 63 primary and 159 secondary measured spacings in the
longitudinal simulation and a total of 196 primary and 388 secondary measured
spacings in the cross-section simulations. We did not estimate secondary spacings
in the shallow melt pools that were composed principally of primary dendrites,
and we averaged the primary (λ1 ) and secondary (λ2 ) spacings over the different

95
5.1 Methods

melt pools sizes, as we did not see any significant differences between melt pool
shapes.

Computational performance

The following tables list the computational efficiency of the different simula-
tions in terms of speedup (wall time ratio) and simulation (wall) time from cellular
automaton (CA) compared to phase-field (PF) simulations. Here we are compar-
ing CA simulations on a single core on a desktop or laptop computer with PF
simulations performed on multiple GPUs (each with thousands of CUDA cores).
For a similar simulation, the speedup from a single core to a single-GPU parallel
implementation is roughly of the order of 20×, and the strong scaling with the
number of GPUs (here 2 to 8) was also observed to be nearly linear. Both effects
are shown to significantly improve the PF performance in the tables below. Also
take note of the considerable differences in computation times when using various
GPU models, for instance, with a factor of ≈ 1.4× between the cross-section refer-
ence PF calculations, performed on two RTX3090 (faster), and the wide/shallow
PF simulations, performed on two RTX2080Ti (slower), despite the simulation size
being nearly identical.

Table 5.2: Computational performance for longitudinal section simulations.

Method: Phase-Field Cellular Automaton
Hardware: 8 × RTX2080Ti 1 × Intel Xeon 6130 core (2.10 GHz)
Wall time hCA /hPF Wall time Speed-up PF/CA
9d16h 6 10d10h 0.93
(232h) 12 1d08h 7.2
24 4h02m 58
48 32m 430
96 4m 3 500
192 33s 25 000
384 4s 210 000
768 1s 840 000

96
 5.1 Methods

Table 5.3: Computational performance for cross-section simulations for refer-

ence melt pool size.
Method: Phase-Field Cellular Automaton
Hardware: 2 × RTX3090 1 × Intel Xeon 6130 core (2.10 GHz)
Wall time hCA /hPF Wall time Speed-up PF/CA
1d19h 3 4d05h 0.43
(43h) 6 13h02m 3.3
9 4h20m 9.9
12 1h35m 27
18 33m 78
24 11m 230
36 3m 860
48 81s 1 900
72 29s 5 300
96 10s 15 000
114 3s 52 000

Table 5.4: Computational performance for cross-section simulations for wide &
shallow melt pool size.
Method: Phase-Field Cellular Automaton
Hardware: 2 × RTX2080Ti 1 × Intel Xeon 6130 core (2.10 GHz)
Wall time hCA /hPF Wall time Speed-up PF/CA
2d11h 3 7d21h 0.31
(59h) 6 1d00h 2.4
9 8h10m 7.2
12 3h25m 19
18 1h08m 52
24 26m 140
36 10m 350
48 3m 1 200
72 58s 3 700
96 19s 11 000
114 6s 35 000

97
5.2 Results and discussion

Table 5.5: Computational performance for cross-section simulations for deep &
narrow melt pool size.
Method: Phase-Field Cellular Automaton
Hardware: 2 × RTX3090 1 × Intel Xeon 6130 core (2.10 GHz)
Wall time hCA /hPF Wall time Speed-up PF/CA
1d18h 3 2d05h 0.79
(42h) 6 6h17m 6.7
9 2h12m 19
12 45m 56
18 13m 190
24 5m 500
36 93s 1 600
48 40s 3 800
72 14s 11 000
96 5s 30 000
114 2s 76 000

5.2 Results and discussion

5.2.1 Longitudinal section

For three different levels of CA grid size, namely hCA /hPF = 9, 48, and 288, as
well as for the well-converged (hPF = 5 nm) PF simulation, the grain maps in the
longitudinal section are shown in Figure 5.6. Due to the small sample size provided
by the number of grains in this single simulation, we chose not to quantitatively
investigate the average grain distributions in the melt pool. In spite of this, the
majority of simulations show a close grain distribution, with a grain orientation,
or preferred growth direction α, measured counterclockwise from the horizontal
x+ direction, close to 60◦ (mostly in shades of green) in most of the melt pool and
closer to 45◦ (orange) around the melt pool’s tail. In addition, the PF simulation
displays a wider distribution of orientations with several large grains at α ≈ 15◦
(purple), 30◦ (red), and 0◦ (blue), which are considerably less present in the CA-
predicted grain maps.

Three regions in Figure 5.6 with the labels A, B, and C can be seen in more
detail in Figure 5.7. To better understand the agreements and differences between

98
 5.2 Results and discussion

PF and CA predictions, the underlying dendritic structures are overlaid on the PF

grain map. This allows identifying the mechanism of grain growth competition at
play at each GB (namely impingement at converging GBs and sidebranching at
diverging GBs [54]).

Region A, which is close to the melt pool’s tail, represents an example of a

region where PF and CA agree well, almost independently of the CA grid. Except
for the diverging GB 4 − 5, most grain boundaries are of the converging type. The
preferred growth orientation is close to α ≈ 45◦ since the main direction of the
temperature gradient is normal to the fusion line (displayed in dashed black line).

Figure 5.6: Selected grain orientations (color map) predicted by phase-field (top
row) and cellular automaton (bottom rows) with different grid coarsening levels
hCA /hPF = 9, 48, and 288. The brighter (washed out) area is outside of the
region Φ considered for the quantitative measurements and comparisons of grain
distributions. The dashed black line represents the limit of the melted region.

99
5.2 Results and discussion

Therefore, all convergent GBs in this region follow the grains with a favorable
orientation. With sidebranches emerging from both grains, the diverging GB 4 − 5
is oriented between grains 4 and 5, closer to the geometrical limit [65]. The only
significant difference between the PF and CA results is the early disappearance of
grain 3 in the PF results, which is caused by the early sidebranching of grain 2
that is more favorably oriented.

Converging (7 − 8, 8 − 9, 8 − 10, 10 − 11, 10 − 12) and diverging (9 − 10, 11 − 12)

GBs are present in Region B at the melt pool’s bottom. The least-favorably ori-
ented grains 9 and 11 are quickly eliminated by sidebranching at diverging GBs
and at converging GBs in PF simulations (considering a temperature gradient es-
sentially vertical with a slight tilt toward the right). In CA simulations, these

Figure 5.7: Comparison of selected grain orientations (color map) predicted by

phase-field (top row) and cellular automaton (bottom rows) with different grid
coarsening levels hCA /hPF = 9, 48, and 288, within the regions A, B, and C
highlighted in Fig. 5.6. The PF grain map is overlaid with the inner grain
dendritic structure. The dashed black lines show the limit of the melted region.

100
 5.2 Results and discussion

grains survive longer. As the melt pool proceeds, the temperature gradient direc-
tion tilts closer to grain 11’s preferred growth direction (α ≈ 45◦ ), and grain 11
even survives to become one of the main grains in the final microstructure (see
Fig. 5.6). The disappearance of grain 7 on the PF map, which was similarly a vic-
tim of early sidebranching from neighboring grains, is another noticeable difference
between PF and CA.

Significant disparities between the predicted maps can be seen in Region C,

which is also at the bottom of the melt pool. Converging GBs there are 15 − 16
and 16−17, whereas diverging GBs are 13−15 and 14−15. Furthermore, region C
includes the low-angle GBs 17 − 18 and 18 − 19. Other grain boundaries including
grain 14 (i.e. 13 − 14 and 14 − 17, the latter of which only appears in the PF
simulation) are not specified because of grain 14’s close to 45◦ misorientation with
its neighbors. In a static temperature gradient, the growth of a degenerate (sea-
weed) microstructure would have occurred at a significantly higher undercooling
and would have been strongly penalized in the grain growth competition [54], due
to the misorientation of the temperature gradient by 45◦ . In spite of strong early
competition with neighbor grain 13, grain 14 survives for a long time in the PF
simulations (see Figure 5.6). The early elimination of grain 15 in PF, which does
not happen in CA simulation at low hCA and even causes the grain to become
one of the major grains in the following grain growth competition, is the source
of a significant discrepancy between PF and CA in this case as well (see Fig. 5.6).
Even with grain 15, the coarsest CA simulation (bottom row) produces an entirely
different result. Due to the initial grain boundary (GB) between grains 15 and
20 being close to the fusion line (black dashed line), a coarse discretization with
hCA /hPF ≥ 288 results in grain 15 being completely melted, which causes grain
15 to be suppressed and replaced by lower grain 20. As a result, the identity
and orientation of initially competing grains may completely change as the spatial
discretization gets coarser.

A region just above the fusion line appears in PF simulations, corresponding

to the initial growth with a planar solid-liquid interface before the morphological
destabilization into dendrites, after which the actual growth competition among

101
5.2 Results and discussion

dendrites starts. This is a notable difference between PF and CA simulations. On

the other hand, in CA simulations, the grains begin to compete from the moment
the simulation begins, and the main orientation of GBs-which is primarily linear
at this scale-initiated directly from the fusion line. These differences between CA
and PF, which are also shown in all of the cross-section simulations reported later,
highlight the need for PF’s level of accuracy for morphological transitions (such as
from planar to dendritic growth) and the prediction of non-planar (rough) GBs.

Figure 5.8 compares the area fractions of the CA and PF grain maps for various
CA grid sizes in the measured area Φ. Vertical lines are also presented for the
measured primary and secondary arms spacings, which are, respectively, λ1 =
1.26 µm and λ2 = 1.03 µm.

From Fig. 5.8, several significant conclusions may be drawn. First, as the
CA grid becomes more coarse, the difference between CA and PF is essentially

<latexit sha1_base64="Kl5HDRJefjvc9sthS2/dkt9DxIg=">AAAB/nicbZDLSsNAFIZPvNZ6i4orN4NF6KomRdRlpSAuK9gLtCFMppN26EwSZiZCCQVfxY0LRdz6HO58G6dtFtr6w8DHf87hnPmDhDOlHefbWlldW9/YLGwVt3d29/btg8OWilNJaJPEPJadACvKWUSbmmlOO4mkWASctoNRfVpvP1KpWBw96HFCPYEHEQsZwdpYvn089LOeFKh+MznPsXE78e2SU3FmQsvg5lCCXA3f/ur1Y5IKGmnCsVJd10m0l2GpGeF0UuyliiaYjPCAdg1GWFDlZbPzJ+jMOH0UxtK8SKOZ+3siw0KpsQhMp8B6qBZrU/O/WjfV4bWXsShJNY3IfFGYcqRjNM0C9ZmkRPOxAUwkM7ciMsQSE20SK5oQ3MUvL0OrWnEvK9X7i1KtnMdRgBM4hTK4cAU1uIMGNIFABs/wCm/Wk/VivVsf89YVK585gj+yPn8ALRSU5w==</latexit>

hhCA /hPF
CA/hPF
10 100 1000
fraction)

1 <latexit sha1_base64="y7Qi/4ih6UUWbWURqYKENk4YSbg=">AAAB8HicdVDLSgMxFL1TX7W+qi7dBKvgqsyUUl0W3LisYB/SDiWTybShSWZIMkIZ+hVuXCji1s9x59+YtiNU0QOBwznnkntPkHCmjet+OoW19Y3NreJ2aWd3b/+gfHjU0XGqCG2TmMeqF2BNOZO0bZjhtJcoikXAaTeYXM/97gNVmsXyzkwT6gs8kixiBBsr3Q+4jYZ4WBuWK27VXQCtkMaSeLlSgRytYfljEMYkFVQawrHWfc9NjJ9hZRjhdFYapJommEzwiPYtlVhQ7WeLhWfo3CohimJlnzRooa5OZFhoPRWBTQpsxvq3Nxf/8vqpia78jMkkNVSS5UdRypGJ0fx6FDJFieFTSzBRzO6KyBgrTIztqGRL+L4U/U86tarXqNZu65XmWV5HEU7gFC7Ag0towg20oA0EBDzCM7w4ynlyXp23ZbTg5DPH8APO+xdqH5AR</latexit>

<latexit sha1_base64="CzTvFkC11wCO0BdR3M2VOMnsVJY=">AAAB8HicbVDLSgMxFL1TX7W+qi7dBKvgapgpUl0W3LisYB/SDiWTybShSWZIMkIp/Qo3LhRx6+e482/MtLPQ6oHA4Zxzyb0nTDnTxvO+nNLa+sbmVnm7srO7t39QPTzq6CRThLZJwhPVC7GmnEnaNsxw2ksVxSLktBtObnK/+0iVZom8N9OUBgKPJIsZwcZKDwNuoxEe+sNqzXMbXg7kud4K8QtSgwKtYfVzECUkE1QawrHWfd9LTTDDyjDC6bwyyDRNMZngEe1bKrGgOpgtFp6jc6tEKE6UfdKghfpzYoaF1lMR2qTAZqxXvVz8z+tnJr4OZkymmaGSLD+KM45MgvLrUcQUJYZPLcFEMbsrImOsMDG2o4otwV89+S/p1F2/4dbvLmvNs6KOMpzAKVyAD1fQhFtoQRsICHiCF3h1lPPsvDnvy2jJKWaO4Recj29pB5AQ</latexit>

<latexit sha1_base64="9aoq6YdzGxcKglK/UpiCmlLB5c0=">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</latexit>

22 11
Moore
Moore (8N)
(8N)
fraction)

0.9 Von Neumann (4N)

Von Neumann (4N)
Higher VKGT
Higher V
(area

KGT
0.8 Lower VKGT
CA-PF (area

Lower V KGT
CA-PF

0.7

0.6
Difference
Di↵erence

0.5

0.4
0.03 5.0
<latexit sha1_base64="4lj5ap2rj0B5jsa4HWWpuKBKXYk=">AAACCXicbVC7SgNBFJ2Nrxhfq5Y2g0GIhWE3hVpGImIZwTwgCWF2cjcZMju7zMwKYUlr46/YWChi6x/Y+TfOJlto4oGBwzn3cuccL+JMacf5tnIrq2vrG/nNwtb2zu6evX/QVGEsKTRoyEPZ9ogCzgQ0NNMc2pEEEngcWt64lvqtB5CKheJeTyLoBWQomM8o0Ubq2/ia+T5IEBRw7eqsfoNLxOxjXxKaTpz27aJTdmbAy8TNSBFlqPftr+4gpHEAQlNOlOq4TqR7CZGaUQ7TQjdWEBE6JkPoGCpIAKqXzJJM8YlRBtgPpXlC45n6eyMhgVKTwDOTAdEjteil4n9eJ9b+ZS9hIoq1yTo/5Mcc6xCnteABk0A1nxhCqGTmr5iOSFqCKa9gSnAXIy+TZqXsnpcrd5VitZTVkUdH6BiVkIsuUBXdojpqIIoe0TN6RW/Wk/VivVsf89Gcle0coj+wPn8ALfmYqA==</latexit>

0.1 <latexit sha1_base64="PMZoKByZqdUv355ZCxbFe9+ayxs=">AAACAXicbVA9TwJBEN3zE/Hr1MbEZiOYYEPuKNQSQ2OJiXwkQMjessCG3b3L7pyRXLDxr9hYaIyt/8LOf+MCVyj4kkle3pvJzLwgEtyA5307K6tr6xubma3s9s7u3r57cFg3Yawpq9FQhLoZEMMEV6wGHARrRpoRGQjWCEaVqd+4Z9rwUN3BOGIdSQaK9zklYKWue5wfdpO2lrhyPcnjQhvYA8gYy/Oum/OK3gx4mfgpyaEU1a771e6FNJZMARXEmJbvRdBJiAZOBZtk27FhEaEjMmAtSxWRzHSS2QcTfGaVHu6H2pYCPFN/TyREGjOWge2UBIZm0ZuK/3mtGPpXnYSrKAam6HxRPxYYQjyNA/e4ZhTE2BJCNbe3YjokmlCwoWVtCP7iy8ukXir6F8XSbSlXLqRxZNAJOkUF5KNLVEY3qIpqiKJH9Ixe0Zvz5Lw4787HvHXFSWeO0B84nz9LDpVt</latexit>

1
hhCA
CA ( (µm)
m)

Figure 5.8: Difference between orientation maps in CA and PF results, as a

fraction area of the measured area Φ, considering CA results using either Moore
(eight-neighbor) or von Neumann (four-neighbor) neighborhood, as well as arti-
ficially higher (V = 10 × VKGT ) or lower (V = VKGT /10) growth velocities, i.e.
respectively multiplying or dividing the right-hand side of Eq. (5.12) by a factor
of 10. Vertical lines show the measured values of primary (λ1 ) and secondary
(λ2 ) dendrite arm spacings.

