A machine learning approach to fracture mechanics problems

Xing Liu
School of Engineering, Brown University
Providence, RI 02912, USA
[email protected]

Christos E. Athanasiou
School of Engineering, Brown University
Providence, RI 02912, USA
[email protected]

Nitin P. Padture
School of Engineering, Brown University
Providence, RI 02912, USA
[email protected]

Brian W. Sheldon*
Corresponding author
School of Engineering, Brown University
Providence, RI 02912, USA
[email protected]

Huajian Gao*
Corresponding author
School of Engineering, Brown University
Providence, RI 02912, USA
College of Engineering, College of Science
Nanyang Technological University, Singapore
+65 6790 9514
[email protected]
 A machine learning approach to fracture mechanics problems

Xing Liu1, Christos E. Athanasiou1, Nitin P. Padture1, Brian W. Sheldon1, Huajian Gao1,2

1School of Engineering, Brown University, Providence, RI 02912, USA

2College of Engineering, College of Science, Nanyang Technological University,

Singapore

Corresponding authors: Brian W. Sheldon ([email protected]), Huajian Gao

([email protected])

Abstract

Analytical and empirical solutions to engineering problems such as materials

characterization are usually preferred because of their convenience in applications.

However, they are not always accessible in complex problems. A new class of solutions,

based on machine learning (ML) models such as regression trees and neural networks

(NNs), are proposed and their feasibility and value are demonstrated through the analysis

of fracture toughness measurements as an example. It is found that both solutions based

on regression trees and NNs can provide accurate results for the specific problem, but

NN-based solutions outperform regression-tree-based solutions in terms of their

simplicity. This example demonstrates that ML solutions are a major improvement over

analytical and empirical solutions in terms of both reliable functionality and rapid

deployment. When analytical solutions are not available, the use of ML solutions can

overcome the limitations of empirical solutions and substantially change the way that

engineering problems are solved.

1
Keywords

Machine learning, Analytical methods, Mechanical properties testing, Fracture

1. Introduction

Engineers often seek analytical solutions for simplicity and reliability, which brings

bring great convenience to engineering applications such as materials characterization,

structural analysis and design. However, analytical solutions cannot always be obtained.

An acceptable compromise is the empirical solution which relies on engineers’

understanding and generalization of experimental and numerical data. For example, in

ASTM standard C1421 for determination of fracture toughness of advanced ceramics,

empirical solutions are provided to evaluate the plane-strain stress intensity factor at the

crack tip. Both analytical and empirical solutions are known for their rapid deployment

and reliability. However, some complicated engineering problems may involve a

nonlinear and complex relationship among higher-dimensional data, and empirical

solutions cannot be readily obtained it is beyond human’s learning ability to develop an

empirical solution. In the case where neither analytical nor empirical solutions are

feasible the following question can be raised: is there any other possible way to obtain a

solution? The computers are endowed with a strong learning ability by machine learning

(ML) algorithms, and, therefore, they could learn by themselves from experiments and

simulations [1,2,3], and come up with solutions based on ML models, referred to as

“machine learning solutions” in Fig. 1. Machine learning (ML) algorithms can be

informed directly by experiments and simulations [1,2,3], and thus provide “machine

learning solutions”. These ML solutions could serve as are a promising substitute for

2
analytical and empirical solutions if it they can provide rapid and accurate results. In this

context, the initial study presented here, summarized in Fig. 1, uses ML solutions for

materials characterization, taking fracture toughness measurements of microfabricated

brittle ceramic microcantilevers as an important example.

Figure 1. ML solutions to engineering problems. (a) Analytical solutions are developed based on a
physical understanding of the problem, which can be inaccessible in complicated cases. An acceptable
compromise that has been used traditionally relies on fitting abundant experimental and/or numerical data
to an empirical solution. This approach can have limited accuracy when dealing with a nonlinear and
complex relationship among high-dimensional data. Therefore, the ML solutions are proposed to overcome
this weakness and provide accurate results rapidly. ML solutions are provided by machines as they learn
from experimental and numerical data and could serve as a promising substitute when analytical and
empirical solutions are not feasible. (b) The specific engineering problem addressed in this work:
determination of fracture toughness by loading (using a nanoindenter) a pre-notched pentagonal cross-
section microcantilever at its end. The microcantilever is milled out of the bulk material, whose dimensions
are , , , , . (c) The ratio of the crack tip plane-strain stress intensity factor, , to the indentation
load, , are is evaluated from empirical solutions [4,5], FEM simulations and ML solutions for different
cantilever lengths, , and , , , = 5.20,4.80,0.90,1.00 . The ML solution can achieve
comparable accuracy to FEM results when the empirical solution fails.

