International Journal of Pavement Engineering

ISSN: 1029-8436 (Print) 1477-268X (Online) Journal homepage: https://www.tandfonline.com/loi/gpav20

Asphalt concrete master curve using dynamic

backcalculation

Gabriel Bazi & Tatiana Bou Assi

To cite this article: Gabriel Bazi & Tatiana Bou Assi (2020): Asphalt concrete master
curve using dynamic backcalculation, International Journal of Pavement Engineering, DOI:
10.1080/10298436.2020.1733567

To link to this article: https://doi.org/10.1080/10298436.2020.1733567

Published online: 02 Mar 2020.

Submit your article to this journal

View related articles

View Crossmark data

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=gpav20
INTERNATIONAL JOURNAL OF PAVEMENT ENGINEERING
https://doi.org/10.1080/10298436.2020.1733567

Asphalt concrete master curve using dynamic backcalculation

Gabriel Bazi and Tatiana Bou Assi
Lebanese American University (LAU), New York, NY, USA

ABSTRACT ARTICLE HISTORY

The dynamic backcalculation of pavement layer properties, and more specifically the ability to determine Received 28 October 2019
the asphalt concrete master curve, using Falling Weight Deflectometer (FWD) data has caught the Accepted 18 February 2020
attention of many researchers for years. The damaged or aged master curve of existing pavement
KEYWORDS
structures allows the accurate prediction of the pavement structural capacity and remaining life – a Falling weight deflectometer;
goal that cannot be achieved through laboratory testing. To address this need, the authors developed finite element modelling;
a dynamic backcalculation application that employs the ABAQUS software for forward calculation using dynamic backcalculation;
implicit dynamic analyses, and the Newton-Raphson’s root-solving algorithm for optimisation. The asphalt concrete master
capabilities of the application were demonstrated for three simulated flexible pavement structures and curve; Newton-Raphson;
two asphalt concrete mixes using 40 combinations. The asphalt concrete moduli at the FWD most Rayleigh damping
dominant frequencies around 17 Hz, and the variables of unbound layers converged to the exact
values in few iterations for all combinations. The master curves, and mainly the maximum moduli, were
reliably determined to less than 1% and 4% error for about 60% and 85% of the combinations,
respectively. The potential for determining the master curve and maximum modulus was shown to be
dependent on the asphalt concrete mix, asphalt concrete temperature and pavement structure.

Introduction The use of dynamic backcalculation for determining the

pavement layer variables, and more specifically the develop-
The Falling Weight Deflectometer (FWD) is a non-destructive
ment of the asphalt concrete master curve, is critical for design-
device that simulates traffic/aircraft loading commonly used to
ing and evaluating long-lasting pavement structures using
assess the pavement structural capacity. The FWD simulates
mechanistic-empirical principles and for adopting perform-
the moving wheel by applying an impulse load having a 30–
ance-based specifications during construction that describe
40 msec. duration and measures the pavement surface deflec-
the desired levels of fundamental engineering properties to pre-
tions at various offsets from the centre of the load. The FWD
dict performance (A2F03 Committee on Management of Qual-
load and deflections are used, along with the pavement layer
ity Assurance 2002). The development of the dynamic
thicknesses, to determine the layer moduli in a process called
backcalculation application called PULSE addressed this need.
backcalculation. The backcalculation process uses a forward
The application employs the ABAQUS (2019) software for for-
calculation model for simulating the pavement structure and
ward calculation using implicit dynamic analyses, and the New-
calculating the surface deflections, in addition to a mathemat-
ton-Raphson’s root-solving algorithm for optimisation.
ical optimisation tool to iteratively improve the layer moduli.
Though PULSE can be used for both flexible and rigid pave-
Static or dynamic models are used for forward calculation, with
ments, it is demonstrated in this paper for flexible pavements.
the latter being able to model the inertial effects, the viscoelastic
The paper is organised as follows: a literature review
behaviour of the asphalt concrete layer and the damping of the
describes existing dynamic backcalculation software. The
materials. The dynamic modelling is vital in reliably predicting
newly developed PULSE application is then illustrated, including
the pavement responses (surface deflections, in this case), especially
the finite element forward calculation model, the linear visco-
under the FWD that applies a load of very short duration.
elastic modelling of the asphalt concrete material, the New-
It is important to note that the all static and dynamic
ton-Raphson’s root-solving algorithm, and the approximation
models, including finite element and finite layer, are based on
of the Jacobian matrix using finite-difference. Then, results of
solid mechanics principles knowing that most pavement
the dynamic backcalculation for three simulated flexible pave-
materials – except for the Portland cement concrete – are par-
ment structures and two asphalt concrete mixes are presented.
ticulate and not solid. The use of solid mechanics to quantify
The paper concludes with a summary and recommendations
the pavement responses is therefore an approximation to reality
for future work.
(Ullidtz 1998). Despite this fact, many studies have shown that
the use of solid mechanics is adequate for the calculation of
pavement responses based on comparisons with field-measured
responses (Sebaaly 2002, Al-Qadi et al. 2004, 2010, Gopalak- Literature review
rishnan and Thompson 2006a, 2006b, Garg et al. 2010, Wang Dynamic analysis of pavement structures is critical for model-
et al. 2013, Bazi et al. 2019a). ling the behaviour of pavement structures under dynamic