102
 5.2 Results and discussion

growing, and when hCA is decreased, the CA results appear to plateau. Even
the best findings for a well-converged CA, obtained for around hCA ≤ 0.2 µm
(i.e. hCA /hPF ≤ 40), still demonstrate a difference of more than 40% from PF.
Compared to the measured dendrite arm spacings of λ1 = 1.26 µm and λ2 =
1.03 µm, this hCA is much lower. This monotonic behavior differs from earlier
findings in bi-crystal configurations [65], [66], which demonstrated that a minimum
can be found between the so-called geometrical limit (at low hCA ) and the favorably-
oriented grain (at high hCA ) asymptotic behaviors. As the cell size approaches
grain size at the highest hCA , the CA prediction fails over more than 90% of the
solidified area. These findings suggest that low hCA , which results in a smaller
difference with PF, is the most suitable choice for the CA grid size, although
the relationship with microstructural length scales is still unclear and needs more
research.

We also looked at how other important CA model parameters, such as the

neighborhood considered for cell captures, as well as the law considered for the
dendrite tip (i.e. square vertices) growth rule, affected these results. The findings
are not significantly changed when changing the capture neighboring algorithm
from Moore (eight neighbors, including diagonals, used as default here) to von
Neumann (four neighbors) specification.

In terms of the growth velocity, we discover that, at least in this particular case,
either raising the growth velocity (by multiplying V from Eq. 5.12 by a factor 10)
or reducing it (by dividing by a factor of 10) has a rather little impact. However,
given that we are in the presence of a relatively high-temperature gradient, it is
important to proceed with caution when making this conclusion. A higher V (∆T )
will have the main effect of moving the solidification front (dendrite tips) closer
to the liquidus temperature, stabilizing at a relatively lower ∆T for the same
V , whereas a lower V (∆T ) will encourage stabilization of the solidification front
deeper in the mushy zone (closer to the solidus temperature). The solidus and
liquidus isotherms are quite close to one another due to the high temperature
gradient considered here, between 106 at the bottom and 107 K/m at the tail of
the melt pool (See Ref. [79], Fig. 5 therein), which reduces the influence of V (∆T ).

103
5.2 Results and discussion

5.2.2 Cross-sections

Effect of CA grid

Figure 5.9 compares the CA and PF grain maps in the measured cross-section
simulation regions for various CA grid sizes: (a) the average of all 15 simulations
or (b) the average of the five simulations for each melt pool shape. With the
exception of the wide and shallow melt pools, which essentially had only primary
dendrites and very few secondary branches, the primary and secondary dendrite
arms spacings in these cases did not show any clear relationship with the melt pool
shape. Because of this, we averaged the spacings overall measurable simulations,
resulting in λ1 = 1.22 µm and λ2 = 0.82 µm.

The observed trends are essentially the same, despite the errors at low hCA ,
between 30% and 40%, appearing to be slightly smaller than in the longitudinal
simulation. In other words, when hCA increases, the difference between PF and
CA maps increases, and the low-hCA plateau is reached at a grid size considerably
smaller than the characteristic microstructural length scales λ1 and λ2 . Despite
this apparent poor PF-CA match, we illustrate in the following subsection that,
statistically speaking, disregarding the precise location or grain shape, the finer
CA simulations provide a good representation of the average statistical distribu-
tion of grain orientation. This is true even though the discrepancy remains when
integrating the difference point-by-point.

Effect of melt pool shape

Figure 5.10 compares the histograms for grain distribution of PF simulation

results (lighter color) with CA predictions (darker color) using fine (hCA /hPF = 3,
left) and coarse (hCA /hPF = 144, right) grids. The histograms are averaged over
five simulations for each of the three different melt pool shapes (rows).

Simple geometrical considerations [144] suggest that, due to the main temper-
ature gradient direction, a transition occurs between a predominant ⟨100⟩ grain
texture in the horizontal built direction (i.e., in 2D, a high density of grains with

104
 5.2 Results and discussion

α ≈ 0◦ ) for a relatively wide/shallow or a deep/narrow melt pool to a predomi-

nant ⟨110⟩ texture (i.e. α ≈ 45◦ ) for melt pool cross-section aspect ratio closer to
unity (i.e. close to circular in shape). Experimental findings, such as those from
the powder-bed laser melting of pure copper [144] and pure molybdenum [145],

<latexit sha1_base64="Kl5HDRJefjvc9sthS2/dkt9DxIg=">AAAB/nicbZDLSsNAFIZPvNZ6i4orN4NF6KomRdRlpSAuK9gLtCFMppN26EwSZiZCCQVfxY0LRdz6HO58G6dtFtr6w8DHf87hnPmDhDOlHefbWlldW9/YLGwVt3d29/btg8OWilNJaJPEPJadACvKWUSbmmlOO4mkWASctoNRfVpvP1KpWBw96HFCPYEHEQsZwdpYvn089LOeFKh+MznPsXE78e2SU3FmQsvg5lCCXA3f/ur1Y5IKGmnCsVJd10m0l2GpGeF0UuyliiaYjPCAdg1GWFDlZbPzJ+jMOH0UxtK8SKOZ+3siw0KpsQhMp8B6qBZrU/O/WjfV4bWXsShJNY3IfFGYcqRjNM0C9ZmkRPOxAUwkM7ciMsQSE20SK5oQ3MUvL0OrWnEvK9X7i1KtnMdRgBM4hTK4cAU1uIMGNIFABs/wCm/Wk/VivVsf89YVK585gj+yPn8ALRSU5w==</latexit> <latexit sha1_base64="yj7Y/Hu0XrvJA0K/ImmixyIWfqM=">AAAB6nicbVA9SwNBEJ2LXzF+RS1tFoMQm3CXQi0DNpYRzQckR9jbzCVL9vaO3T0hHPkJNhaK2PqL7Pw3bpIrNPHBwOO9GWbmBYng2rjut1PY2Nza3inulvb2Dw6PyscnbR2nimGLxSJW3YBqFFxiy3AjsJsopFEgsBNMbud+5wmV5rF8NNME/YiOJA85o8ZKD1V6OShX3Jq7AFknXk4qkKM5KH/1hzFLI5SGCap1z3MT42dUGc4Ezkr9VGNC2YSOsGeppBFqP1ucOiMXVhmSMFa2pCEL9fdERiOtp1FgOyNqxnrVm4v/eb3UhDd+xmWSGpRsuShMBTExmf9NhlwhM2JqCWWK21sJG1NFmbHplGwI3urL66Rdr3lXtfp9vdKo5nEU4QzOoQoeXEMD7qAJLWAwgmd4hTdHOC/Ou/OxbC04+cwp/IHz+QOCg401</latexit>

hCA
hCA/h
/hPF
PF (a)
3 6 9 12 18 24 36 48 72 96 144
(area fraction)

0.8

0.7
CA-PF CA-PF

0.6
Di↵erenceDifference

0.5

0.4

0.3

0.03
<latexit sha1_base64="y7Qi/4ih6UUWbWURqYKENk4YSbg=">AAAB8HicdVDLSgMxFL1TX7W+qi7dBKvgqsyUUl0W3LisYB/SDiWTybShSWZIMkIZ+hVuXCji1s9x59+YtiNU0QOBwznnkntPkHCmjet+OoW19Y3NreJ2aWd3b/+gfHjU0XGqCG2TmMeqF2BNOZO0bZjhtJcoikXAaTeYXM/97gNVmsXyzkwT6gs8kixiBBsr3Q+4jYZ4WBuWK27VXQCtkMaSeLlSgRytYfljEMYkFVQawrHWfc9NjJ9hZRjhdFYapJommEzwiPYtlVhQ7WeLhWfo3CohimJlnzRooa5OZFhoPRWBTQpsxvq3Nxf/8vqpia78jMkkNVSS5UdRypGJ0fx6FDJFieFTSzBRzO6KyBgrTIztqGRL+L4U/U86tarXqNZu65XmWV5HEU7gFC7Ag0towg20oA0EBDzCM7w4ynlyXp23ZbTg5DPH8APO+xdqH5AR</latexit> <latexit sha1_base64="CzTvFkC11wCO0BdR3M2VOMnsVJY=">AAAB8HicbVDLSgMxFL1TX7W+qi7dBKvgapgpUl0W3LisYB/SDiWTybShSWZIMkIp/Qo3LhRx6+e482/MtLPQ6oHA4Zxzyb0nTDnTxvO+nNLa+sbmVnm7srO7t39QPTzq6CRThLZJwhPVC7GmnEnaNsxw2ksVxSLktBtObnK/+0iVZom8N9OUBgKPJIsZwcZKDwNuoxEe+sNqzXMbXg7kud4K8QtSgwKtYfVzECUkE1QawrHWfd9LTTDDyjDC6bwyyDRNMZngEe1bKrGgOpgtFp6jc6tEKE6UfdKghfpzYoaF1lMR2qTAZqxXvVz8z+tnJr4OZkymmaGSLD+KM45MgvLrUcQUJYZPLcFEMbsrImOsMDG2o4otwV89+S/p1F2/4dbvLmvNs6KOMpzAKVyAD1fQhFtoQRsICHiCF3h1lPPsvDnvy2jJKWaO4Recj29pB5AQ</latexit>

0.1 2 1 1
<latexit sha1_base64="4lj5ap2rj0B5jsa4HWWpuKBKXYk=">AAACCXicbVC7SgNBFJ2Nrxhfq5Y2g0GIhWE3hVpGImIZwTwgCWF2cjcZMju7zMwKYUlr46/YWChi6x/Y+TfOJlto4oGBwzn3cuccL+JMacf5tnIrq2vrG/nNwtb2zu6evX/QVGEsKTRoyEPZ9ogCzgQ0NNMc2pEEEngcWt64lvqtB5CKheJeTyLoBWQomM8o0Ubq2/ia+T5IEBRw7eqsfoNLxOxjXxKaTpz27aJTdmbAy8TNSBFlqPftr+4gpHEAQlNOlOq4TqR7CZGaUQ7TQjdWEBE6JkPoGCpIAKqXzJJM8YlRBtgPpXlC45n6eyMhgVKTwDOTAdEjteil4n9eJ9b+ZS9hIoq1yTo/5Mcc6xCnteABk0A1nxhCqGTmr5iOSFqCKa9gSnAXIy+TZqXsnpcrd5VitZTVkUdH6BiVkIsuUBXdojpqIIoe0TN6RW/Wk/VivVsf89Gcle0coj+wPn8ALfmYqA==</latexit>

<latexit sha1_base64="PMZoKByZqdUv355ZCxbFe9+ayxs=">AAACAXicbVA9TwJBEN3zE/Hr1MbEZiOYYEPuKNQSQ2OJiXwkQMjessCG3b3L7pyRXLDxr9hYaIyt/8LOf+MCVyj4kkle3pvJzLwgEtyA5307K6tr6xubma3s9s7u3r57cFg3Yawpq9FQhLoZEMMEV6wGHARrRpoRGQjWCEaVqd+4Z9rwUN3BOGIdSQaK9zklYKWue5wfdpO2lrhyPcnjQhvYA8gYy/Oum/OK3gx4mfgpyaEU1a771e6FNJZMARXEmJbvRdBJiAZOBZtk27FhEaEjMmAtSxWRzHSS2QcTfGaVHu6H2pYCPFN/TyREGjOWge2UBIZm0ZuK/3mtGPpXnYSrKAam6HxRPxYYQjyNA/e4ZhTE2BJCNbe3YjokmlCwoWVtCP7iy8ukXir6F8XSbSlXLqRxZNAJOkUF5KNLVEY3qIpqiKJH9Ixe0Zvz5Lw4787HvHXFSWeO0B84nz9LDpVt</latexit>

hCA
hCA ((µm)
m)
<latexit sha1_base64="Kl5HDRJefjvc9sthS2/dkt9DxIg=">AAAB/nicbZDLSsNAFIZPvNZ6i4orN4NF6KomRdRlpSAuK9gLtCFMppN26EwSZiZCCQVfxY0LRdz6HO58G6dtFtr6w8DHf87hnPmDhDOlHefbWlldW9/YLGwVt3d29/btg8OWilNJaJPEPJadACvKWUSbmmlOO4mkWASctoNRfVpvP1KpWBw96HFCPYEHEQsZwdpYvn089LOeFKh+MznPsXE78e2SU3FmQsvg5lCCXA3f/ur1Y5IKGmnCsVJd10m0l2GpGeF0UuyliiaYjPCAdg1GWFDlZbPzJ+jMOH0UxtK8SKOZ+3siw0KpsQhMp8B6qBZrU/O/WjfV4bWXsShJNY3IfFGYcqRjNM0C9ZmkRPOxAUwkM7ciMsQSE20SK5oQ3MUvL0OrWnEvK9X7i1KtnMdRgBM4hTK4cAU1uIMGNIFABs/wCm/Wk/VivVsf89YVK585gj+yPn8ALRSU5w==</latexit> <latexit sha1_base64="pIewZvisZw0ILX8HUGjd+s0iw3Y=">AAAB6nicbVA9SwNBEJ2LXzF+RS1tFoMQm3CXQi0DNpYRzQckR9jbzCVL9vaO3T0hHPkJNhaK2PqL7Pw3bpIrNPHBwOO9GWbmBYng2rjut1PY2Nza3inulvb2Dw6PyscnbR2nimGLxSJW3YBqFFxiy3AjsJsopFEgsBNMbud+5wmV5rF8NNME/YiOJA85o8ZKD9XgclCuuDV3AbJOvJxUIEdzUP7qD2OWRigNE1Trnucmxs+oMpwJnJX6qcaEsgkdYc9SSSPUfrY4dUYurDIkYaxsSUMW6u+JjEZaT6PAdkbUjPWqNxf/83qpCW/8jMskNSjZclGYCmJiMv+bDLlCZsTUEsoUt7cSNqaKMmPTKdkQvNWX10m7XvOuavX7eqVRzeMowhmcQxU8uIYG3EETWsBgBM/wCm+OcF6cd+dj2Vpw8plT+APn8weECI02</latexit>

hCA
hCA/h PF
/hPF (b)
3 6 9 12 18 24 36 48 72 96 144
(area fraction)