3
2. Experimental

This investigation is based on a mechanical test that is used to measure the mode-I

fracture toughness, , in small specimens [4,5,6,7]. In this method, a pentagonal cross-

section microcantilever is cut with focused ion beam (FIB) micromachining in Helios

NanoLab 450 system (FEI, Oregon, USA), as shown in Fig. 1b. To induce a well-defined

controlled fracture event, a sharp pre-notch is milled at a distance, , from the fixed end

of the cantilever. The pre-notch is required to be sufficiently sharp to guarantee the

validity of the fracture toughness measurement. Therefore, a two-step strategy is adopted,

i.e., a coarse cut with a moderate current followed by a second finer cut with a lowest

possible current, 1 pA. This strategy restricts the notch radius well below 50 nm and

eliminates the artifacts caused by ion-implantation damage around the notch tip during

FIB milling. In this way, the notch tip is sharp enough to behave like an ideal crack and

ensure the validity of the measurement [8,9]. After the microcantilever fabrication, a

nanoindenter (Hysitron TI 900, Triboindenter, Minneapolis, USA) equipped with a

Berkovich tip is used to apply a controlled load at the free end of the cantilever and the

load-displacement response is recorded. Because the fracture of the specimen geometry is

inherently unstable, a catastrophic crack growth would be observed, i.e., a peak load

followed by a displacement burst. Therefore, the fracture toughness is correlated with the

critical load at the onset of fracture. An accurate evaluation of requires theoretical

solutions or numerical simulations.

In this study, three polycrystalline silicon specimens (~ 1 grain size, American

Elements, California, USA) were tested as reference samples. A scanning electron

microscope image (SEM) of a microcantilever is shown in Fig. 2a and the fracture

4
toughness measurements are summarized in Fig 2b. The specimen sizes are sufficiently

larger than the grain sizes, thus a homogeneous polycrystalline microstructure is sampled

during the measurement.

Figure 2. Fracture toughness tests on polycrystalline silicon specimens. (a) The scanning electron
microscope (SEM) image shows the exact geometry of the polysilicon specimens. (b) Fracture toughness
values of these three specimens were evaluated from empirical solutions, FEM simulations and ML
solutions for comparison. The dimensions of the polysilicon specimens are listed in Table A.1.

3. Theory/calculation

3.1 Problem description

An accurate analysis of the fracture toughness measurement is based on linear elastic

fracture mechanics and requires a solution that describes the plane-strain stress intensity

factor at the crack tip for a given load. The exact microcantilever geometry is limited by

practical considerations related to the use of the FIB instrument and by the need to obtain

a well-defined controlled fracture event during loading. With the geometries that are

typically used, it is impossible to derive analytical or empirical solutions that are highly

accurate over the full range of relevant sample dimensions. An early attempt [4,5], based

on an oversimplified functional form and limited numerical results, fails as it deviates

significantly from the results of finite element method (FEM), whose accuracy is

5
guaranteed as shown in Fig. 1c. It is certainly possible to obtain accurate solutions by

running an FEM simulation for each specimen. However, this is time-consuming and

impractical if there are a large number of specimens with variations in the specimen

dimensions, which is are essential to capture the statistical distribution of fracture

toughness. In this context, a ML solution can provide a valuable tool to accelerate the

implementation of a specific fracture toughness test. More generally, this provides a tool

with both portable deployment (comparable to analytical/empirical solutions) and high

accuracy (comparable to FEM simulations). For the example investigated here, the ML

solution should capture the relationship among the plane-strain stress intensity factor, ,

at the notch tip, the indentation load, , and the specimen dimensions, , , , , .