CONTACT Gabriel Bazi [email protected]

© 2020 Informa UK Limited, trading as Taylor & Francis Group
2 G. BAZI AND T. B. ASSI

loading. Sebaaly et al. (1985, 1986) were the pioneers in indicat- the current prediction accuracies are not sufficient to recommend
ing that the dynamic analysis and inertial effects are important these models for practical implementation.
in the prediction of pavement responses. The potential for developing the asphalt concrete master
Dynamic analyses of pavement structures either use the curve was investigated in this study using an application that
finite element or finite layer models for the calculation of pave- uses a finite element model in the time domain and the New-
ment responses: deflections, stresses or strains. Several software ton-Raphson’s method. The limitations are explained based on
tools were developed for the backcalculation of pavement layer the evaluation of simulated flexible pavement structures using
properties using dynamic analyses. Those programmes require synthetic deflection time histories. This method does not require
the time histories of the FWD load and deflection sensors, and any constraints on the variables, as opposed to the previously
they are based on either frequency or time domain methods. In reported methods. The results are promising, but further inves-
the dynamic procedure, the deflection time histories at each tigation and research are needed to apply this method to field
FWD sensor are used. In addition to the peak deflections measured FWD data. Nonetheless, the current application is
used in static analyses, the time delay or lag between the peak readily available for use with field data to reliably determine
of the deflection and the peak of the load is also utilised. the pavement layer variables (moduli and Rayleigh damping
For backcalculation in the time domain, the deflection time coefficients) and asphalt concrete moduli at the FWD dominant
histories calculated from forward analysis are matched with the frequencies, as demonstrated in a recent study using seasonal
measured deflection time histories at one or more sensors. For FWD testing for an LTPP section (Bazi et al. 2020).
backcalculation in the frequency domain, the measured time
histories of applied load and deflections are transformed into
Dynamic backcalculation application
the frequency domain through Fourier transform. Pavement
properties are then determined by matching the calculated fre- A dynamic backcalculation application called PULSE was devel-
quency response functions with those obtained from measured oped using Visual Basic for Applications (VBA). The appli-
data at desired frequencies at one or more sensors (Luo 2008). cation can be used for both flexible and rigid pavements and
Dynamic backcalculation software tools were summarised has the capability of determining the asphalt concrete master
in a recent FHWA report by Chatti et al. (2017), and those soft- curve for flexible pavements.
ware tools either use the wave propagation theory or finite The asphalt concrete layer for flexible pavements is mod-
element method. Magnuson et al. (1991), Kang (1998), Al- elled as a linear viscoelastic (LVE) material, and all other layers
Khoury et al. (2001) and Losa (2002) developed dynamic back- including the Portland cement concrete layer for rigid pave-
calculation programmes based on the frequency domain ments are modelled as linear elastic with damping.
methods. Ling et al. (1991), Matsui et al. (2000) and Turkiyyah The ABAQUS finite element (FE) software is used for for-
(2005) developed programmes based on the time domain ward calculation to quantify the surface deflections vs. time
methods. Finally, Uzan (1994a) and Chatti et al. (2004a, through an implicit dynamic analysis. The multivariate New-
2004b) developed programmes having the flexibility of either ton-Raphson (NR) method is used to improve the variables
using the time or frequency domain methods. Uzan (1994b) (LVE coefficients, layer moduli and damping coefficients) itera-
compared the two approaches and indicated that the time tively to match five parameters for every FWD deflection time
domain scheme is preferred over the frequency domain when history. These five parameters are essential to capture the shape
the deflection histories do not decay to zero at the end of the and magnitude of each deflection pulse.
sampling window (typically 60 msec.). Those programmes The Microsoft Excel Solver was initially used (i) to deter-
use a variety of optimisation techniques including, but not lim- mine the Prony series coefficients for modelling the asphalt
ited to, the linear and non-linear least-square optimisation and concrete as LVE, and (ii) to solve the system of equations in
the Newton’s method. the NR method for improving the variables for every iteration.
Few attempts were made in the past to develop the asphalt The Excel Solver uses the generalised reduced gradient method
concrete master curve from FWD data. Michigan State Univer- (GRG) (Lasdon et al. 1976), an algorithm that falls under the
sity (MSU) led this effort (Kutay et al. 2011, Zaabar et al. 2014) constrained non-linear optimisation (programming) category.
and developed a procedure based on a time domain dynamic This algorithm is very powerful but requires unnecessary com-
solution with viscoelastic asphalt concrete layer as a forward putational power for solving simple problems such as the Prony
routine. Kutay et al. (2011) indicated that the master curve series coefficients or NR method.
may be obtained for frequencies above 10−3 Hz and reported The MATLAB package (2018) replaced the Excel Solver in
modest success in the backcalculation process. Recommen- the latest version of the PULSE application. This upgrade
dations for improvements in FWD technology and test pro- reduced the computational time from several minutes to a
cedures were made. Zaabar et al. (2014) used a genetic couple of milliseconds, which can be attributed to the high-per-
algorithm as the search core for backcalculation analysis. The formance programming language and specific functions used in
method appears to be robust, but it requires constraints (limits) MATLAB.
on the sigmoidal function coefficients and those constraints are A flowchart of the process followed by the PULSE application
not readily available unless laboratory testing is performed. Gopa- is shown in Figure 1.
lakrishnan et al. (2014) also made a similar effort using the same The application does not place any constraints, either upper
layered viscoelastic forward analysis tool developed by MSU or lower, on the variables and it efficiently converges to the
along with a neural network methodology, but indicated that exact values.
 INTERNATIONAL JOURNAL OF PAVEMENT ENGINEERING 3