0.8
<latexit sha1_base64="10devUDZN/BMf6gvQORl/Njg3uU=">AAACHnicbVDLSgMxFM34rOOr6tJNsAquykzFx7KgC5dV7AM6Q8lk7rShmcyQZIQy9Evc+CtuXCgiuNK/MX2A2nogcHLOvSTnBClnSjvOl7WwuLS8slpYs9c3Nre2izu7DZVkkkKdJjyRrYAo4ExAXTPNoZVKIHHAoRn0L0d+8x6kYom404MU/Jh0BYsYJdpIneKpF/FM9STr9rR9CxFIEBSw57XLzgkTvt1kIcif6xVACrJTLDllZww8T9wpKaEpap3ihxcmNItBaMqJUm3XSbWfE6kZ5TC0vUxBSmifdKFtqCAxKD8fxxviI6OEOEqkOULjsfp7IyexUoM4MJMx0T01643E/7x2pqMLP2cizbRJPXkoyjjWCR51hUMmgWo+MIRQycxfMe0RSag2jdqmBHc28jxpVMruWblyUylVD6d1FNA+OkDHyEXnqIquUQ3VEUUP6Am9oFfr0Xq23qz3yeiCNd3ZQ39gfX4DB0KhDQ==</latexit>

Reference
Reference
0.7 Wider
Wider
Deeper
Deeper
CA-PFCA-PF

0.6
Di↵erenceDifference

0.5

0.4

0.3

0.03
<latexit sha1_base64="y7Qi/4ih6UUWbWURqYKENk4YSbg=">AAAB8HicdVDLSgMxFL1TX7W+qi7dBKvgqsyUUl0W3LisYB/SDiWTybShSWZIMkIZ+hVuXCji1s9x59+YtiNU0QOBwznnkntPkHCmjet+OoW19Y3NreJ2aWd3b/+gfHjU0XGqCG2TmMeqF2BNOZO0bZjhtJcoikXAaTeYXM/97gNVmsXyzkwT6gs8kixiBBsr3Q+4jYZ4WBuWK27VXQCtkMaSeLlSgRytYfljEMYkFVQawrHWfc9NjJ9hZRjhdFYapJommEzwiPYtlVhQ7WeLhWfo3CohimJlnzRooa5OZFhoPRWBTQpsxvq3Nxf/8vqpia78jMkkNVSS5UdRypGJ0fx6FDJFieFTSzBRzO6KyBgrTIztqGRL+L4U/U86tarXqNZu65XmWV5HEU7gFC7Ag0towg20oA0EBDzCM7w4ynlyXp23ZbTg5DPH8APO+xdqH5AR</latexit> <latexit sha1_base64="CzTvFkC11wCO0BdR3M2VOMnsVJY=">AAAB8HicbVDLSgMxFL1TX7W+qi7dBKvgapgpUl0W3LisYB/SDiWTybShSWZIMkIp/Qo3LhRx6+e482/MtLPQ6oHA4Zxzyb0nTDnTxvO+nNLa+sbmVnm7srO7t39QPTzq6CRThLZJwhPVC7GmnEnaNsxw2ksVxSLktBtObnK/+0iVZom8N9OUBgKPJIsZwcZKDwNuoxEe+sNqzXMbXg7kud4K8QtSgwKtYfVzECUkE1QawrHWfd9LTTDDyjDC6bwyyDRNMZngEe1bKrGgOpgtFp6jc6tEKE6UfdKghfpzYoaF1lMR2qTAZqxXvVz8z+tnJr4OZkymmaGSLD+KM45MgvLrUcQUJYZPLcFEMbsrImOsMDG2o4otwV89+S/p1F2/4dbvLmvNs6KOMpzAKVyAD1fQhFtoQRsICHiCF3h1lPPsvDnvy2jJKWaO4Recj29pB5AQ</latexit>

0.1 2 1 1
<latexit sha1_base64="4lj5ap2rj0B5jsa4HWWpuKBKXYk=">AAACCXicbVC7SgNBFJ2Nrxhfq5Y2g0GIhWE3hVpGImIZwTwgCWF2cjcZMju7zMwKYUlr46/YWChi6x/Y+TfOJlto4oGBwzn3cuccL+JMacf5tnIrq2vrG/nNwtb2zu6evX/QVGEsKTRoyEPZ9ogCzgQ0NNMc2pEEEngcWt64lvqtB5CKheJeTyLoBWQomM8o0Ubq2/ia+T5IEBRw7eqsfoNLxOxjXxKaTpz27aJTdmbAy8TNSBFlqPftr+4gpHEAQlNOlOq4TqR7CZGaUQ7TQjdWEBE6JkPoGCpIAKqXzJJM8YlRBtgPpXlC45n6eyMhgVKTwDOTAdEjteil4n9eJ9b+ZS9hIoq1yTo/5Mcc6xCnteABk0A1nxhCqGTmr5iOSFqCKa9gSnAXIy+TZqXsnpcrd5VitZTVkUdH6BiVkIsuUBXdojpqIIoe0TN6RW/Wk/VivVsf89Gcle0coj+wPn8ALfmYqA==</latexit>

<latexit sha1_base64="PMZoKByZqdUv355ZCxbFe9+ayxs=">AAACAXicbVA9TwJBEN3zE/Hr1MbEZiOYYEPuKNQSQ2OJiXwkQMjessCG3b3L7pyRXLDxr9hYaIyt/8LOf+MCVyj4kkle3pvJzLwgEtyA5307K6tr6xubma3s9s7u3r57cFg3Yawpq9FQhLoZEMMEV6wGHARrRpoRGQjWCEaVqd+4Z9rwUN3BOGIdSQaK9zklYKWue5wfdpO2lrhyPcnjQhvYA8gYy/Oum/OK3gx4mfgpyaEU1a771e6FNJZMARXEmJbvRdBJiAZOBZtk27FhEaEjMmAtSxWRzHSS2QcTfGaVHu6H2pYCPFN/TyREGjOWge2UBIZm0ZuK/3mtGPpXnYSrKAam6HxRPxYYQjyNA/e4ZhTE2BJCNbe3YjokmlCwoWVtCP7iy8ukXir6F8XSbSlXLqRxZNAJOkUF5KNLVEY3qIpqiKJH9Ixe0Zvz5Lw4787HvHXFSWeO0B84nz9LDpVt</latexit>

hCA
hCA (µm)
( m)

Figure 5.9: Difference between orientation maps in CA and PF results, as a frac-

tion area of the measured area Φ, considering CA results in the cross-section sim-
ulations: (a) averaged over all 15 simulations, (b) averaged over five simulations
for each melt pool shape (reference, wider/shallower, and deeper/narrower). The
shaded areas associated with each curve represents the distribution between the
five different runs (between minimum and maximum CA-PF difference).

105
5.2 Results and discussion

supported this trend. Even though the total number of simulations and grains
reported here is not statistically sufficient to detect such a transition in grain tex-

<latexit sha1_base64="jSrRQx5UyMCVnUWEtf/A16bGvOs=">AAACAnicbZDLSgMxFIYz9VbrbdSVuAkWxVWdqaJuhEpBXFawF2iHIZOmbWiSGZKMUIbBja/ixoUibn0Kd76NaTsLbf0h8PGfczg5fxAxqrTjfFu5hcWl5ZX8amFtfWNzy97eaagwlpjUcchC2QqQIowKUtdUM9KKJEE8YKQZDKvjevOBSEVDca9HEfE46gvaoxhpY/n23sBPOpLD6nV6kmHtJoVX8NS3i07JmQjOg5tBEWSq+fZXpxvimBOhMUNKtV0n0l6CpKaYkbTQiRWJEB6iPmkbFIgT5SWTE1J4aJwu7IXSPKHhxP09kSCu1IgHppMjPVCztbH5X60d696ll1ARxZoIPF3UixnUIRznAbtUEqzZyADCkpq/QjxAEmFtUiuYENzZk+ehUS6556Xy3VmxcpTFkQf74AAcAxdcgAq4BTVQBxg8gmfwCt6sJ+vFerc+pq05K5vZBX9kff4A6u+VvQ==</latexit> <latexit sha1_base64="o5me8FCr4jfBfWsAkxfSeRU+s6Y=">AAACBHicbZDLSgMxFIYz9VbrbdRlN8GiuKozpagboVIQlxXsBdphyKSZNjTJDElGKEMXbnwVNy4UcetDuPNtTNtZaOsPgY//nMPJ+YOYUaUd59vKrayurW/kNwtb2zu7e/b+QUtFicSkiSMWyU6AFGFUkKammpFOLAniASPtYFSf1tsPRCoaiXs9jonH0UDQkGKkjeXbxaGf9iSH9evJWYaNmwm8gm616tslp+zMBJfBzaAEMjV8+6vXj3DCidCYIaW6rhNrL0VSU8zIpNBLFIkRHqEB6RoUiBPlpbMjJvDYOH0YRtI8oeHM/T2RIq7UmAemkyM9VIu1qflfrZvo8NJLqYgTTQSeLwoTBnUEp4nAPpUEazY2gLCk5q8QD5FEWJvcCiYEd/HkZWhVyu55uXJXLdVOsjjyoAiOwClwwQWogVvQAE2AwSN4Bq/gzXqyXqx362PemrOymUPwR9bnD+CIljc=</latexit>

Reference hCA /hPF = 3 hCA /hPF = 144

<latexit sha1_base64="Hi9z9TX/osI2UT4+aYqUE7Ctacc=">AAAB9HicbVDLSgNBEOyNrxhfUY9eBoOQ07KbQ/QY8OIxinlAsoTZSW8yZPbhzGwgLPkOLx4U8erHePNvnCQraLSgoajqprvLTwRX2nE+rcLG5tb2TnG3tLd/cHhUPj5pqziVDFssFrHs+lSh4BG2NNcCu4lEGvoCO/7keuF3pigVj6N7PUvQC+ko4gFnVBvJ6/sBucMAJUYMB+WKYztLEMeuf5NccXNSgRzNQfmjP4xZGmKkmaBK9Vwn0V5GpeZM4LzUTxUmlE3oCHuGRjRE5WXLo+fkwihDEsTSVKTJUv05kdFQqVnom86Q6rFa9xbif14v1cGVl/EoSbX5arUoSAXRMVkkQIZcItNiZghlkptbCRtTSZk2OZVMCO76y39Ju2a7dbt2W6s0qnkcRTiDc6iCC5fQgBtoQgsYPMAjPMOLNbWerFfrbdVasPKZU/gF6/0LWf6RwA==</latexit>

0.18 <latexit sha1_base64="rq2y4Tr2M7GS3S5kcpsQs6ftVl8=">AAAB6XicbVA9SwNBEJ3zM8avqKXNYhSswl0KtYyksYxiPiA5wt5mLlmyt3fs7gnhyD+wsVDE1n9k579xk1yhiQ8GHu/NMDMvSATXxnW/nbX1jc2t7cJOcXdv/+CwdHTc0nGqGDZZLGLVCahGwSU2DTcCO4lCGgUC28G4PvPbT6g0j+WjmSToR3QoecgZNVZ6qN/2S2W34s5BVomXkzLkaPRLX71BzNIIpWGCat313MT4GVWGM4HTYi/VmFA2pkPsWipphNrP5pdOyYVVBiSMlS1pyFz9PZHRSOtJFNjOiJqRXvZm4n9eNzXhjZ9xmaQGJVssClNBTExmb5MBV8iMmFhCmeL2VsJGVFFmbDhFG4K3/PIqaVUr3lWlel8t187zOApwCmdwCR5cQw3uoAFNYBDCM7zCmzN2Xpx352PRuubkMyfwB87nDxZDjPk=</latexit> <latexit sha1_base64="8TCsve5hWMDXC5fOvtgfSr9ukOE=">AAAB6XicbVBNS8NAEJ3Ur1q/qh69LFbBU0l6UI8FQTxWsR/QhrLZTtqlm03Y3Qgl9B948aCIV/+RN/+N2zYHbX0w8Hhvhpl5QSK4Nq777RTW1jc2t4rbpZ3dvf2D8uFRS8epYthksYhVJ6AaBZfYNNwI7CQKaRQIbAfjm5nffkKleSwfzSRBP6JDyUPOqLHSQ+O2X664VXcOskq8nFQgR6Nf/uoNYpZGKA0TVOuu5ybGz6gynAmclnqpxoSyMR1i11JJI9R+Nr90Ss6tMiBhrGxJQ+bq74mMRlpPosB2RtSM9LI3E//zuqkJr/2MyyQ1KNliUZgKYmIye5sMuEJmxMQSyhS3txI2oooyY8Mp2RC85ZdXSatW9S6rtftapX6Wx1GEEziFC/DgCupwBw1oAoMQnuEV3pyx8+K8Ox+L1oKTzxzDHzifPzGYjQs=</latexit>

0.18 <latexit sha1_base64="rq2y4Tr2M7GS3S5kcpsQs6ftVl8=">AAAB6XicbVA9SwNBEJ3zM8avqKXNYhSswl0KtYyksYxiPiA5wt5mLlmyt3fs7gnhyD+wsVDE1n9k579xk1yhiQ8GHu/NMDMvSATXxnW/nbX1jc2t7cJOcXdv/+CwdHTc0nGqGDZZLGLVCahGwSU2DTcCO4lCGgUC28G4PvPbT6g0j+WjmSToR3QoecgZNVZ6qN/2S2W34s5BVomXkzLkaPRLX71BzNIIpWGCat313MT4GVWGM4HTYi/VmFA2pkPsWipphNrP5pdOyYVVBiSMlS1pyFz9PZHRSOtJFNjOiJqRXvZm4n9eNzXhjZ9xmaQGJVssClNBTExmb5MBV8iMmFhCmeL2VsJGVFFmbDhFG4K3/PIqaVUr3lWlel8t187zOApwCmdwCR5cQw3uoAFNYBDCM7zCmzN2Xpx352PRuubkMyfwB87nDxZDjPk=</latexit> <latexit sha1_base64="8TCsve5hWMDXC5fOvtgfSr9ukOE=">AAAB6XicbVBNS8NAEJ3Ur1q/qh69LFbBU0l6UI8FQTxWsR/QhrLZTtqlm03Y3Qgl9B948aCIV/+RN/+N2zYHbX0w8Hhvhpl5QSK4Nq777RTW1jc2t4rbpZ3dvf2D8uFRS8epYthksYhVJ6AaBZfYNNwI7CQKaRQIbAfjm5nffkKleSwfzSRBP6JDyUPOqLHSQ+O2X664VXcOskq8nFQgR6Nf/uoNYpZGKA0TVOuu5ybGz6gynAmclnqpxoSyMR1i11JJI9R+Nr90Ss6tMiBhrGxJQ+bq74mMRlpPosB2RtSM9LI3E//zuqkJr/2MyyQ1KNliUZgKYmIye5sMuEJmxMQSyhS3txI2oooyY8Mp2RC85ZdXSatW9S6rtftapX6Wx1GEEziFC/DgCupwBw1oAoMQnuEV3pyx8+K8Ox+L1oKTzxzDHzifPzGYjQs=</latexit>