3.2 Data set preparation based on FEM simulations

Developing a ML solution requires a data set that includes desired inputs and outputs

for a large number of samples. These data are then used to train appropriate ML models,

which can evolve to optimize performance. The quality and quantity of the data dictate

the accuracy that can be achieved. Therefore, it is of great importance to build a good

data set, comprising a combination of both input and target data. The way that these data

are defined can greatly affect the ML process. The fracture toughness measurements used

here are assumed to strictly follow the assumptions of linear-elastic fracture mechanics,

such that the calculation of the plane-strain stress intensity factor, , is a boundary value

problem in linear elasticity. The boundary conditions of the problem imply that the crack

tip plane-strain stress intensity factor, , is directly proportional to the indentation load,

, and the ratio between them depends only on the specimen dimensions, , , , , ,

and is independent of elastic properties (Young’s modulus, , and Poisson’s ratio, ).

6
Dimensional analysis indicates that five independent dimensionless variables,

, , , ;&
! " $%
!.' (" () , are relevant. Based on this, the following structure of the ML
"

solutions is employed:

, Input variables: - = .- , -/ , -0 , -1 2 = 3 , , , 4
. .12
+Target variable: 5 =
* .6 7 7/

Based on the specimen sizes used for typical measurements, the parameter space of

the input variables is defined as:

∈ 90.1, 0.8:, ∈ 91.0, 3.0:, ∈ 90.1, 0.4:, ∈ 92.0, 5.0:. .22

In order to fully explore the parameter space of input variables, a grid-search strategy

is adopted when the data set for ML is generated. Each input domain -< in

.- , -/ , -0 , -1 2 = = , , !
, ">, is discretized into < uniform intervals and a grid of all

possible input variables is then constructed inside the parameter space. This gives ? =

. @ 12. / @ 12. 0 @ 12. 1 @ 12 different samples. A dense grid of

. , /, 0, 12 = .35, 100, 10, 102 is then used to generate a data set for ML with

? = 439,956 samples. Since the target variable in this problem is continuous and

bounded over the whole input parametric space, the accuracy of the ML solutions over

the continuous space can be estimated in a reliable manner by sampling these discrete

points.

FEM is used to evaluate the target variable 5 for each sample. The direct domain J-

integral [10,11] and elastic compliance [12] are the two conventional methods to evaluate

the energy release rate, B , and the plane-strain stress intensity factor, =

7
C B ⁄.1 − / 2, at the crack tip. This first evaluates B from the direct domain J-integral.

The elastic compliance method calculates the rate of change of the compliance with crack

extension for a given indentation load, which then gives and hence the rate of potential

energy loss, which is equivalent to B. Both methods provide an accurate value of the

target, 5 = , for a given input, - = = , , , " > . However, the

$% !
& " !.' (" ()

computational cost of the J-integral method is significantly lower than the elastic

compliance method since the latter requires multiple simulations for different crack

lengths. It is crucial to improve the computational efficiency of the data generation,

especially when the size of the data set is huge, and thus the J-integral method is clearly

the best choice. This was implemented through user-defined subroutines in FEAP [13].

An encastre boundary condition were applied to the fixed end of the cantilever and a

displacement-controlled boundary condition to the free end. More than 10,000 full

integration elements (C3D20) were employed with the mesh being refined near the crack

tip. The average wall time for each simulation was 120 seconds when running in parallel

≈ 31 days,
10F,F6G⋅ / ⋅ G
0/ ⋅G ⋅G ⋅/1
on 16 CPUs. The generation of the whole data set took roughly

with 20 jobs running simultaneously on 320 CPUs. The elapsed time can be further

shortened if more CPUs are utilized.

There is a wide variety of ML models and algorithms, and one needs to find the most

appropriate one for their specific problems. After screening the popular ML models, it is

found that regression trees (RTs) and their ensembles, neural networks (NNs) are most

suitable for this problem.