where: E = time- and temperature-dependent relaxation mod-

ulus; δ, α, β, and γ = fit constants; 10δ = minimum modulus;
10δ+α = maximum modulus; γ = steepness of the function; fr =
reduced test frequency: fr = f×a(T) or log( fr) = log( f ) + log[a
(T )] with f = frequency, T = Temperature and log[a(T )] =
shift factor.
The sigmoidal function has no minima or maxima at finite
times or frequencies. The upper part of the function approaches
asymptotically the maximum modulus of the mix at cold temp-
eratures, which is dependent on the limiting binder modulus
(glassy modulus). The lower part approaches asymptotically a
Figure 1. PULSE application flowchart.
limiting equilibrium value at high temperatures, which is
dependent on the aggregate gradation (Kim 2008). The func-
tion has one inflection point at which the slope is the largest.
In the coming sections, the dynamic FE model is introduced The LVE behaviour for the asphalt concrete layer is mod-
along with the Rayleigh damping coefficients. The modelling of elled in the ABAQUS software using the Prony series with 14
the asphalt concrete as linear viscoelastic material and the deter- coefficients – a limit set by ABAQUS. One coefficient rep-
mination of the Prony series coefficients from the sigmoidal func- resents the long-term equilibrium modulus and the remaining
tion coefficients is illustrated. The Newton-Raphson method is 13 coefficients represent the elastic stiffnesses at different vis-
later explained in depth, and the method is applied for dynamic cous relaxation times having one-decade intervals.
backcalculation. The Jacobian matrix needed for the Newton’s The dynamic backcalculation of the asphalt concrete LVE
method is determined using the finite-difference derivative behaviour is achieved by only evaluating the four master curve
approximation. Finally, the parameters needed to capture the sigmoidal coefficients as opposed to the 14 Prony series coeffi-
shape and magnitude of each FWD sensor’s pulse are defined. cients, a process that minimises the number of unknows from
14 to 4 (Zaabar et al. 2014). The Prony series coefficients are
Dynamic FE model determined from the sigmoidal function coefficients using the
MATLAB ‘lsqnonneg’ function. This function uses a special
A 2D axisymmetric FE model was developed using the ABA-
case of the linear least squares formulation with linear inequality
QUS software for use in the forward calculation. The asphalt
constraint, requiring the Prony coefficients to be nonnegative.
concrete layer was modelled as LVE, and damping was con-
sidered for all other layers. Damping is defined as the energy
dissipation behaviour of materials and structures that are sub- Reduced sigmoidal function
jected to time-variable loading. Damping is typically an In equation 2, the sigmoidal function’s exponent b + g. log ( fr )
approximation in the sense that it models the energy absorbing can be re-written in the form of b′ + g. log (f ), where
characteristics of the structure without attempting to model the b′ = b + g.log[a(T)]. This substitution allows the determi-
physical mechanisms that cause them (ABAQUS 2019). nation of the 4 sigmoidal coefficients at any temperature T,
without including a time-temperature superposition model.
Rayleigh damping coefficients Equation 2 can then be re-written as follows:
Damping is considered in the FE model in the form of Rayleigh a
damping, where the damping matrix is formed as a linear com- logE = d + ′ (3)
1 + eb +g. log (f )
bination of the mass and the stiffness matrices according to
equation 1 (Chopra 2012): Figure 2 provides a graphical representation of how each of the
sigmoidal function’s coefficients affects the master curve. In
[C] = a[M] + b[K] (1) addition to the master curve, the inflection point (IP) having
the coordinates shown in equations 4 and 5 is represented by
where: [C] = Damping matrix; [M] = Mass matrix; [K] = Stiff-
a ‘+’ sign for each curve.
ness matrix; α (1/sec.) and β (sec.) = Rayleigh damping
coefficients. b′
Abscissa: log(f ) = − (4)
g
Asphalt concrete linear viscoelastic behaviour d + (d + a)
Ordinate: log|E| =
Several viscoelastic response functions can be used to charac- 2
terise the LVE behaviour of asphalt concrete with the most fun- log (Minimum E) + log(Maximum E)
=
damental ones being the relaxation modulus E(t), creep 2
compliance D(t), and complex modulus E* (Kim 2008). The (5)
A review of Figure 2 shows that the coefficient δ shifts the mas-
master curve is typically constructed by fitting the complex ter curve up or down by the same scale (or ratio of moduli), the
modulus data using a sigmoidal function having the following coefficient α shifts the maximum modulus up or down without
form (equation 2). affecting the minimum modulus, the coefficient b′ shifts the
a inflection point horizontally (left for cooler temperatures and
logE = d + (2)
1 + eb+g. log ( fr ) right for warmer temperatures, when E is plotted vs.
4 G. BAZI AND T. B. ASSI

Figure 2. Effect of sigmoidal coefficients variation: (a) Effect of δ, (b) Effect of α, (c) Effect of β’ and (d) Effect of γ.