CA
CA-Reference PF-Reference PF CA
CA-Reference PF-Reference PF
0.16 0.16
area fraction

0.14 0.14

0.12 0.12

Relative Population
Relative Population

0.1 0.1
Relative

0.08 0.08

0.06 0.06

0.04 0.04
<latexit sha1_base64="bAo6jjTrszvYSVpKVPMgRK/sDaM=">AAAB/3icbVC7SgNBFL0bXzG+ooKNzWAUrMJuCrUM2FhGMTGQLGF2cjcZMvtgZjYQ1hT+io2FIrb+hp1/42yyhSYeGDiccw/3zvFiwZW27W+rsLK6tr5R3Cxtbe/s7pX3D1oqSiTDJotEJNseVSh4iE3NtcB2LJEGnsAHb3Sd+Q9jlIpH4b2exOgGdBBynzOqjdQrH92hMHSMhJoY8SVlc6NiV+0ZyDJxclKBHI1e+avbj1gSYKiZoEp1HDvWbkql5kzgtNRNFMaUjegAO4aGNEDlprP7p+TMKH3iR9K8UJOZ+juR0kCpSeCZyYDqoVr0MvE/r5No/8pNeRgnGkM2X+QnguiIZGWQPpfItJgYQpnk5lbChjTrwFRWMiU4i19eJq1a1bmo1m5rlfppXkcRjuEEzsGBS6jDDTSgCQwe4Rle4c16sl6sd+tjPlqw8swh/IH1+QPEGJXZ</latexit>

0.02 0.02

0 0
0 9 18 27 36 45 54 63 72 81 90 0 9 18 27 36 45 54 63 72 81 90
Wider
<latexit sha1_base64="QGtzB1ZaNCjHyfB9PCzLcDcOJdI=">AAAB8HicdVBNS8NAEJ3Ur1q/qh69LBahp5L2UD0WvHisYD+kDWWzmbRLd5OwuxFK6K/w4kERr/4cb/4bt22EKvpg4PHeDDPz/ERwbVz30ylsbG5t7xR3S3v7B4dH5eOTro5TxbDDYhGrvk81Ch5hx3AjsJ8opNIX2POn1wu/94BK8zi6M7MEPUnHEQ85o8ZK90M/JD0eoBqVK27NXYKskeaK1HOlAjnao/LHMIhZKjEyTFCtB3U3MV5GleFM4Lw0TDUmlE3pGAeWRlSi9rLlwXNyYZWAhLGyFRmyVNcnMiq1nknfdkpqJvq3txD/8gapCa+8jEdJajBiq0VhKoiJyeJ7EnCFzIiZJZQpbm8lbEIVZcZmVLIhfH9K/ifdRq3erDVuG5VWNY+jCGdwDlWowyW04Aba0AEGEh7hGV4c5Tw5r87bqrXg5DOn8APO+xdSPZAE</latexit>

Grain Orientation Grain Orientation

0.2 <latexit sha1_base64="rq2y4Tr2M7GS3S5kcpsQs6ftVl8=">AAAB6XicbVA9SwNBEJ3zM8avqKXNYhSswl0KtYyksYxiPiA5wt5mLlmyt3fs7gnhyD+wsVDE1n9k579xk1yhiQ8GHu/NMDMvSATXxnW/nbX1jc2t7cJOcXdv/+CwdHTc0nGqGDZZLGLVCahGwSU2DTcCO4lCGgUC28G4PvPbT6g0j+WjmSToR3QoecgZNVZ6qN/2S2W34s5BVomXkzLkaPRLX71BzNIIpWGCat313MT4GVWGM4HTYi/VmFA2pkPsWipphNrP5pdOyYVVBiSMlS1pyFz9PZHRSOtJFNjOiJqRXvZm4n9eNzXhjZ9xmaQGJVssClNBTExmb5MBV8iMmFhCmeL2VsJGVFFmbDhFG4K3/PIqaVUr3lWlel8t187zOApwCmdwCR5cQw3uoAFNYBDCM7zCmzN2Xpx352PRuubkMyfwB87nDxZDjPk=</latexit> <latexit sha1_base64="8TCsve5hWMDXC5fOvtgfSr9ukOE=">AAAB6XicbVBNS8NAEJ3Ur1q/qh69LFbBU0l6UI8FQTxWsR/QhrLZTtqlm03Y3Qgl9B948aCIV/+RN/+N2zYHbX0w8Hhvhpl5QSK4Nq777RTW1jc2t4rbpZ3dvf2D8uFRS8epYthksYhVJ6AaBZfYNNwI7CQKaRQIbAfjm5nffkKleSwfzSRBP6JDyUPOqLHSQ+O2X664VXcOskq8nFQgR6Nf/uoNYpZGKA0TVOuu5ybGz6gynAmclnqpxoSyMR1i11JJI9R+Nr90Ss6tMiBhrGxJQ+bq74mMRlpPosB2RtSM9LI3E//zuqkJr/2MyyQ1KNliUZgKYmIye5sMuEJmxMQSyhS3txI2oooyY8Mp2RC85ZdXSatW9S6rtftapX6Wx1GEEziFC/DgCupwBw1oAoMQnuEV3pyx8+K8Ox+L1oKTzxzDHzifPzGYjQs=</latexit>
0.2 <latexit sha1_base64="rq2y4Tr2M7GS3S5kcpsQs6ftVl8=">AAAB6XicbVA9SwNBEJ3zM8avqKXNYhSswl0KtYyksYxiPiA5wt5mLlmyt3fs7gnhyD+wsVDE1n9k579xk1yhiQ8GHu/NMDMvSATXxnW/nbX1jc2t7cJOcXdv/+CwdHTc0nGqGDZZLGLVCahGwSU2DTcCO4lCGgUC28G4PvPbT6g0j+WjmSToR3QoecgZNVZ6qN/2S2W34s5BVomXkzLkaPRLX71BzNIIpWGCat313MT4GVWGM4HTYi/VmFA2pkPsWipphNrP5pdOyYVVBiSMlS1pyFz9PZHRSOtJFNjOiJqRXvZm4n9eNzXhjZ9xmaQGJVssClNBTExmb5MBV8iMmFhCmeL2VsJGVFFmbDhFG4K3/PIqaVUr3lWlel8t187zOApwCmdwCR5cQw3uoAFNYBDCM7zCmzN2Xpx352PRuubkMyfwB87nDxZDjPk=</latexit> <latexit sha1_base64="8TCsve5hWMDXC5fOvtgfSr9ukOE=">AAAB6XicbVBNS8NAEJ3Ur1q/qh69LFbBU0l6UI8FQTxWsR/QhrLZTtqlm03Y3Qgl9B948aCIV/+RN/+N2zYHbX0w8Hhvhpl5QSK4Nq777RTW1jc2t4rbpZ3dvf2D8uFRS8epYthksYhVJ6AaBZfYNNwI7CQKaRQIbAfjm5nffkKleSwfzSRBP6JDyUPOqLHSQ+O2X664VXcOskq8nFQgR6Nf/uoNYpZGKA0TVOuu5ybGz6gynAmclnqpxoSyMR1i11JJI9R+Nr90Ss6tMiBhrGxJQ+bq74mMRlpPosB2RtSM9LI3E//zuqkJr/2MyyQ1KNliUZgKYmIye5sMuEJmxMQSyhS3txI2oooyY8Mp2RC85ZdXSatW9S6rtftapX6Wx1GEEziFC/DgCupwBw1oAoMQnuEV3pyx8+K8Ox+L1oKTzxzDHzifPzGYjQs=</latexit>

CA
CA-Wider PF-Wider PF CA
CA-Wider PF-Wider PF
0.18 0.18
area fraction

0.16 0.16

0.14 0.14
Relative Population
Relative Population

0.12 0.12

0.1 0.1
Relative

0.08 0.08

0.06 0.06

0.04 0.04
<latexit sha1_base64="bAo6jjTrszvYSVpKVPMgRK/sDaM=">AAAB/3icbVC7SgNBFL0bXzG+ooKNzWAUrMJuCrUM2FhGMTGQLGF2cjcZMvtgZjYQ1hT+io2FIrb+hp1/42yyhSYeGDiccw/3zvFiwZW27W+rsLK6tr5R3Cxtbe/s7pX3D1oqSiTDJotEJNseVSh4iE3NtcB2LJEGnsAHb3Sd+Q9jlIpH4b2exOgGdBBynzOqjdQrH92hMHSMhJoY8SVlc6NiV+0ZyDJxclKBHI1e+avbj1gSYKiZoEp1HDvWbkql5kzgtNRNFMaUjegAO4aGNEDlprP7p+TMKH3iR9K8UJOZ+juR0kCpSeCZyYDqoVr0MvE/r5No/8pNeRgnGkM2X+QnguiIZGWQPpfItJgYQpnk5lbChjTrwFRWMiU4i19eJq1a1bmo1m5rlfppXkcRjuEEzsGBS6jDDTSgCQwe4Rle4c16sl6sd+tjPlqw8swh/IH1+QPEGJXZ</latexit>

0.02 0.02

0 0
0 9 18 27 36 45 54 63 72 81 90 0 9 18 27 36 45 54 63 72 81 90
Deeper
<latexit sha1_base64="y30tNdVjYgSl5DZ0eFJxfDGp0Tg=">AAAB8XicbVA9SwNBEJ2LXzF+RS1tFoOQKtyliJYBLSwjmA9MjrC3mUuW7O0du3tCCPkXNhaK2Ppv7Pw37iVXaPTBwOO9GWbmBYng2rjul1PY2Nza3inulvb2Dw6PyscnHR2nimGbxSJWvYBqFFxi23AjsJcopFEgsBtMrzO/+4hK81jem1mCfkTHkoecUWOlh0EQkhvEBNWwXHFrDTcDcWvuGvFyUoEcrWH5czCKWRqhNExQrfuemxh/TpXhTOCiNEg1JpRN6Rj7lkoaofbny4sX5MIqIxLGypY0ZKn+nJjTSOtZFNjOiJqJXvcy8T+vn5rwyp9zmaQGJVstClNBTEyy98mIK2RGzCyhTHF7K2ETqigzNqSSDcFbf/kv6dRrXqNWv6tXmtU8jiKcwTlUwYNLaMIttKANDCQ8wQu8Otp5dt6c91VrwclnTuEXnI9vAvCQaA==</latexit>

Grain Orientation Grain Orientation

0.18 <latexit sha1_base64="rq2y4Tr2M7GS3S5kcpsQs6ftVl8=">AAAB6XicbVA9SwNBEJ3zM8avqKXNYhSswl0KtYyksYxiPiA5wt5mLlmyt3fs7gnhyD+wsVDE1n9k579xk1yhiQ8GHu/NMDMvSATXxnW/nbX1jc2t7cJOcXdv/+CwdHTc0nGqGDZZLGLVCahGwSU2DTcCO4lCGgUC28G4PvPbT6g0j+WjmSToR3QoecgZNVZ6qN/2S2W34s5BVomXkzLkaPRLX71BzNIIpWGCat313MT4GVWGM4HTYi/VmFA2pkPsWipphNrP5pdOyYVVBiSMlS1pyFz9PZHRSOtJFNjOiJqRXvZm4n9eNzXhjZ9xmaQGJVssClNBTExmb5MBV8iMmFhCmeL2VsJGVFFmbDhFG4K3/PIqaVUr3lWlel8t187zOApwCmdwCR5cQw3uoAFNYBDCM7zCmzN2Xpx352PRuubkMyfwB87nDxZDjPk=</latexit> <latexit sha1_base64="8TCsve5hWMDXC5fOvtgfSr9ukOE=">AAAB6XicbVBNS8NAEJ3Ur1q/qh69LFbBU0l6UI8FQTxWsR/QhrLZTtqlm03Y3Qgl9B948aCIV/+RN/+N2zYHbX0w8Hhvhpl5QSK4Nq777RTW1jc2t4rbpZ3dvf2D8uFRS8epYthksYhVJ6AaBZfYNNwI7CQKaRQIbAfjm5nffkKleSwfzSRBP6JDyUPOqLHSQ+O2X664VXcOskq8nFQgR6Nf/uoNYpZGKA0TVOuu5ybGz6gynAmclnqpxoSyMR1i11JJI9R+Nr90Ss6tMiBhrGxJQ+bq74mMRlpPosB2RtSM9LI3E//zuqkJr/2MyyQ1KNliUZgKYmIye5sMuEJmxMQSyhS3txI2oooyY8Mp2RC85ZdXSatW9S6rtftapX6Wx1GEEziFC/DgCupwBw1oAoMQnuEV3pyx8+K8Ox+L1oKTzxzDHzifPzGYjQs=</latexit>
0.18 <latexit sha1_base64="rq2y4Tr2M7GS3S5kcpsQs6ftVl8=">AAAB6XicbVA9SwNBEJ3zM8avqKXNYhSswl0KtYyksYxiPiA5wt5mLlmyt3fs7gnhyD+wsVDE1n9k579xk1yhiQ8GHu/NMDMvSATXxnW/nbX1jc2t7cJOcXdv/+CwdHTc0nGqGDZZLGLVCahGwSU2DTcCO4lCGgUC28G4PvPbT6g0j+WjmSToR3QoecgZNVZ6qN/2S2W34s5BVomXkzLkaPRLX71BzNIIpWGCat313MT4GVWGM4HTYi/VmFA2pkPsWipphNrP5pdOyYVVBiSMlS1pyFz9PZHRSOtJFNjOiJqRXvZm4n9eNzXhjZ9xmaQGJVssClNBTExmb5MBV8iMmFhCmeL2VsJGVFFmbDhFG4K3/PIqaVUr3lWlel8t187zOApwCmdwCR5cQw3uoAFNYBDCM7zCmzN2Xpx352PRuubkMyfwB87nDxZDjPk=</latexit> <latexit sha1_base64="8TCsve5hWMDXC5fOvtgfSr9ukOE=">AAAB6XicbVBNS8NAEJ3Ur1q/qh69LFbBU0l6UI8FQTxWsR/QhrLZTtqlm03Y3Qgl9B948aCIV/+RN/+N2zYHbX0w8Hhvhpl5QSK4Nq777RTW1jc2t4rbpZ3dvf2D8uFRS8epYthksYhVJ6AaBZfYNNwI7CQKaRQIbAfjm5nffkKleSwfzSRBP6JDyUPOqLHSQ+O2X664VXcOskq8nFQgR6Nf/uoNYpZGKA0TVOuu5ybGz6gynAmclnqpxoSyMR1i11JJI9R+Nr90Ss6tMiBhrGxJQ+bq74mMRlpPosB2RtSM9LI3E//zuqkJr/2MyyQ1KNliUZgKYmIye5sMuEJmxMQSyhS3txI2oooyY8Mp2RC85ZdXSatW9S6rtftapX6Wx1GEEziFC/DgCupwBw1oAoMQnuEV3pyx8+K8Ox+L1oKTzxzDHzifPzGYjQs=</latexit>