8
4. Results

4.1 Regression-tree-based solutions

The tree-based method for regression can capture the highly non-linear relationship

between the input variables, - = .- , -/ , -0 , -1 2 , and the target variable, 5 . The

regression tree is built through dividing the input parametric space into distinct, non-

overlapping regions known as leaf nodes. Each leaf node is labeled with the average

value of targets of the training samples that fall into the region, as shown in Fig. 3. A

given input would flow from the root node, through the internal branch nodes, into the

terminal leaf node, and produce a prediction of target variable by the label of the leaf

node. An optimized CART algorithm [14] based on information gain is implemented in

the open-source package Scikit-learn [15] to grow the regression trees. The complexity

and size of a grown regression tree is controlled by the maximum tree depth and the

minimum leaf size. The tree depth represents the maximum number of edges from the

root node to the leaf node and is limited to a moderate interval 4~8. The leaf size

represents the number of samples required for each leaf node and a reasonable lower

bound is set to 50. Using these parameters overfitting can be effectively eliminated.

The structure and performance of the regression tree solutions are summarized in Fig.

3 and Table A.2. The accuracy of the solutions is measured by the maximum value of the

absolute percentage error:

J = KL5MNOP − 5QNPR S⁄5MNOP K × 100%, .32

over all the samples in the data set, where 5QNPR is the target value predicted by the ML

solution and 5MNOP is the target value of the data set which is initially obtained from FEM

9
simulations. As a limit is set on the minimum leaf size, the accuracy of these regression

trees gets saturated before reaching an acceptable low value, as shown in Fig. 3 and Table

10
A.2.

Although a single regression tree is not powerful enough to address the complex

Figure 3. Regression-tree-based solutions. (a) A regression tree is a hierarchically organized structure

with each branch node splitting the input parametric space, -. Eventually the parametric space is divided
into distinct and non-overlapping regions, i.e., the leaf nodes at the end of each branch. The value of target
variable, 5, is then predicted by the region that the input falls in. A regression tree is built through dividing

the input parametric space, i.e., - = .- , -/ , -0 , -1 2 = = , , !

, " > , into distinct and non-overlapping

regions; each region is represented by a leaf node and corresponds to a value of the prediction of target
variable. (b) A regression-tree ensemble combines a set of weak regression trees and make predictions by
aggregating responses from each single regression tree. (c) Ensemble methods, such as GBRT, can achieve
much higher accuracy (i.e., smaller absolute percentage error) than a single regression tree.

11
problem, it can be used as a building block to construct a better model with stronger

learning ability, i.e., a regression tree ensemble, as shown in Fig. 3. Given an input of

- = .- , -/ , -0 , -1 2, each regression tree in the ensemble provides a response which is

then synthesized into a final prediction of the target variable 5. Gradient Tree Boosting

algorithm [16,17] is implemented to build the regression tree ensembles which allows for

the optimization of arbitrary differentiable loss functions.

A series of gradient-boosted regression trees (GBRT) with least squares loss are

constructed with different numbers and sizes of the basic regression trees, and their

structures and performance are summarized in Fig. 3 and Table A.2. Interestingly, the

accuracy of prediction is greatly improved. An accuracy of less than 5% of relative error

can be achieved by combining 512 regression trees of depth 6.

Thus, the learning capability of the regression tree ensemble method in producing an

accurate ML solution is demonstrated. It is worth mentioning that a regression tree

ensemble typically contains a large number of basic trees which drastically increases the

complexity of the corresponding ML solution.

4.2 Neural-network-based solutions

NNs are computing systems with interconnected nodes inspired by biological neural

networks in the human brain. NNs are known for their power in approximating the

nonlinear relationships among high-dimensional data (inputs and outputs) according to

the universal approximation theorem [18]. A simple NN consists of an input layer, an

output layer, and hidden layers in-between. These layers are interconnected to form a

network architecture, as shown in Fig. 4. Feedforward NN is adopted, in which each node

12
in one layer is connected to all the nodes in the adjacent layers and information is fed

between layers only in the forward direction. As the input data - = .- , -/ , -0 , -1 2 is fed

into the NN through the input layer, each node in the next hidden layer processes data

from the input layer and feeds the next layer through an activation function. Finally, the

output layer collects data from the last hidden layer and produce the target data, 5. The

architecture of a NN is controlled by the number of hidden layers and the number of

nodes in each hidden layer. Simple NNs with 1 or 2 hidden layers with the rectified linear

activation function (ReLU) “ReLU” activation function, i.e., single-layer and multilayer

perceptrons, are employed in this study.