frequency), and finally the coefficient γ influences the steepness function is concave or concave downward. The IP coincides
or the slope of the curve at the inflection point. with the maximum slope and zero curvature.
It will be demonstrated later that the master curve, including
the maximum modulus, can efficiently be determined in few
Slope and curvature of the sigmoidal function iterations through dynamic backcalculation whenever the
The first and second derivatives of the sigmoidal function FWD dominant frequencies are to the right of the minimum
shown in equation 3 were determined symbolically in curvature, which is represented by an arrow in Figure 3. It is
MATLAB, and are represented by equations 6 and 7, respect- worth noting that the maximum modulus, located to the far
ively. The curvature k, that is the amount by which a curve right, is outside the range of FWD dominant frequencies.
deviates from being straight line, is calculated using the first
and second derivatives in accordance with equation 8.
Newton’s method for approximating roots
′
∂log(E) a.g.eb +g.log(f ) In numerical analysis, the Newton’s method – also known as
First derivative: =− (6)
∂log(f ) ′
(1 + eb +g.log(f ) )
2
the Newton-Raphson method – is a root-finding algorithm
′ which produces iteratively better approximations to the roots
∂2 log(E) 2a.g2 .e2b +2g.log(f )
Second derivative: = of a function. In optimisation, the Newton’s method is used
∂log(f )2 ′
(1 + eb +g.log(f ) )
3
to determine the stationary points of a function (minima or
′
a.g2 .eb +g.log(f ) maxima) by finding the roots of the function’s first derivative.
− ′ 2 (7) The Newton’s method uses the first few terms of the Taylor
(1 + eb +g.log(f ) )
series (equation 9) of a real-valued function at x = xn .
∂2 log(E)
f ′′ (xn )
∂log(f )2 f (x) = f (xn ) + f ′ (xn ) × (x − xn ) + × (x − xn )2
Curvature k = 
 1.5 (8) 2!
∂log(E) 2 + ... (9)
1+
∂log(f )
For finding the roots of a function, the linear approximation
Figure 3 provides a graphical representation of a sigmoidal (first two terms) of the Taylor series are used and f (x) is set
function along with its slope and curvature plotted vs. fre- equal to zero (equation 10). The offset Dx needed to improve
quency. The slope, which is represented by the first derivative, the root for every iteration is shown in equation 11.
is shown to always be positive. The curvature, which is almost f (x) = f (xn ) + f ′ (xn ) × (x − xn ) = 0 ⇒ x = xn+1
equal to the second derivative, is shown to be positive to the left
of the IP where the sigmoidal function is convex or concave f (xn )
= xn − (10)
upward, and it is negative to the right of the IP where the f ′ (xn )
 INTERNATIONAL JOURNAL OF PAVEMENT ENGINEERING 5

Figure 3. Sigmoidal function along with its slope and curvature.

function f(x) = x 3 + x−1 is better approximated after every iter-

f (xn ) ation starting with an initial guess x0 = −0.6 and using
Dx = xn+1 − xn = − (11)
f ′ (xn ) equations 10 or 11 (Sauer 2012).

For finding the stationary points (minima or maxima), the

function’s first derivative is determined using the quadratic Multivariate Newton’s method for dynamic
approximation (first three terms) of the Taylor series and set backcalculation
 shown in equation 12, is calculated
equal to zero. The offset Dx, The one-variable Newton’s method provides the main outline
using the first and second derivatives. of the multivariate Newton’s method needed in the dynamic
′ backcalculation problem. The analogue of the first derivative
 = − f (xn )
Dx (12) in the one-variable case is the Jacobian matrix, also known
′′
f (xn )
as the gradient matrix. Equations 13 and 14 show the system
The process starts with an initial guess and a better approxi- of equations setup to solve the dynamic backcalculation pro-
mation of the root is obtained after every iteration. This process blem with n evaluation parameters and m variables. Hari-
is repeated as many times as necessary to get the desired accu- chandran et al. (1993) and Chatti et al. (2004a, 2004b) have
racy. Figure 4 provides an illustration of how the root of a used a similar system setup for static and dynamic

Figure 4. Two iterations of Newton’s method for f (x) = x 3 + x−1 = 0 starting with x0 = −0.6 (x1 = 0.27307692 after iteration 1, x2 = 0.85046676, and root = 0.68232780
after iteration 6).
6 G. BAZI AND T. B. ASSI

backcalculation problems, respectively. forward-, backward- and central-difference derivative approxi-