CA
CA-Deeper PF-Deeper PF CA
CA-Deeper PF-Deeper PF
0.16 0.16
area fraction

0.14 0.14

0.12 0.12
Relative Population
Relative Population

0.1 0.1

0.08 0.08
Relative

0.06 0.06

0.04 0.04
<latexit sha1_base64="bAo6jjTrszvYSVpKVPMgRK/sDaM=">AAAB/3icbVC7SgNBFL0bXzG+ooKNzWAUrMJuCrUM2FhGMTGQLGF2cjcZMvtgZjYQ1hT+io2FIrb+hp1/42yyhSYeGDiccw/3zvFiwZW27W+rsLK6tr5R3Cxtbe/s7pX3D1oqSiTDJotEJNseVSh4iE3NtcB2LJEGnsAHb3Sd+Q9jlIpH4b2exOgGdBBynzOqjdQrH92hMHSMhJoY8SVlc6NiV+0ZyDJxclKBHI1e+avbj1gSYKiZoEp1HDvWbkql5kzgtNRNFMaUjegAO4aGNEDlprP7p+TMKH3iR9K8UJOZ+juR0kCpSeCZyYDqoVr0MvE/r5No/8pNeRgnGkM2X+QnguiIZGWQPpfItJgYQpnk5lbChjTrwFRWMiU4i19eJq1a1bmo1m5rlfppXkcRjuEEzsGBS6jDDTSgCQwe4Rle4c16sl6sd+tjPlqw8swh/IH1+QPEGJXZ</latexit>

0.02 0.02

0 0
0 9 18 <latexit sha1_base64="3J4zMX0dwnr2MszUFcr93Iksa+A=">AAAB+nicbVC7TsMwFL0pr1JeLYwsFgWJqUo6AGMlBhiLRB9SG1WO67RWHTuyHVAV+iksDCDEypew8Te4aQZoOdKVjs651773BDFn2rjut1NYW9/Y3Cpul3Z29/YPypXDtpaJIrRFJJeqG2BNORO0ZZjhtBsriqOA004wuZ77nQeqNJPi3kxj6kd4JFjICDZWGpQrNwozgaRiVJhcq7o1NwNaJV5OqpCjOSh/9YeSJJF9gHCsdc9zY+OnWBlGOJ2V+ommMSYTPKI9SwWOqPbTbPUZOrPKEIVS2RIGZerviRRHWk+jwHZG2Iz1sjcX//N6iQmv/JSJODFUkMVHYcKRkWieAxoyRYnhU0swUczuisgYK0yMTatkQ/CWT14l7XrNu6jV7+rVxmkeRxGO4QTOwYNLaMAtNKEFBB7hGV7hzXlyXpx352PRWnDymSP4A+fzB1r8k/k=</latexit>
27 36 45 54 63 72 81 90 0 9 18 <latexit sha1_base64="3J4zMX0dwnr2MszUFcr93Iksa+A=">AAAB+nicbVC7TsMwFL0pr1JeLYwsFgWJqUo6AGMlBhiLRB9SG1WO67RWHTuyHVAV+iksDCDEypew8Te4aQZoOdKVjs651773BDFn2rjut1NYW9/Y3Cpul3Z29/YPypXDtpaJIrRFJJeqG2BNORO0ZZjhtBsriqOA004wuZ77nQeqNJPi3kxj6kd4JFjICDZWGpQrNwozgaRiVJhcq7o1NwNaJV5OqpCjOSh/9YeSJJF9gHCsdc9zY+OnWBlGOJ2V+ommMSYTPKI9SwWOqPbTbPUZOrPKEIVS2RIGZerviRRHWk+jwHZG2Iz1sjcX//N6iQmv/JSJODFUkMVHYcKRkWieAxoyRYnhU0swUczuisgYK0yMTatkQ/CWT14l7XrNu6jV7+rVxmkeRxGO4QTOwYNLaMAtNKEFBB7hGV7hzXlyXpx352PRWnDymSP4A+fzB1r8k/k=</latexit>
27 36 45 54 63 72 81 90
GrainGrain
orientation
Orientation Grain
Grainorientation
Orientation

Figure 5.10: Distribution of grain orientations, averaged over five simulations

for three different melt pool shapes, namely the reference case fitted to FE
results (top row), a wider and shallower melt pool (central row), and a deeper
and narrower melt pool (bottom row), comparing PF and CA results for a fine
(hCA /hPF = 3, left column) and a coarse (hCA /hPF = 144, right column) CA
grid. (PF results are the same in left and right columns, only duplicated for
readability).

106
 5.2 Results and discussion

ture, Fig. 5.10 shows clear trends in this direction when using PF results, which
are thought to be the most accurate. Particularly, the wide and shallow melt
pools (center row) clearly show a low density of ⟨11⟩ (α ≈ 45◦ ) oriented grains
and a stronger ⟨10⟩ (α ≈ 0◦ ) texture compared to the reference (top row) and
deeper/narrower (bottom row) melt pools.

The agreement between PF and CA is often quite good for the reference melt
pool shape (Fig. 5.10, top row), which has the aspect ratio that is closest to unity
(w0 /d0 = 1.33). This is true regardless of the size of the CA grid. The typical grain
distribution in such a melt pool is shown in Figure 5.11. The main direction of the
temperature gradient across the melt pool is clearly followed by the grain orienta-
tion that was used for this simulation. As a consequence, the sides and center of
the melt pool have a ⟨10⟩ texture (expressed with respect to the build direction,
in this case vertical), with intermediate orientations smoothly changing along the
width of the melt pool. Both of the CA simulations accurately reflect this pattern.
In general, the following are the main differences between PF and CA simulations:
(1) the early planar growth region in the PF simulation is present just above the
fusion line, (2) the early elimination of grains with the least favorable orientation
in PF that gradually develop in CA simulations (as indicated in Section 5.2.1),
and (3) the grain boundaries in PF simulations have a significantly higher level of
roughness, while all GBs from CA simulations are essentially straight lines. The
simulation is shown in Fig. 5.11 and this characteristic distribution of grain orien-
tations inside the melt pool is reflective of the pattern shown in all five simulations
(for completeness, all of them are shown at the end of this chapter in Fig.5.14).

The agreement between PF and CA is excellent for fine CA grids (left column)
in the wider (i.e. shallower) melt pool (Fig. 5.10, center row), with an aspect ratio
w0 /d0 = 5.33, but a significant discrepancy appears for coarser CA grids (right
column).

Figure 5.12 shows the selected grain maps for a particular initial grain distri-
bution, which are representative of the behavior seen in all five configurations (all
of them shown in Fig. 5.15). The temperature gradient in this case is oriented

107
5.2 Results and discussion

upwards. Therefore, it is not unexpected to see that most of the selected grains in
the PF simulations had orientations that are α = 90◦ ± 30◦ . The distribution of
grain orientations across the melt pool is more scattered with a coarser CA grid
(bottom row), and several grains with α ≈ 45◦ even show up within the solidified
melt pool. Similar to the previously mentioned situations, only in PF simulations,
which also exhibit notably rougher GBs, is the initial planar growth followed by

Figure 5.11: Grain orientation maps for the reference melt pool dimension given
by FE thermal simulation (w0 = 128 µm, d0 = 96 µm), predicted by phase field
(top row) compared to CA with fine (central row) and coarse (bottom row) grid.
The white dashed lines identify the melted region. Arrows at the top illustrate
the predominant selected grain orientations within the melt pool.

108
 5.2 Results and discussion

Figure 5.12: Grain orientation maps for a wide and shallow melt pool (w0 =
256 µm, d0 = 48 µm), predicted by phase field (top row) compared to CA with
fine (central row) and coarse (bottom row) grid. The white dashed lines identify
the melted region. Arrows at the top illustrate the predominant selected grain
orientations within the melt pool.

early elimination of unfavorably orientated grains (resulting in small square grains

just above the fusion line) observed.

The agreement between CA and PF grain orientation densities is significantly

worse for the deeper (i.e. narrower) melt pools (Fig. 5.10, bottom row), and
it only slightly improves with lower hCA . Figure 5.13 illustrates that both PF
and CA qualitatively match fairly well in terms of texture distribution within the
melt pool, corresponding with the primary temperature gradient direction within,
despite a considerably higher discrepancy in their relative amounts. While the
right side of the melt pool has a significant density of grain with an orientation
of 55◦ ≤ α ≤ 90◦ , the left side is mostly composed of grains with orientation
0◦ ≤ α ≤ 35◦ . In the bottom part of Fig. 5.13, grain orientations that fall between
25◦ and 35◦ (red) and between 65◦ and 75◦ (green) are highlighted, illustrating the
left-right segregation of grain orientations. The GBs in the deeper melt pool are
more curved than those in the reference and wide melt pool cases because they
follow the rotation of the main temperature gradient direction during the melt
pool solidification. In contrast, the GBs in the reference and wide melt pool cases
are essentially straight lines. This general rotation of the grain upwards from the

109
5.3 Conclusions and perspectives

Figure 5.13: Grain orientation maps for a deep and narrow melt pool (w0 =
64 µm, d0 = 192 µm), predicted by phase field (left column) compared to CA
with fine (central column) and coarse (right column) grid. The white dashed
lines identify the melted region. The bottom row highlights grains with orien-
tations 25◦ ≤ α ≤ 35◦ (red) and 65◦ ≤ α ≤ 75◦ (green).

horizontal also occurs in the PF simulation, although the morphologies of the GBs
are rougher than in CA simulations. This behavior is representative of what was
seen in each of the five simulated deep melt pools (see Fig. 5.16).

5.3 Conclusions and perspectives

In this study, we systematically compared the grain microstructures from the

polycrystalline solidification of a melt pool that is predicted by phase-field and
cellular automaton techniques. We studied 2D simulations of solidification in the

110
 5.3 Conclusions and perspectives

longitudinal direction [79], as well as different melt pool shapes in the cross-section
normal to the scanning direction, using a binary Ni-5wt%Nb alloy as the model
alloy and a thermal field calculated by finite elements thermal simulations for an
Inconel 718 alloy [79].

Our simulations showed that when the CA grid is refined, the accurate match-
ing of grain structures — that is, comparing specific grain and GB morphologies
— essentially improves. The best CA results, when comparing/integrating grain
orientation maps point-by-point, capture the PF grain maps over, at most, 70% of
the melted region. The statistical distribution of crystal orientations in the whole
melt pools, averaged over several runs, is predicted quite well by CA simulations
using finer grids, even when specific details at the scale of individual grains and
GBs are not accurately reproduced. We found that the deeper melt pools had
worse matching, which we attribute to the stronger change in temperature profile
and gradient directions during solidification. This suggests that the agreement
may also be influenced by the shape of the melt pool.

As most grain boundaries tend to follow the classical favorably-oriented grain

criterion [65], [79], converging and low angle GBs’ trajectories are usually (but not
always) well captured when looking at the detailed GB scale. High-angle divergent
GBs have a greater disparity.

Regarding the transient growth regime and the morphology of the solid-liquid
interface, there are some striking differences between PF and CA. In fact, PF
simulations consistently result in an initial transient solidification period with a
nearly planar solid-liquid interface until its velocity increases enough to cause a
morphological destabilization into cells or dendrites. CA simulations, in which the
grains compete right away after the simulation begins, do not capture this (i.e.
grains compete directly from the fusion line). In PF simulations, the transient
destabilization of the planar interface into dendrites also results in intense side-
branching activity. From this activity, the least favorably oriented grains – those
with higher misorientation of their preferred growth direction with the tempera-
ture gradient – are frequently quickly eliminated by favorably oriented neighbors.

111
5.3 Conclusions and perspectives

Simulations using CA do not capture the dynamics of sidebranching that occurs

once the interface becomes morphologically unstable. Additionally, compared to
the GBs emerging from PF simulations, the CA-predicted GBs are significantly
smoother — most often straight lines.

Regarding how the shape of the melt pool affects the distribution of grain orien-
tations, melt pools with an aspect ratio close to unity (i.e., approximately circular)
show good agreement between the CA and PF statistical distribution of grain ori-
entations as well as their distribution within the melt pool. For melt pools with
greater aspect ratios, the CA-PF agreement decreases and becomes more grid-
dependent, although the distribution of grain orientations overall within the melt
pool continues to be reasonable. The statistical sample from our 15 configurations
was insufficient to clearly show how the melt pool aspect ratio influences the tran-
sition from ⟨100⟩ to ⟨110⟩ [144]. However, PF and to a lesser extent fine-grid CA
simulations indicate that such experimentally observed transition [144], [145] can
be predicted, should the number of simulations (and thus grains) be increased to
a larger and more statistically significant amount.

Furthermore, neither the considered CA neighborhood algorithm (Moore or

von Neumann) nor the dendritic growth kinetics law V (∆T ) applied at the grain
vertices showed a strong effect on the simulation results. This latter observation is
attributed to the high-temperature gradient, which reduces the distance between
the solidus and liquidus isotherms and, as a result, reduces the influence of the
growth law on the location of the solidification front between TL and TS .

It is important to point out some of the study’s limitations, some of which are
the subject of follow-up research. First, solidification simulations presented here,
while based on the 3D calculation of the temperature field, are two-dimensional,
thus not capturing the full complexity of grain growth competition in an actual
melt pool. However, the current study would not be possible in 3D at this time
since the grid size needed for well-converged PF simulations already results in
relatively expensive 2D simulations, which are handled here using an advanced
multi-GPU parallelized implementation. Fluid flow and nucleation, two additional

112
 5.3 Conclusions and perspectives

significant phenomenon that occur during melt pool solidification, were ignored.
The latter would need special attention because nucleation parameters (such as
undercooling and site density) are most often used as calibration parameters in
both PF and CA simulations. Finally, more advanced CA formulations, e.g. using
irregular quadrangles instead of square building blocks [65], [138], [139], could also
have a key influence, considering the high-temperature gradients encountered in
AM-relevant conditions.

In spite of these limitations, we trust that the presented results bring some use-
ful insight into the applicability and limitations of both modeling methods. They
clearly show that the accuracy of PF is still required to precisely capture transient
growth kinetics and solid-liquid interface morphological evolution while consider-
ing microstructural details at the scale of individual grains or GBs. Instead of the
classical Ivantsov- or KGT-based ones, more advanced growth kinetics [139] may
improve the prediction of transient growth regimes in CA simulations. However,
the level of detail provided by PF by itself is unquestionably necessary for sim-
ulating the morphological stability of a solid-liquid interface. This is even more
essential in AM-relevant conditions, which can result in “absolute stability” [146],
[147], restabilization of a planar interface at high growth velocity, which has been
shown in a variety of rapidly solidified metals in the form of banded microstruc-
tures [148]–[150]. However, our results demonstrate that CA-based methods are
capable of statistically reasonable grain texture predictions which could be useful,
for example, to predict averaged anisotropic mechanical properties over representa-
tive volume elements [151]. The significant speedup and upscaling provided by CA
modeling offers a very promising path for the prediction of actual three-dimensional
AM microstructures while similar quantitative PF simulations in 3D are currently
computationally out of reach.