Due to the difficulty in their learning process, NNs must be complemented by a

robust ML algorithm, such that, they can learn and evolve continuously as new data is fed

to them. Nadam algorithm [19] with Log-Cosh loss function is adopted for the training of

NNs using the open-source platform TensorFlow r2.0 [20]. To fully exploit the learning

capability of NNs, a large portion of the data set is utilized for training while keeping a

small portion for validation and test. Therefore, the data set is split into training data set

(70%), validation data set (15%), and test data set (15%). By collecting and reviewing the

accuracy of these three data sets during the training process as shown in Fig. A.1, it is

demonstrated that all the NNs are converged and not overfitted.

The structure and performance of the NN-based solutions are summarized in Fig. 4

and Table A.3. It turns out that solutions with high accuracy can be achieved even with

simple NNs: a NN with only 29 neurons (4/16/8/1) can produce an accuracy of less than

5% of relative error. As presented in Fig. 4, one can choose appropriate NN-based

solutions to tune the accuracy according to the application needs. Surprisingly, by using a

13
 large data set, i.e., ? = 439,956, even with a small number of neurons, i.e., 29, the

strong learning ability of a NN is demonstrated (Fig. 4).

Figure 4. NN-based solutions. A NN is composed of an input layer, hidden layers, and an output layer. In

this context, the input layer consists of 4 nodes corresponding to - = = , , !

, ">; the output layer has a

single node for 5 = &

$%
!.' (" () . NNs with various hidden layer structures are examined: (a) single-layer
"

perceptrons denotated “ V< ⁄V ⁄VW ”, and (b) multilayer perceptrons denotated “ V< ⁄V ⁄V/ ⁄VW ”, where
V< , V , V/ , VW are the number of nodes in the input layer, hidden layers, and output layer, respectively. (c)
ML solutions with high accuracy (i.e., small absolute percentage error) can be obtained even with simple
NNs. One can choose appropriate NN-based solutions to tune the accuracy according to the application
needs.

14
4.3 Portable deployment of ML solutions

It is concluded that both the regression-tree-based solutions and NN-based solutions

can provide accurate results for the specific problem. However, NN-based solutions

outperform regression-tree-based solutions in terms of their simplicity (Fig. 3 and Fig. 4).

Indeed, a regression-tree-based solution with less than 5% of relative error involves

hundreds of regression trees and each regression tree consists of hundreds of nodes.

Therefore, it is challenging to obtain an overall view of the redundant structure of the

solution. In contrast, a NN-based solution can achieve the same level of accuracy only

with a simple interconnected network. As simplicity of a ML solution is essential for

rapid and robust deployment, the NN-based solutions are preferred.

The optimal ML solution can be exported in an open-standard file format such as

Javascript object notation (JSON), and shared among researchers and engineers in the

community. The standard data-interchange format sets the scene for researchers to

operate in various development platforms. The optimal ML solution can also be

established as an open-access standard for specific engineering problems. A graphical

user interface can be further developed for accessing and visualizing the problem and the

established standard ML solution. In this context, the ML solution for the fracture

toughness measurements has been created and shared through a web-based application,

named “SIF Calculator” [21]. The ML solution can achieve comparable accuracy to FEM

results as shown in Fig. 1c and Fig. 2b. The overview of the deployment and positioning

of the ML solutions is illustrated in Fig. 5. The ML solution can achieve the

functionalities of an empirical solution.

15
Figure 5. Functionality and positioning of a ML solution to engineering problems. The optimal ML
solution can be exported in an open-standard file format, such as JSON, and shared among researchers for
further operations and development. It can also be established as an engineering standard and stored online
for public access. Cloud or web-based applications can be developed to interactively provide fast and
accurate solutions to engineers.