⎡ ⎤ mations can be used, with the latter requiring twice the calls to a
∂P1 ∂P1 i mechanistic model (ABAQUS, in this case). For example, a
⎡ ⎤i ··· ⎡ ⎤i ⎡ ⎤
P1 ⎢ ∂V1 ∂Vm ⎥
⎢ ⎥ ∂V1 P1 dynamic backcalculation problem with 8 variables requires 9
⎢ . ⎥ ⎢ . .. .. ⎥ ⎢ .. ⎥ ⎢ .. ⎥
⎣ .. ⎦ +⎢ .. . . ⎥ ⎣ . ⎦=⎣ . ⎦ (13) calls to ABAQUS to generate the Jacobian matrix using the for-
⎢ ⎥
Pn ⎣ ∂Pn ∂Pn ⎦ ∂Vm Pn
ward- or backward-difference derivative approximations and
··· 16 calls using the central-difference derivative approximation.
∂V1 ∂Vm
The three finite-difference approximation formulas are shown
⎡ ⎤ below for iteration i where h is the step size.
∂P1 ∂P1 i
⎡ ⎤ ⎡ ⎤ ⎡ ⎤i
⎢ ∂V1 · · · ∂Vm ⎥ ∂V1 i P1 P1 f (xi+1 ) − f (xi )
⎢ ⎥ Forward-difference: f ′ (xi ) =
⎢ . . . ⎥ ⎢ .. ⎥ ⎢ .. ⎥ ⎢ . ⎥ xi+1 − xi
⎢ .. .. .. ⎥ ⎣ . ⎦ = ⎣ . ⎦ − ⎣ .. ⎦
⎢ ⎥
⎣ ∂Pn ∂Pn ⎦ ∂Vm Pn Pn f (xi+1 ) − f (xi )
··· = (16)
∂V1 ∂Vm h
⎡ i⎤
P1 − P1 f (xi ) − f (xi−1 )
⎢ .. ⎥ Backward-difference: f ′ (xi ) =
=⎣
. i⎦ xi − xi−1
(14)
Pn − Pn f (xi ) − f (xi−1 )
= (17)
h
where: P 1:n = Calculated parameters; P1:n = Measured par-
∂P1:n f (xi+1 ) − f (xi−1 )
ameters; V1:m = Variables; = First-order derivative Central-difference: f ′ (xi ) =
∂V1:m xi+1 − xi−1
(slope); i = Iteration number. f (xi+1 ) − f (xi−1 )
Equation 14 has the form of AX = b for iteration i, where A = (18)
2h
is the Jacobian (gradient) matrix, X is the offset array, and b is
the error array or the difference between the measured and cal- Figure 5 shows the forward-, backward- and central-differ-
culated parameters ⎡ array. ⎤ ence derivative approximations along with the exact solution
∂V1 i for a function at x = 0.75 using a step size h = 0.5. The step
⎢ ⎥ size h is the length of the interval over which the finite-differ-
The offset array ⎣ ... ⎦ is determined for every iteration
ence approximation is made. The first derivative determined
∂Vm using the central-difference approximation is the closest to
by multiplying the inverse of the Jacobian matrix by the error
⎡ i⎤ the exact solution. Halving the step size approximately halves
P1 − P1 the error of the backward and forward differences and quarters
⎢ .. ⎥
array ⎣
. i⎦
and adding it to the previously determined the error of the centred difference (Chapra 2018).
The step size h should be selected to minimise the total
Pn − Pn
numerical error (Chapra 2018). The total numerical error is
variables. The process is repeated until the changes in the vari-
the summation of the truncation and roundoff errors. The