113
 5.3 Conclusions and perspectives
PF (hPF = 10 nm)
PF1 CA (hPFCA1-0003
/hCA = 3) CA (hPF /hCA = 144)
CA1-0144
9.7x10-5 9.7x10-5 90 9.7x10-5 90 90

60 60 60

y (m)

y (m)

y (m)
30 30 30

PF2 CA2-0003 CA2-0144

0 0 0 0 0 0
9.7x10-5 0 9.7x10 -5 0 90
0.000257 -5 0 90
0.000257
9.7x10 0.000257 90
x (m) x (m) x (m)

60 y (m) 60 60
y (m)

y (m)

y (m)
9.7x10-5
30 30 30

0
PF3 CA3-0003 CA3-0144
0 0 0 0 0 0

0
9.7x10-5 0 9.7x10 -5 0 90
0.000257 -5 0 90
0.000257
9.7x10 0.000257 90
x (m) x (m) x (m)

60 60 60
y (m)

y (m)

y (m)
114

30 30 30

PF4 CA4-0003 CA4-0144

0 0 0 0 0 0
9.7x10-5 0 9.7x10 -5 0 90
0.000257 -5 0 90
0.000257
9.7x10 0.000257 90
x (m) x (m) x (m)

60 60 60

CA5-0144
y (m)

y (m)

y (m)
x (m)
30 30 30

PF5 CA5-0003 CA5-0144

0 0 0 0 0 0
9.7x10-5 0 9.7x10 -5 0 90
0.000257 -5 0 90
0.000257
9.7x10 0.000257 90
x (m) x (m) x (m)

60 60 60
y (m)

y (m)

y (m)
30 30 30
0.000257

0 0 0 0 0 0
0 0.000257
0 0.000257
0 0.000257
x (m) x (m) x (m)
0

30

60

90

Figure 5.14: Simulations of cross-section melt pool solidification for the reference case (i.e. thermal field as extracted
from FE results) for five different initial grain distributions (rows) as predicted by phase-field (left column), and
cellular automaton with fine (central column) and coarse (right column) spatial discretization. Domain size ≈
257 µm × 97 µm.
 y (m)

9.7x10-5
0
0

CA5-0144
x (m)
PF (hPF = 10 nm)
PF1 CA (hPFCA1-0003
/hCA = 3) CA (hPF /hCA = 144)
CA1-0144
4.9x10-5 4.9x10-5 90 4.9x10-5 90 90
y (m) y (m) y (m) y (m) y (m)

y (m) y (m) y (m) y (m) y (m)

y (m) y (m) y (m) y (m) y (m)

60 60 60
PF2 30 CA2-0003 30 CA2-0144 30
0
4.9x10-5 0 4.9x10 0
-5
0.000513
0 0
90 4.9x10 0
-5
0.000513
0 0
90 0.000513 0
90
xPF3
(m) 60 x (m) 60 x (m) 60
30 CA3-0003 30 CA3-0144 30
0
4.9x10-5 0 4.9x10 0
-5
0.000513
0 0
90 4.9x10 0
-5
0.000513
0 0
90 0.000513 0
90
xPF4
(m) 60 x (m) 60 x (m) 60
30 CA5-0144 30 CA4-0144 30
0
4.9x10-5 0 4.9x10 0
-5
0.000513
0 0
90 4.9x10 0
-5
0.000513
0 0
90 0.000513 0
90
xPF5
(m) 60 x (m) 60 x (m) 60
30 CA5-0003 30 CA5-0144 30
0
4.9x10-5 0 4.9x10 0
-5
0.000513
0 0
90 4.9x10 0
-5
0.000513
0 0
90 0.000513 0
90
60 60 60
115

x (m) x (m) x (m)

30 30 30

0.000257
0 0 0
0.000513
0 0 0
0.000513
0 0 0.000513 0
x (m) x (m) x (m)

5.3 Conclusions and perspectives

0

30

60

90
Figure 5.15: Simulations of cross-section melt pool solidification for the reference case (i.e. thermal field as extracted
from FE results) for five different initial grain distributions (rows) as predicted by phase-field (left column), and
cellular automaton with fine (central column) and coarse (right column) spatial discretization. Domain size ≈
513 µm × 49 µm.
5.3 Conclusions and perspectives

PF (hPF = 10 nm)
PF1
CA (hPF /hCA = 3) CA1-0003
CA (hPF /hCA = 144) CA1-0144
0.000193 90 0.000193 90 0.000193 90

60 60 60
y (m)

y (m)

y (m)
30 30 30

0 PF2 0 0 CA2-0003 0 0 CA2-0144 0

0.0001930 0.00012990 0.0001930 0.00012990 0.0001930 0.00012990
x (m) x (m) x (m)

60 60 60
y (m)

y (m)

y (m)
30 30 30

0 PF3 0 0 CA3-0003 0 0 CA3-0144 0

0.0001930 0.00012990 0.0001930 0.00012990 0.0001930 0.00012990
x (m) y (m)
x (m) x (m)
9.7x10-5

60 60 60
y (m)

y (m)

y (m)
0
0

30 30 30

0 PF4 0 0 CA4-0003 0 0 CA4-0144 0

0.0001930 0.00012990 0.0001930 0.00012990 0.0001930 0.00012990
x (m) x (m) x (m)

60 60 60
y (m)

y (m)

y (m)

30 30 30
CA5-0144
x (m)

0 PF5 0 0 CA5-0003 0 0 CA5-0144 0

0.0001930 0.00012990 0.0001930 0.00012990 0.0001930 0.00012990
x (m) x (m) x (m)

60 60 60
y (m)

y (m)

y (m)

30 30 30
0.000257

0 0 0 0 0 0
0 0.000129 0 0.000129 0 0.000129
x (m) x (m) x (m)
0

30

60

90

Figure 5.16: Simulations of cross-section melt pool solidification of the

deeper/narrower melt pool for five different initial grain distributions (rows)
as predicted by phase-field (left column), and cellular automaton with fine
(central column) and coarse (right column) spatial discretization. Domain size
≈ 129 µm × 193 µm.
116
Chapter 6
Summary and Perspectives

6.1 Summary

In this study, a multiscale modeling framework was presented for the simula-
tion of metallic alloy powder-bed fusion. The following methods are coupled in this
framework: (1) CalPhaD calculations, (2) 3D finite element thermal and thermo-
mechanical frameworks, (3) 2D phase-field model, and (4) 2D cellular automaton
model. The main achievements of the current study are:

• The CalPhaD calculations of temperature-dependent properties and phase

diagram, was used to calculate alloy properties.

• An in-house finite element code was used for three-dimensional thermal and
thermomechanical SLM process simulations. The results show that in order
to predict the size and shape of the melt pool, it is important to consider
temperature-dependent parameters as well as applying specific properties in
the powder bed and dense regions.

• 2D phase-field simulations of the microstructure development due to the so-

lidification of polycrystalline materials in the melt pool scale were developed.

117
6.2 Perspectives

We highlighted similarities and differences with equivalent simulations that

are usually performed on a smaller subset of the melt pool.

• A 2D cellular automaton model was implemented to simulate the develop-

ment of grain microstructures during the polycrystalline solidification of an
entire melt pool. The results were compared against the phase-field results,
focusing on longitudinal section as well as different melt pool shapes in the
cross-section normal to the laser’s scanning direction.

6.2 Perspectives

In the context of Integrated Computational Materials Engineering (ICME) for

powder-bed fusion processes, this work arguably represents a significant step for-
ward. It nonetheless has certain limitations, most of which point to potential
future directions — some of which are highlighted below.

• The coupling of thermomechanics, which includes fluid dynamics and plas-

ticity, is the next step for macroscopic simulations. The extension of the
method to powder-bed melting in keyhole mode would be possible with fluid
flow simulation, which would also enable defect formation prediction. Ad-
ditionally, it will allow predicting the important aspects of print quality,
including residual stresses and part distortion.

• The pseudo-binary alloy approximation, the absence of solute trapping, and

the lack of solid-state microstructure growth are the key limitations of the
microstructure PF simulations. The use of dedicated models is required for
the extension to multicomponent alloys or solute trapping. Nucleation is
another important factor to be considered since it affects how much grain
structure is columnar or equiaxed.

• Due to the two-dimensional simulations, the pseudo-binary alloy approxima-

tion, and the fact that a statistical (high-throughput) exploration would be

118
 6.2 Perspectives

necessary to extract statistically relevant trends and conclusions, the result-

ing microstructure analysis provided here remains semi-quantitative and it
does not properly represent the complexity of grain growth competition in a
real melt pool. Therefore, one of the key tasks for the future could be to de-
velop a reliable 3D model (most likely using CA technique) that can predict
how microstructures would develop during solidification of melt pool. In this
respect, due to the high-temperature gradients seen in AM conditions, more
advanced CA formulations, such as employing irregular quadrangles rather
than square building blocks may have a significant effect.

119
120
Bibliography

[1] B. Schoinochoritis, D. Chantzis, and K. Salonitis, “Simulation of metal-

lic powder bed additive manufacturing processes with the finite element
method: A critical review,” Proceedings of the Institution of Mechanical
Engineers, Part B: Journal of Engineering Manufacture, vol. 231, no. 1,
pp. 96–117, 2017.
[2] B. Vrancken, “Study of residual stresses in selective laser melting,” KU
Leuven, 2016.
[3] C. Y. Yap, C. K. Chua, Z. L. Dong, et al., “Review of selective laser melt-
ing: Materials and applications,” Applied Physics Reviews, vol. 2, no. 4,
p. 041 101, 2015.
[4] N. Contuzzi, S. Campanelli, and A. Ludovico, “3d finite element analysis
in the selective laser melting process,” International Journal of Simulation
Modelling, vol. 10, no. 3, pp. 113–122, 2011.
[5] Y. Zhang, G. Guillemot, M. Bernacki, and M. Bellet, “Macroscopic thermal
finite element modeling of additive metal manufacturing by selective laser
melting process,” Computer Methods in Applied Mechanics and Engineer-
ing, vol. 331, pp. 514–535, 2018.
[6] A. Hussein, L. Hao, C. Yan, and R. Everson, “Finite element simulation of
the temperature and stress fields in single layers built without-support in
selective laser melting,” Materials & Design (1980-2015), vol. 52, pp. 638–
647, 2013.

121
BIBLIOGRAPHY

[7] C. Fu and Y. Guo, “3-dimensional finite element modeling of selective laser

melting ti-6al-4v alloy,” in 2014 International Solid Freeform Fabrication
Symposium, University of Texas at Austin, 2014, pp. 1129–1144.
[8] I. A. Roberts, C. Wang, R. Esterlein, M. Stanford, and D. Mynors, “A three-
dimensional finite element analysis of the temperature field during laser
melting of metal powders in additive layer manufacturing,” International
Journal of Machine Tools and Manufacture, vol. 49, no. 12-13, pp. 916–
923, 2009.
[9] I. A. Roberts, “Investigation of residual stresses in the laser melting of
metal powders in additive layer manufacturing,” International Journal of
Simulation Modelling, 2012.
[10] Y. Huang, L. Yang, X. Du, and Y. Yang, “Finite element analysis of ther-
mal behavior of metal powder during selective laser melting,” International
Journal of Thermal Sciences, vol. 104, pp. 146–157, 2016.
[11] Q. Chen, G. Guillemot, C.-A. Gandin, and M. Bellet, “Three-dimensional
finite element thermomechanical modeling of additive manufacturing by se-
lective laser melting for ceramic materials,” Additive Manufacturing, vol. 16,
pp. 124–137, 2017.
[12] R. Li, Y. Shi, J. Liu, H. Yao, and W. Zhang, “Effects of processing pa-
rameters on the temperature field of selective laser melting metal powder,”
Powder Metallurgy and Metal Ceramics, vol. 48, no. 3-4, pp. 186–195, 2009.
[13] D. Zhang, Q. Cai, J. Liu, L. Zhang, and R. Li, “Select laser melting of w–ni–
fe powders: Simulation and experimental study,” The International Journal
of Advanced Manufacturing Technology, vol. 51, no. 5-8, pp. 649–658, 2010.
[14] D. Riedlbauer, M. Drexler, D. Drummer, P. Steinmann, and J. Mergheim,
“Modelling, simulation and experimental validation of heat transfer in selec-
tive laser melting of the polymeric material pa12,” Computational Materials
Science, vol. 93, pp. 239–248, 2014.

122
 BIBLIOGRAPHY

[15] L. E. Criales, Y. M. Arısoy, and T. Özel, “Sensitivity analysis of material

and process parameters in finite element modeling of selective laser melt-
ing of inconel 625,” The International Journal of Advanced Manufacturing
Technology, vol. 86, no. 9-12, pp. 2653–2666, 2016.
[16] M. Chiumenti, E. Neiva, E. Salsi, et al., “Numerical modelling and ex-
perimental validation in selective laser melting,” Additive Manufacturing,
vol. 18, pp. 171–185, 2017.
[17] T. Bartel, I. Guschke, and A. Menzel, “Towards the simulation of selective
laser melting processes via phase transformation models,” Computers &
Mathematics with Applications, 2018.
[18] Y. Li, K. Zhou, P. Tan, S. B. Tor, C. K. Chua, and K. F. Leong, “Modeling
temperature and residual stress fields in selective laser melting,” Interna-
tional Journal of Mechanical Sciences, vol. 136, pp. 24–35, 2018.
[19] Y. Li and D. Gu, “Parametric analysis of thermal behavior during selective
laser melting additive manufacturing of aluminum alloy powder,” Materials
& Design, vol. 63, pp. 856–867, 2014.
[20] ——, “Thermal behavior during selective laser melting of commercially pure
titanium powder: Numerical simulation and experimental study,” Additive
Manufacturing, vol. 1, pp. 99–109, 2014.
[21] S. Roy, “Heat transfer model and finite element formulation for simula-
tion of selective laser melting additive manufacturing,” Ph.D. dissertation,
Rensselaer Polytechnic Institute, 2017.
[22] M. Matsumoto, M. Shiomi, K. Osakada, and F. Abe, “Finite element anal-
ysis of single layer forming on metallic powder bed in rapid prototyping
by selective laser processing,” International Journal of Machine Tools and
Manufacture, vol. 42, no. 1, pp. 61–67, 2002.
[23] L. Wang, X. Jiang, Y. Zhu, X. Zhu, J. Sun, and B. Yan, “An approach to
predict the residual stress and distortion during the selective laser melting
of alsi10mg parts,” The International Journal of Advanced Manufacturing
Technology, vol. 97, no. 9-12, pp. 3535–3546, 2018.