5. Discussion

In this work, an innovative class of solutions based on ML models are proposed to

solve engineering problems, especially materials characterization, which could serve as a

promising substitute when analytical and empirical solutions are not accessible. The

feasibility and advantage of the ML solutions are demonstrated through the application in

small-scale fracture toughness measurements of pre-notched pentagonal cross-section

cantilevers. This ML approach can be easily adopted for other experimental

characterization methods with complex specimen geometries, such as cantilever-based

16
methods [22,23,24,25] and indentation-based methods [26,27,28]. The development

process consists of 4 basic steps,

1. Definition of a well-posed problem. Solving a complex engineering problem requires

the development of a mathematical description and the understanding of what physical

quantities are involved. Dimensional analysis can assist in finding all the relevant

dimensionless numbers from these quantities, and the essence of the solution is to

understand the relationships among these quantities.

2. Preparation of a data set. The dimensionless quantities involved can be assorted into

input variables and target variables based on their physical meanings. It is necessary to

explicitly define the parameter space of the input variables, which covers the whole

region of interest (ROI). Since the reliability of the ML solutions is of high priority, these

solutions are restrained from doing extrapolation and the performance within ROI is

emphasized. The quality of the data set directly controls the accuracy of the solutions.

Reliable experimental and numerical data are essential.

3. Model selection and training. There are a large variety of ML models that can be

employed. Based on the data set and type of problem, the appropriate ML models should

be selected. The optimal solution is achieved by evaluating their accuracy and simplicity.

4. Deployment of ML solutions. The optimal ML solution can be shared among the

community in an open-standard format. Engineers can obtain rapid and accurate results

by interacting with cloud or web-based applications which can be integrated with the ML

solution.

17
6. Conclusions

ML provides an alternative way to solve complex engineering problems, due to its

advantages in dealing with the nonlinear and complex relationship among high-

dimensional physical quantities. More specifically in this work, it is demonstrated that

ML greatly accelerates data interpretation in fracture toughness measurements. The

simplicity of employing ML solutions also paves the way creates a foundation for the

standardization of standardizing these types of novel experimental methods. Besides

materials characterization, ML solutions have a wide range of possible applications in

materials science. Possible examples include e.g., predictions of surface roughness and

temperature rise (target variables) in metal cutting under different machining conditions

(input variables), and predictions of the mechanical properties of 3D printing products

(target variables) under different manufacturing conditions (input variables). In summary

Therefore, ML solutions can substantially change the way that engineering problems are

solved.

Acknowledgements

We acknowledge the financial support from the US Department of Energy Basic Energy

Sciences Grant # DE-SC0018113 for this research. This work was conducted using the

computational resources and services at the Brown University Center for Computation

and Visualization.

18
Appendix A

Table A.1. Fracture toughness measurements of polysilicon specimens.

X Y Z [\ [] ^_` bcd .? ⋅ .6
2
. 2 . 2 . 2 . 2 . 2 . a2 Empirical1 FEM ML

PolySi 1 5.00 0.90 5.00 1.00 10.00 2.30 1.73 0.97 0.97

PolySi 2 4.80 0.90 5.20 1.00 10.00 2.35 1.77 0.99 0.99

PolySi 3 5.20 1.00 5.50 1.00 11.00 2.50 1.78 0.98 0.98
1
The empirical solution was developed in [4, 5].

19
Table A.2. Construction and performance of regression-tree-based solutions.

Avg No. of Min No. of samples

No. of trees Tree depth Max. APE1
leaf nodes at a leaf node
RT 1 4 16 8580 29.706%