ables are sufficiently small. The Singular Value Decomposition
truncation errors result from using an approximation in place
(SVD) process is used to invert the Jacobian matrix for deter-
of an exact mathematical procedure, while the roundoff errors
mining the offset array since the system is overdetermined,
arise because digital computers cannot represent some quan-
i.e. it has more rows than columns. The SVD process produces
tities exactly. The truncation error generally increases as the
a solution identical to the least squares method, where the sum
step size increases, while the roundoff error decreases as the
of squared errors (SSE) is minimised. Alternatively, the system
step size increases (Chapra 2018).
shown in equation 14 can be scaled by the magnitude of each
The forward-difference derivative approximation was used
parameter to minimise the root mean squared relative error
in the dynamic backcalculation problem as it requires less
(RMSRE) using the SVD process. The scaling of parameters
calls to the mechanistic model and accurate results were also
is common in optimisation problems, especially when the par-
obtained. The derivative approximation used a step size h
ameters are of differing orders of magnitudes and units. Mini-
equal to 10% the magnitude of each variable. An evaluation
mising the RMSRE objective function in accordance with
of step size found the selection to be optimal.
equation 15, as opposed to SSE, has shown improved stability
One quasi-Newton method, known as the Broyden’s
and faster speed of convergence.
method, was briefly evaluated. This method is more efficient
 2
 n  than the Newton-Raphson method as it only computes the
1  P − P 
RMSRE (%) = 
i i
× 100 (15) whole Jacobian at the first iteration and does rank-one updates
n i=1 Pi at other iterations with only one call to the mechanistic model
for every iteration. This method appears to be promising based
Finite-Difference Approximation of Jacobian Matrix: Whenever on initial evaluations.
the analytical expressions of the partial derivatives are not Evaluated Parameters: Five parameters were considered for
available, the Jacobian matrix is estimated numerically for capturing the shape and magnitude of the deflection time his-
every iteration using the finite-difference approximation. The tory for every FWD sensor. Those parameters, shown in bold in
 INTERNATIONAL JOURNAL OF PAVEMENT ENGINEERING 7

Figure 5. Finite-difference approximations of first derivatives.