123
BIBLIOGRAPHY

[24] T. Mukherjee, W. Zhang, and T. DebRoy, “An improved prediction of resid-

ual stresses and distortion in additive manufacturing,” Computational Ma-
terials Science, vol. 126, pp. 360–372, 2017.
[25] P. Conti, F. Cianetti, and P. Pilerci, “Parametric finite elements model of
slm additive manufacturing process,” Procedia Structural Integrity, vol. 8,
pp. 410–421, 2018.
[26] M. F. Zaeh and G. Branner, “Investigations on residual stresses and defor-
mations in selective laser melting,” Production Engineering, vol. 4, no. 1,
pp. 35–45, 2010.
[27] L. Papadakis, A. Loizou, J. Risse, S. Bremen, and J. Schrage, “A compu-
tational reduction model for appraising structural effects in selective laser
melting manufacturing: A methodical model reduction proposed for time-
efficient finite element analysis of larger components in selective laser melt-
ing,” Virtual and Physical Prototyping, vol. 9, no. 1, pp. 17–25, 2014.
[28] C. Li, C. Fu, Y. Guo, and F. Fang, “Fast prediction and validation of
part distortion in selective laser melting,” Procedia Manufacturing, vol. 1,
pp. 355–365, 2015.
[29] ——, “A multiscale modeling approach for fast prediction of part distor-
tion in selective laser melting,” Journal of Materials Processing Technology,
vol. 229, pp. 703–712, 2016.
[30] C. Li, J. Liu, and Y. Guo, “Prediction of residual stress and part distortion
in selective laser melting,” Procedia CIRP, vol. 45, pp. 171–174, 2016.
[31] D. Gu and B. He, “Finite element simulation and experimental investiga-
tion of residual stresses in selective laser melted ti–ni shape memory alloy,”
Computational Materials Science, vol. 117, pp. 221–232, 2016.
[32] B. Cheng, S. Shrestha, and Y. K. Chou, “Stress and deformation evaluations
of scanning strategy effect in selective laser melting,” in ASME 2016 11th
International Manufacturing Science and Engineering Conference, Ameri-
can Society of Mechanical Engineers, 2016, V003T08A009–V003T08A009.

124
 BIBLIOGRAPHY

[33] C. Luo, J. Qiu, Y. Yan, J. Yang, C. Uher, and X. Tang, “Finite element
analysis of temperature and stress fields during the selective laser melting
process of thermoelectric snte,” Journal of Materials Processing Technology,
vol. 261, pp. 74–85, 2018.
[34] D. Pitassi, E. Savoia, V. Fontanari, et al., “Finite element thermal analysis
of metal parts additively manufactured via selective laser melting,” in Finite
Element Method-Simulation, Numerical Analysis and Solution Techniques,
IntechOpen, 2017, pp. 123–154.
[35] N. Hodge, R. Ferencz, and J. Solberg, “Implementation of a thermome-
chanical model for the simulation of selective laser melting,” Computational
Mechanics, vol. 54, no. 1, pp. 33–51, 2014.
[36] C. Bruna-Rosso, A. G. Demir, and B. Previtali, “Selective laser melting
finite element modeling: Validation with high-speed imaging and lack of
fusion defects prediction,” Materials & Design, vol. 156, pp. 143–153, 2018.
[37] B. Cheng and K. Chou, “Melt pool evolution study in selective laser melt-
ing,” in 26th Annual International Solid Freeform Fabrication Symposium-
An Additive Manufacturing Conference, Austin, TX, USA, 2015, pp. 1182–
1194.
[38] A. Foroozmehr, M. Badrossamay, E. Foroozmehr, et al., “Finite element
simulation of selective laser melting process considering optical penetration
depth of laser in powder bed,” Materials & Design, vol. 89, pp. 255–263,
2016.
[39] R. Andreotta, L. Ladani, and W. Brindley, “Finite element simulation of
laser additive melting and solidification of inconel 718 with experimen-
tally tested thermal properties,” Finite Elements in Analysis and Design,
vol. 135, pp. 36–43, 2017.
[40] L.-E. Loh, C.-K. Chua, W.-Y. Yeong, et al., “Numerical investigation and
an effective modelling on the selective laser melting (slm) process with
aluminium alloy 6061,” International Journal of Heat and Mass Transfer,
vol. 80, pp. 288–300, 2015.

125
BIBLIOGRAPHY

[41] M. M. Francois, A. Sun, W. E. King, et al., “Modeling of additive manufac-

turing processes for metals: Challenges and opportunities,” Current Opinion
in Solid State and Materials Science, vol. 21, no. 4, pp. 198–206, 2017.
[42] C. Körner, M. Markl, and J. A. Koepf, “Modeling and simulation of mi-
crostructure evolution for additive manufacturing of metals: A critical re-
view,” Metallurgical and Materials Transactions A, vol. 51, no. 10, pp. 4970–
4983, 2020.
[43] T. DebRoy, T. Mukherjee, H. Wei, J. Elmer, and J. Milewski, “Metallurgy,
mechanistic models and machine learning in metal printing,” Nature Re-
views Materials, vol. 6, no. 1, pp. 48–68, 2021.
[44] W. Yan, W. Ge, Y. Qian, et al., “Multi-physics modeling of single/multiple-
track defect mechanisms in electron beam selective melting,” Acta Materi-
alia, vol. 134, pp. 324–333, 2017.
[45] A. Otto and M. Schmidt, “Towards a universal numerical simulation model
for laser material processing,” Physics Procedia, vol. 5, pp. 35–46, 2010.
[46] S. A. Khairallah, A. T. Anderson, A. Rubenchik, and W. E. King, “Laser
powder-bed fusion additive manufacturing: Physics of complex melt flow
and formation mechanisms of pores, spatter, and denudation zones,” Acta
Materialia, vol. 108, pp. 36–45, 2016.
[47] C. Panwisawas, B. Perumal, R. M. Ward, et al., “Keyhole formation and
thermal fluid flow-induced porosity during laser fusion welding in titanium
alloys: Experimental and modelling,” Acta Materialia, vol. 126, pp. 251–
263, 2017.
[48] Y. Zhang, Q. Chen, G. Guillemot, C.-A. Gandin, and M. Bellet, “Numerical
modelling of fluid and solid thermomechanics in additive manufacturing by
powder-bed fusion: Continuum and level set formulation applied to track-
and part-scale simulations,” Comptes Rendus Mécanique, vol. 346, no. 11,
pp. 1055–1071, 2018.

126
 BIBLIOGRAPHY

[49] C. Tang, K. Le, and C. Wong, “Physics of humping formation in laser pow-
der bed fusion,” International Journal of Heat and Mass Transfer, vol. 149,
p. 119 172, 2020.
[50] L.-X. Lu, N. Sridhar, and Y.-W. Zhang, “Phase field simulation of powder
bed-based additive manufacturing,” Acta Materialia, vol. 144, pp. 801–809,
2018.
[51] Y. Yang, O. Ragnvaldsen, Y. Bai, M. Yi, and B.-X. Xu, “3d non-isothermal
phase-field simulation of microstructure evolution during selective laser sin-
tering,” Npj Computational Materials, vol. 5, no. 1, pp. 1–12, 2019.
[52] M. Yang, L. Wang, and W. Yan, “Phase-field modeling of grain evolutions in
additive manufacturing from nucleation, growth, to coarsening,” Npj Com-
putational Materials, vol. 7, no. 1, pp. 1–12, 2021.
[53] A. F. Chadwick and P. W. Voorhees, “The development of grain structure
during additive manufacturing,” Acta Materialia, vol. 211, p. 116 862, 2021.
[54] D. Tourret and A. Karma, “Growth competition of columnar dendritic
grains: A phase-field study,” Acta Materialia, vol. 82, pp. 64–83, 2015.
[55] D. Tourret, Y. Song, A. J. Clarke, and A. Karma, “Grain growth compe-
tition during thin-sample directional solidification of dendritic microstruc-
tures: A phase-field study,” Acta Materialia, vol. 122, pp. 220–235, 2017.
[56] T. M. Rodgers, J. D. Madison, and V. Tikare, “Simulation of metal additive
manufacturing microstructures using kinetic monte carlo,” Computational
Materials Science, vol. 135, pp. 78–89, 2017.
[57] H. Wei, J. Elmer, and T. DebRoy, “Crystal growth during keyhole mode
laser welding,” Acta Materialia, vol. 133, pp. 10–20, 2017.
[58] M. Rappaz and C.-A. Gandin, “Probabilistic modelling of microstructure
formation in solidification processes,”Acta Metallurgica et Materialia, vol. 41,
no. 2, pp. 345–360, 1993.

127
BIBLIOGRAPHY

[59] C.-A. Gandin and M. Rappaz, “A coupled finite element-cellular automa-

ton model for the prediction of dendritic grain structures in solidification
processes,” Acta Metallurgica et Materialia, vol. 42, no. 7, pp. 2233–2246,
1994.
[60] S. Chen, G. Guillemot, and C.-A. Gandin,“Three-dimensional cellular automaton-
finite element modeling of solidification grain structures for arc-welding pro-
cesses,” Acta Materialia, vol. 115, pp. 448–467, 2016.
[61] A. Rai, M. Markl, and C. Körner, “A coupled cellular automaton–lattice
boltzmann model for grain structure simulation during additive manufac-
turing,” Computational Materials Science, vol. 124, pp. 37–48, 2016.
[62] J. Koepf, D. Soldner, M. Ramsperger, J. Mergheim, M. Markl, and C.
Körner, “Numerical microstructure prediction by a coupled finite element
cellular automaton model for selective electron beam melting,” Computa-
tional Materials Science, vol. 162, pp. 148–155, 2019.
[63] Y. Lian, Z. Gan, C. Yu, D. Kats, W. K. Liu, and G. J. Wagner, “A cel-
lular automaton finite volume method for microstructure evolution during
additive manufacturing,” Materials & Design, vol. 169, p. 107 672, 2019.
[64] M. S. Mohebbi and V. Ploshikhin, “Implementation of nucleation in cellular
automaton simulation of microstructural evolution during additive manu-
facturing of al alloys,” Additive Manufacturing, vol. 36, p. 101 726, 2020.
[65] A. Pineau, G. Guillemot, D. Tourret, A. Karma, and C.-A. Gandin,“Growth
competition between columnar dendritic grains–cellular automaton versus
phase field modeling,” Acta Materialia, vol. 155, pp. 286–301, 2018.
[66] E. Dorari, K. Ji, G. Guillemot, C.-A. Gandin, and A. Karma, “Growth
competition between columnar dendritic grains–the role of microstructural
length scales,” Acta Materialia, vol. 223, p. 117 395, 2022.
[67] C.-A. Gandin and M. Rappaz, “A 3d cellular automaton algorithm for
the prediction of dendritic grain growth,” Acta Materialia, vol. 45, no. 5,
pp. 2187–2195, 1997.

128
 BIBLIOGRAPHY

[68] W. J. Boettinger, J. A. Warren, C. Beckermann, and A. Karma, “Phase-field

simulation of solidification,” Annual Review of Materials Research, vol. 32,
no. 1, pp. 163–194, 2002.
[69] L.-Q. Chen, “Phase-field models for microstructure evolution,” Annual Re-
view of Materials Research, vol. 32, no. 1, pp. 113–140, 2002.
[70] N. Moelans, B. Blanpain, and P. Wollants, “An introduction to phase-field
modeling of microstructure evolution,” Calphad, vol. 32, no. 2, pp. 268–294,
2008.
[71] I. Steinbach, “Phase-field models in materials science,” Modelling and Sim-
ulation in Materials Science and Engineering, vol. 17, no. 7, p. 073 001,
2009.
[72] ——, “Phase-field model for microstructure evolution at the mesoscopic
scale,” Annual Review of Materials Research, vol. 43, pp. 89–107, 2013.
[73] D. Tourret, H. Liu, and J. LLorca, “Phase-field modeling of microstructure
evolution: Recent applications, perspectives and challenges,” Progress in
Materials Science, vol. 123, p. 100 810, 2022.
[74] T. Keller, G. Lindwall, S. Ghosh, et al., “Application of finite element,
phase-field, and calphad-based methods to additive manufacturing of ni-
based superalloys,” Acta Materialia, vol. 139, pp. 244–253, 2017.
[75] V. Fallah, M. Amoorezaei, N. Provatas, S. Corbin, and A. Khajepour,
“Phase-field simulation of solidification morphology in laser powder deposi-
tion of ti–nb alloys,” Acta Materialia, vol. 60, no. 4, pp. 1633–1646, 2012.
[76] K. Karayagiz, L. Johnson, R. Seede, et al., “Finite interface dissipation
phase field modeling of ni–nb under additive manufacturing conditions,”
Acta Materialia, vol. 185, pp. 320–339, 2020.
[77] J. Berry, A. Perron, J.-L. Fattebert, et al., “Toward multiscale simulations
of tailored microstructure formation in metal additive manufacturing,” Ma-
terials Today, vol. 51, pp. 65–86, 2021.

129
BIBLIOGRAPHY

[78] F. Yu, Y. Wei, and X. Liu, “The evolution of polycrystalline solidification in

the entire weld: A phase-field investigation,” International Journal of Heat
and Mass Transfer, vol. 142, p. 118 450, 2019.
[79] S. Elahi, R. Tavakoli, A. Boukellal, T. Isensee, I. Romero, and D. Tourret,
“Multiscale simulation of powder-bed fusion processing of metallic alloys,”
Computational Materials Science, vol. 209, p. 111 383, 2022.
[80] D. Portillo, D. del Pozo, D. Rodrı́guez-Galán, J. Segurado, and I. Romero,
“Muesli-a material universal library,” Advances in Engineering Software,
vol. 105, pp. 1–8, 2017.
[81] V. Thomée, Galerkin finite element methods for parabolic problems, ser. Springer
Series in Computational Mathematics. Berlin, Heidelberg: Springer Science
& Business Media, 1997, vol. 25.
[82] W. Kurz and D. J. Fisher, Fundamentals of solidification. 1984.
[83] K. C. Mills, Recommended values of thermophysical properties for selected
commercial alloys. Woodhead Publishing, 2002.
[84] G. L. Knapp, N. Raghavan, A. Plotkowski, and T. Debroy, “Experiments
and simulations on solidification microstructure for inconel 718 in powder
bed fusion electron beam additive manufacturing,” Additive Manufacturing,
vol. 25, pp. 511–521, 2019.
[85] L. C. Wei, L. E. Ehrlich, M. J. Powell-Palm, C. Montgomery, J. Beuth, and
J. A. Malen, “Thermal conductivity of metal powders for powder bed ad-
ditive manufacturing,” Additive Manufacturing, vol. 21, pp. 201–208, 2018.
[86] F. Thümmler, R. Oberacker, et al., An introduction to powder metallurgy.
Institute of Materials London, 1993, vol. 490.
[87] J. Ning, W. Wang, X. Ning, D. E. Sievers, H. Garmestani, and S. Y. Liang,
“Analytical thermal modeling of powder bed metal additive manufacturing
considering powder size variation and packing,” Materials, vol. 13, no. 8,
p. 1988, 2020.