RT 1 5 32 3240 23.828%

RT 1 6 64 1296 19.982%

RT 1 7 128 240 17.440%

RT 1 8 256 88 17.146%

GBRT 256 4 16 430 20.348%

GBRT 256 5 32 106 17.019%

GBRT 256 6 64 51 14.327%

GBRT 256 7 128 50 11.483%

GBRT 256 8 255 50 9.724%

GBRT 512 4 16 93 7.873%

GBRT 512 5 32 50 6.018%

GBRT 512 6 64 50 4.122%

GBRT 512 7 127 50 3.334%

GBRT 512 8 251 50 2.171%

1
Max. APE: maximum absolute percentage error.

20
Table A.3. Construction and performance of NN-based solutions.

NN No. of trainable Training Validation Test Fine-tuned

architecture parameters Max. APE1 Max. APE Max. APE Max. APE1

4/16/1 97 9.003% 8.438% 8.858% 8.866%

4/32/1 193 6.046% 5.169% 5.357% 5.693%

4/64/1 385 2.907% 2.638% 1.556% 2.764%

4/128/1 769 1.755% 1.762% 1.675% 1.590%

4/256/1 1537 1.691% 1.266% 1.402% 1.476%

4/512/1 3073 1.359% 0.905% 1.101% 1.150%

4/8/8/1 121 7.589% 7.208% 7.177% 7.192%

4/8/16/1 201 5.993% 5.072% 5.299% 5.742%

4/16/8/1 225 4.076% 3.359% 3.561% 3.789%

4/16/16/1 369 4.109% 3.457% 3.989% 3.785%

4/32/32/1 1249 1.618% 1.394% 1.512% 1.558%

4/64/64/1 4545 1.130% 1.043% 1.119% 1.111%

1
Max. APE: maximum absolute percentage error.

21
Figure A.1. Training history of NNs. The maximum absolute percentage error (max. APE) on the training
and validation data sets are recorded over the number of epochs for some simple NNs: (a) 4/16/1, (b)
4/32/1, (c) 4/64/1, (d) 4/128/1, (e) 4/256/1, (f) 4/512/1, (g) 4/8/8/1, (h) 4/8/16/1, (i) 4/16/8/1, (j) 4/16/16/1,
(k) 4/32/32/1 and (l) 4/64/64/1. Each plot shows that the training and validation accuracy reach the same
plateau, which indicates that the NN is converged and not overfitted.

22
References

[1] M. Mozaffar, R. Bostanabad, W. Chen, K. Ehmann, J. Cao, M.A. Bessa, Deep

learning predicts path-dependent plasticity, Proceedings of the National Academy of
Sciences 116 (2019) 26414-26420. https://doi.org/10.1073/pnas.1911815116.
[2] L. Liang, M. Liu, C. Martin, W. Sun, A deep learning approach to estimate stress
distribution: a fast and accurate surrogate of finite-element analysis, J. R. Soc. Interface
15 (2018) 20170844. https://doi.org/10.1098/rsif.2017.0844.
[3] G.X. Gu, C.T. Chen, M.J. Buehler, De novo composite design based on machine
learning algorithm, Extreme Mechanics Letters 18 (2018) 19-28.
https://doi.org/10.1016/j.eml.2017.10.001.
[4] D. Di Maio, S.G. Roberts, Measuring fracture toughness of coatings using focused-
ion-beam-machined microbeams, Journal of Materials Research 20 (2005) 299-302.
https://doi.org/10.1557/JMR.2005.0048.
[5] D. Di Maio, S.G. Roberts, Measuring fracture toughness of coatings using focused-
ion-beam-machined microbeams–ERRATUM, Journal of Materials Research 32 (2017)
1211. https://doi.org/10.1557/jmr.2017.84.
[6] C.E. Athanasiou et al., High toughness carbon-nanotube-reinforced ceramics via ion-
beam engineering of interfaces, Carbon (In Press).
https://doi.org/10.1016/j.carbon.2020.02.075.
[7] B.N. Jaya, V. Jayaram, Fracture testing at small-length scales: From plasticity in Si to
brittleness in Pt, JOM 68 (2016) 94-108. https://doi.org/10.1007/s11837-015-1489-2.
[8] S. Wurster, C. Motz, R. Pippan, Characterization of the fracture toughness of micro-
sized tungsten single crystal notched specimens. Philosophical Magazine 92 (2012)
1803–1825. https://doi.org/10.1080/14786435.2012.658449
[9] F. Iqbal, J. Ast, M. Göken, K. Durst, In situ micro-cantilever tests to study fracture
properties of NiAl single crystals, Acta Materialia 60 (2012) 1193-1200.
https://doi.org/10.1016/j.actamat.2011.10.060.
[10] C.F. Shih, B. Moran, T. Nakamura, Energy release rate along a three-dimensional
crack front in a thermally stressed body, International Journal of fracture 30 (1986) 79-
102. https://doi.org/10.1007/BF00034019.