130
 BIBLIOGRAPHY

[88] S. S. Sih and J. W. Barlow, “The prediction of the emissivity and thermal
conductivity of powder beds,” Particulate Science and Technology, vol. 22,
no. 4, pp. 427–440, 2004.
[89] G. Damköhler, “Einflüsse der strömung, diffusion und des wärmeüberganges
auf die leistung von reaktionsöfen.: I. Allgemeine gesichtspunkte für die
übertragung eines chemischen prozesses aus dem kleinen ins große,”Zeitschrift
für Elektrochemie und Angewandte Physikalische Chemie, vol. 42, no. 12,
pp. 846–862, 1936.
[90] P. Zehner and E. Schlünder, “Wärmeleitfähigkeit von schüttungen bei mäßi-
gen temperaturen,” Chemie Ingenieur Technik, vol. 42, no. 14, pp. 933–941,
1970.
[91] K. Karayagiz, A. Elwany, G. Tapia, et al., “Numerical and experimental
analysis of heat distribution in the laser powder bed fusion of ti-6al-4v,”
IISE Transactions, vol. 51, no. 2, pp. 136–152, 2019.
[92] A. Iveković, M. L. Montero-Sistiaga, J. Vleugels, J.-P. Kruth, and K. Van-
meensel, “Crack mitigation in laser powder bed fusion processed hastelloy x
using a combined numerical-experimental approach,” Journal of Alloys and
Compounds, vol. 864, p. 158 803, 2021.
[93] Z. Xiang, M. Yin, G. Dong, X. Mei, and G. Yin, “Modeling of the thermal
physical process and study on the reliability of linear energy density for
selective laser melting,” Results in Physics, vol. 9, pp. 939–946, 2018.
[94] F. P. Beer, E. R. Johnston, J. T. DeWolf, and D. F. Mazurek, Statics and
mechanics of materials. McGraw-Hill Education New York, 2017.
[95] Y. A. Cengel, M. A. Boles, and M. Kanoğlu, Thermodynamics: an engi-
neering approach. McGraw-hill New York, 2011, vol. 5.
[96] S. M. Tawfik, M. N. Nasr, and H. A. El Gamal, “Finite element modelling
for part distortion calculation in selective laser melting,” Alexandria Engi-
neering Journal, 2018.

131
BIBLIOGRAPHY

[97] J. Eiken, B. Böttger, and I. Steinbach, “Multiphase-field approach for mul-

ticomponent alloys with extrapolation scheme for numerical application,”
Physical Review E, vol. 73, no. 6, p. 066 122, 2006.
[98] B. Nestler and A. Choudhury, “Phase-field modeling of multi-component
systems,” Current Opinion in Solid State and Materials Science, vol. 15,
no. 3, pp. 93–105, 2011.
[99] M. Ohno, “Quantitative phase-field modeling of nonisothermal solidifica-
tion in dilute multicomponent alloys with arbitrary diffusivities,” Physical
Review E, vol. 86, no. 5, p. 051 603, 2012.
[100] I. Steinbach, L. Zhang, and M. Plapp, “Phase-field model with finite inter-
face dissipation,” Acta Materialia, vol. 60, no. 6-7, pp. 2689–2701, 2012.
[101] L. Zhang and I. Steinbach, “Phase-field model with finite interface dissipa-
tion: Extension to multi-component multi-phase alloys,” Acta Materialia,
vol. 60, no. 6-7, pp. 2702–2710, 2012.
[102] J. Kundin, L. Mushongera, and H. Emmerich, “Phase-field modeling of mi-
crostructure formation during rapid solidification in inconel 718 superalloy,”
Acta Materialia, vol. 95, pp. 343–356, 2015.
[103] T. Pinomaa and N. Provatas, “Quantitative phase field modeling of solute
trapping and continuous growth kinetics in quasi-rapid solidification,” Acta
Materialia, vol. 168, pp. 167–177, 2019.
[104] S. Kavousi and M. A. Zaeem, “Quantitative phase-field modeling of solute
trapping in rapid solidification,” Acta Materialia, vol. 205, p. 116 562, 2021.
[105] S. G. Kim, W. T. Kim, P.-R. Cha, et al., “Phase-field model with relaxation
of the partition coefficient,” Computational Materials Science, vol. 188,
p. 110 184, 2021.
[106] B. Echebarria, R. Folch, A. Karma, and M. Plapp, “Quantitative phase-field
model of alloy solidification,” Physical Review E, vol. 70, no. 6, p. 061 604,
2004.

132
 BIBLIOGRAPHY

[107] K. Glasner, “Nonlinear preconditioning for diffuse interfaces,” Journal of

Computational Physics, vol. 174, no. 2, pp. 695–711, 2001.
[108] A. Karma and W. Rappel, “Quantitative phase-field modeling of dendritic
growth in two and three dimensions,” Physical Review E, vol. 57, no. 4,
p. 4323, 1998.
[109] I. Steinbach, F. Pezzolla, B. Nestler, et al., “A phase field concept for multi-
phase systems,” Physica D: Nonlinear Phenomena, vol. 94, no. 3, pp. 135–
147, 1996.
[110] R. Bridson, “Fast poisson disk sampling in arbitrary dimensions,” SIG-
GRAPH sketches, vol. 10, no. 1, 2007.
[111] Y. Zhang, L. Wu, X. Guo, et al., “Additive manufacturing of metallic mate-
rials: A review,” Journal of Materials Engineering and Performance, vol. 27,
no. 1, pp. 1–13, 2018.
[112] M. Walbrühl, A. Blomqvist, and P. A. Korzhavyi, “Atomic diffusion in
liquid nickel: First-principles modeling,” The Journal of Chemical Physics,
vol. 148, no. 24, p. 244 503, 2018.
[113] J. J. Hoyt, M. Asta, and A. Karma, “Atomistic and continuum modeling
of dendritic solidification,” Materials Science and Engineering: R: Reports,
vol. 41, no. 6, pp. 121–163, 2003.
[114] Q. Jiang and H. Lu, “Size dependent interface energy and its applications,”
Surface Science Reports, vol. 63, no. 10, pp. 427–464, 2008.
[115] E. Asadi, M. A. Zaeem, S. Nouranian, and M. I. Baskes, “Two-phase solid–
liquid coexistence of ni, cu, and al by molecular dynamics simulations using
the modified embedded-atom method,” Acta Materialia, vol. 86, pp. 169–
181, 2015.
[116] W. Boettinger, L. Bendersky, R. Schaefer, and F. Biancaniello, “On the for-
mation of dispersoids during rapid solidification of an ai-fe-ni alloy,” Metal-
lurgical Transactions A, vol. 19, no. 4, pp. 1101–1107, 1988.

133
BIBLIOGRAPHY

[117] A. Barbieri and J. Langer, “Predictions of dendritic growth rates in the

linearized solvability theory,” Physical Review A, vol. 39, no. 10, p. 5314,
1989.
[118] M. J. Aziz, “Model for solute redistribution during rapid solidification,”
Journal of Applied Physics, vol. 53, no. 2, pp. 1158–1168, 1982.
[119] M. J. Aziz and T. Kaplan, “Continuous growth model for interface motion
during alloy solidification,” Acta Metallurgica, vol. 36, no. 8, pp. 2335–2347,
1988.
[120] N. Ahmad, A. Wheeler, W. J. Boettinger, and G. B. McFadden, “Solute
trapping and solute drag in a phase-field model of rapid solidification,”
Physical Review E, vol. 58, no. 3, p. 3436, 1998.
[121] A. Clarke, D. Tourret, Y. Song, et al., “Microstructure selection in thin-
sample directional solidification of an al-cu alloy: In situ x-ray imaging and
phase-field simulations,” Acta Materialia, vol. 129, pp. 203–216, 2017.
[122] S. Sakane, T. Takaki, R. Rojas, et al., “Multi-gpus parallel computation of
dendrite growth in forced convection using the phase-field-lattice boltzmann
model,” Journal of Crystal Growth, vol. 474, pp. 154–159, 2017.
[123] T. Takaki, S. Sakane, M. Ohno, Y. Shibuta, and T. Aoki, “Large–scale
phase–field lattice boltzmann study on the effects of natural convection on
dendrite morphology formed during directional solidification of a binary
alloy,” Computational Materials Science, vol. 171, p. 109 209, 2020.
[124] N. Yamanaka, S. Sakane, and T. Takaki, “Multi-phase-field lattice boltz-
mann model for polycrystalline equiaxed solidification with motion,” Com-
putational Materials Science, vol. 197, p. 110 658, 2021.
[125] T. Shimokawabe, T. Aoki, T. Takaki, et al., “Peta-scale phase-field sim-
ulation for dendritic solidification on the tsubame 2.0 supercomputer,” in
Proceedings of 2011 International Conference for High Performance Com-
puting, Networking, Storage and Analysis, 2011, pp. 1–11.

134
 BIBLIOGRAPHY

[126] M. Greenwood, K. Shampur, N. Ofori-Opoku, et al., “Quantitative 3d phase

field modelling of solidification using next-generation adaptive mesh refine-
ment,” Computational Materials Science, vol. 142, pp. 153–171, 2018.
[127] S. Sakane, T. Takaki, and T. Aoki, “Parallel-gpu-accelerated adaptive mesh
refinement for three-dimensional phase-field simulation of dendritic growth
during solidification of binary alloy,” Materials Theory, vol. 6, no. 1, pp. 1–
19, 2022.
[128] J. Baiges, M. Chiumenti, C. A. Moreira, M. Cervera, and R. Codina, “An
adaptive finite element strategy for the numerical simulation of additive
manufacturing processes,” Additive Manufacturing, vol. 37, p. 101 650, 2021.
[129] Y. Wen, J. Simmons, C. Shen, C. Woodward, and Y. Wang, “Phase-field
modeling of bimodal particle size distributions during continuous cooling,”
Acta Materialia, vol. 51, no. 4, pp. 1123–1132, 2003.
[130] J. Zhu, T. Wang, A. Ardell, S. Zhou, Z. Liu, and L. Chen,“Three-dimensional
phase-field simulations of coarsening kinetics of γ particles in binary ni–al
alloys,” Acta Materialia, vol. 52, no. 9, pp. 2837–2845, 2004.
[131] G. Boussinot, A. Finel, and Y. Le Bouar, “Phase-field modeling of bimodal
microstructures in nickel-based superalloys,” Acta Materialia, vol. 57, no. 3,
pp. 921–931, 2009.
[132] M. Cottura, Y. Le Bouar, A. Finel, B. Appolaire, K. Ammar, and S. Forest,
“A phase field model incorporating strain gradient viscoplasticity: Applica-
tion to rafting in ni-base superalloys,” Journal of the Mechanics and Physics
of Solids, vol. 60, no. 7, pp. 1243–1256, 2012.
[133] M. A. Ali, J. V. Görler, and I. Steinbach, “Role of coherency loss on raft-
ing behavior of ni-based superalloys,” Computational Materials Science,
vol. 171, p. 109 279, 2020.
[134] L. Gránásy, G. I. Tóth, J. A. Warren, et al., “Phase-field modeling of crystal
nucleation in undercooled liquids–a review,” Progress in Materials Science,
vol. 106, p. 100 569, 2019.

135
BIBLIOGRAPHY

[135] C. Gandin, J.-L. Desbiolles, M. Rappaz, P. Thevoz, et al.,“A three-dimensional

cellular automation-finite element model for the prediction of solidifica-
tion grain structures,” Metallurgical and Materials Transactions A, vol. 30,
no. 12, pp. 3153–3165, 1999.
[136] T. Carozzani, H. Digonnet, and C.-A. Gandin, “3d cafe modeling of grain
structures: Application to primary dendritic and secondary eutectic solidi-
fication,” Modelling and Simulation in Materials Science and Engineering,
vol. 20, no. 1, p. 015 010, 2011.
[137] H. Takatani, C.-A. Gandin, and M. Rappaz, “Ebsd characterisation and
modelling of columnar dendritic grains growing in the presence of fluid
flow,” Acta Materialia, vol. 48, no. 3, pp. 675–688, 2000.
[138] T. Carozzani, “Développement d’un modèle 3d automate cellulaire-éléments
finis (cafe) parallèle pour la prédiction de structures de grains lors de la
solidification d’alliages métalliques,” Ph.D. dissertation, Ecole Nationale
Supérieure des Mines de Paris, 2012.
[139] R. Fleurisson, O. Senninger, G. Guillemot, and C.-A. Gandin, “Hybrid cel-
lular automaton-parabolic thick needle model for equiaxed dendritic solid-
ification,” Journal of Materials Science & Technology, vol. 124, pp. 26–40,
2022.
[140] K. Teferra and D. J. Rowenhorst, “Optimizing the cellular automata finite
element model for additive manufacturing to simulate large microstruc-
tures,” Acta Materialia, vol. 213, p. 116 930, 2021.
[141] D. Sgandurra, “An introduction to cellular automata,” 2009.
[142] W. 1. Kurz, B. Giovanola, and R. Trivedi, “Theory of microstructural de-
velopment during rapid solidification,” Acta Metallurgica, vol. 34, no. 5,
pp. 823–830, 1986.
[143] Y. Shibuta, M. Ohno, and T. Takaki, “Solidification in a supercomputer:
From crystal nuclei to dendrite assemblages,” Jom, vol. 67, no. 8, pp. 1793–
1804, 2015.

136
 BIBLIOGRAPHY

[144] S. D. Jadhav, S. Dadbakhsh, L. Goossens, J. Kruth, J. Van Humbeeck,

and K. Vanmeensel, “Influence of selective laser melting process parame-
ters on texture evolution in pure copper,” Journal of Materials Processing
Technology, vol. 270, pp. 47–58, 2019.
[145] M. Higashi and T. Ozaki, “Selective laser melting of pure molybdenum:
Evolution of defect and crystallographic texture with process parameters,”
Materials & Design, vol. 191, p. 108 588, 2020.
[146] W. W. Mullins and R. Sekerka, “Stability of a planar interface during solid-
ification of a dilute binary alloy,” Journal of Applied Physics, vol. 35, no. 2,
pp. 444–451, 1964.
[147] W. Kurz and R. Trivedi, “Rapid solidification processing and microstructure
formation,” Materials Science and Engineering: A, vol. 179, pp. 46–51, 1994.
[148] M. Carrard, M. Gremaud, M. Zimmermann, and W. Kurz, “About the
banded structure in rapidly solidified dendritic and eutectic alloys,” Acta
Metallurgica et Materialia, vol. 40, no. 5, pp. 983–996, 1992.
[149] W. Kurz and R. Trivedi, “Banded solidification microstructures,” Metallur-
gical and Materials Transactions A, vol. 27, no. 3, pp. 625–634, 1996.
[150] J. T. McKeown, K. Zweiacker, C. Liu, et al., “Time-resolved in situ mea-
surements during rapid alloy solidification: Experimental insight for additive
manufacturing,” Jom, vol. 68, no. 3, pp. 985–999, 2016.
[151] C. M. Pilgar, A. M. Fernandez, S. Lucarini, and J. Segurado, “Effect of
printing direction and thickness on the mechanical behavior of slm fabri-
cated hastelloy-x,” International Journal of Plasticity, vol. 153, p. 103 250,
2022.

137