23
[11] T.D. Nguyen, S. Govindjee, P.A. Klein, H. Gao, A material force method for
inelastic fracture mechanics, Journal of the Mechanics and Physics of Solids 53 (2005)
91-121. https://doi.org/10.1016/j.jmps.2004.06.010.
[12] T.L. Anderson, Fracture Mechanics: Fundamentals and Applications, CRC Press,
Boca Raton, 2005.
[13] R. L. Taylor, FEAP - A Finite Element Analysis Program, University of California,
Berkeley, 2014.
[14] L. Breiman, J.H. Friedman, R.A. Olshen, C.J. Stones, Classification and Regression
Trees, CRC, Boca Raton, 1984.
[15] F. Pedregosa et al., Scikit-learn: Machine learning in Python, Journal of Machine
Learning Research 12 (2011) 2825-2830.
[16] J.H. Friedman, Greedy function approximation: a gradient boosting machine, Annals
of Statistics 29 (2001) 1189-1232.
[17] J.H. Friedman, Stochastic gradient boosting, Computational Statistics & Data
Analysis 38 (2002) 367-378. https://doi.org/10.1016/S0167-9473(01)00065-2.
[18] K. Hornik, Approximation capabilities of multilayer feedforward networks, Neural
Networks 4 (1991) 251-257. https://doi.org/10.1016/0893-6080(91)90009-T.
[19] T. Dozat, Incorporating Nesterov Momentum into Adam, Proceedings of the
International Conference on Learning Representation (2018).
[20] M. Abadi et al., TensorFlow: large-scale machine learning on heterogeneous
distributed systems, Preprint at http://arxiv.org/abs/1603.04467 (2016).
[21] X. Liu, SIF Calculator. https://hint1412.github.io/XLiu.github.io/SIF/, 2019
(accessed 30 December 2019).
[22] D.E. Armstrong, A.J. Wilkinson, S.G. Roberts, Measuring anisotropy in Young’s
modulus of copper using microcantilever testing, Journal of Materials Research 24 (2009)
3268-3276. https://doi.org/10.1557/jmr.2009.0396.
[23] M.G. Mueller et al., Fracture toughness testing of nanocrystalline alumina and fused
quartz using chevron-notched microbeams, Acta Materialia 86 (2015) 385-395.
https://doi.org/10.1016/j.actamat.2014.12.016.

24
[24] G. Žagar, V. Pejchal, M.G. Mueller, L. Michelet, A. Mortensen, Fracture toughness
measurement in fused quartz using triangular chevron-notched micro-cantilevers, Scripta
Materialia 112 (2016) 132-135. https://doi.org/10.1016/j.scriptamat.2015.09.032.
[25] A.D. Norton, S. Falco, N. Young, J. Severs, R.I. Todd, Microcantilever investigation
of fracture toughness and subcritical crack growth on the scale of the microstructure in
Al2O3, Journal of the European Ceramic Society 35 (2015) 4521-4533.
https://doi.org/10.1016/j.jeurceramsoc.2015.08.023.
[26] X. Wang et al., High damage tolerance of electrochemically lithiated silicon, Nature
Communications 6 (2015) 8417. https://doi.org/10.1038/ncomms9417.
[27] M. Sebastiani, K.E. Johanns, E.G. Herbert, F. Carassiti, G.M. Pharr, A novel pillar
indentation splitting test for measuring fracture toughness of thin ceramic coatings,
Philosophical Magazine 95 (2015) 1928-1944.
https://doi.org/10.1080/14786435.2014.913110.
[28] M. Sebastiani, K.E. Johanns, E.G. Herbert, G.M. Pharr, Measurement of fracture
toughness by nanoindentation methods: Recent advances and future challenges, Current
Opinion in Solid State and Materials Science 19 (2015) 324-333.
https://doi.org/10.1016/j.cossms.2015.04.003.

25