Figure 6, include the peak deflection (DPeak) that occurs at a would be an 8 × 1 matrix and the error array would be a
time TPeak; the times to the left and right of the peak that cor- 35 × 1 matrix.
respond to 50% of the peak deflection TLeft and TRight, respect- The RMSRE calculated using equation 15 gives a measure of
ively; and the duration of the pulse at 50% of the peak deflection the total error by considering the five groups of parameters.
calculated as the difference between TRight and TLeft (Dur- The error can be sub-divided to separately quantify the error
ation). The duration, time lag between peak deflections rep- for each of the five groups (DPeak, Duration, etc.).
resented by TPeak, and the magnitudes of the peak deflections
are significantly influenced by the pavement layers variables
(Bazi et al. 2019b). Therefore, the selected parameters can Dynamic backcalculation results
reliably determine those variables through backcalculation. The capabilities of the PULSE application were demonstrated in
The FWD measured data or calculated data using a dynamic this paper for a combination of simulated flexible pavement
FE model are not continuous and are only available at discrete structures and asphalt concrete mixes. Those combinations
intervals, therefore the use of data fitting through high-degree consisted of three flexible pavement structures with varying
polynomial equations and first derivative for determining asphalt concrete layer thicknesses ranging from 7.5–18-cm
TPeak are critical in accurately quantifying all five parameters. for mix A and one flexible structure with 18-cm thick asphalt
For an FWD setup with 7 sensors and a pavement structure concrete layer for mix B. Ten temperatures ranging from −1
requiring the determination of 8 variables, the Jacobian matrix to 37.5°C at 5.5°C intervals with two additional temperatures
would be a 35 × 8 matrix consisting of 35 rows (7 sensors × 5 at 18.5 and 24°C were considered for each structure resulting
parameters) and 8 columns (8 variables), the offset array in a total of 40 combinations as shown in Table 1. The

Figure 6. Sample deflection time history with evaluated parameters.

8 G. BAZI AND T. B. ASSI

Table 1. Flexible pavement structures1.

Asphalt Concrete
Layer Thickness Sigmoidal Function Aggregate Base Layer Subgrade
h1 (cm) Coefficients2 Temperature (°C) Thickness and Properties Properties FWD Pulse Load
7.5 Mix A: δ=−0.127, α=4.523, β= −1, 4.5, 10, 15.5, 18.5, 21, h2 = 30-cm; E2 = 275 MPa E3 = 35 MPa & 40 msec. haversine pulse; FWD load
12.5 −1.188, γ=−0.494 & 24, 26.5, 32 & 37.5 & β = 0.002 β = 0.002 level = 40 kN & radius of loaded
18 Ea=181,000 area = 15-cm
18 Mix B: δ=0.705, α=3.703, β=
−1.417, γ=−0.548 &
Ea=170,500
1
Ei = modulus of elasticity and hi = layer thickness of ith layer; i = 1, 2 and 3 correspond to the surface, base and subgrade layers, respectively.
2
Ea is the activation energy used in the shift factor in accordance with the Arrhenius law (Ea = 181,000 J/mol for mix A & 170,500 for mix B). Reference temperature = 20°C
(293.15 K).

sigmoidal function coefficients were determined for mix A the FWD most dominant frequency. This finding is consistent
using dynamic modulus data for a dense graded mix. Mix B with previous research. Sebaaly et al. (1985, 1986) reported that
is stiffer – representing a mix with harder asphalt binder or the FWD impulse load contains frequencies in the range of 0–
an aged mix. At a frequency of 10 Hz and a temperature of 45.5 Hz as obtained using a Fourier series expansion, with 45%
25°C, mix A has a modulus of 3,250 MPa and mix B has a mod- in the range of 4.5–18.2 Hz. The dominant frequency was
ulus of 6,750 MPa. reported to be around 17 Hz by Kim et al. (2000) and about
10–25 Hz by Chatti and Lei (2012).
The PULSE application improved the variables for every iter-
Backcalculated layer variables
ation, and the process was terminated whenever the error
Each of the 40 combinations had 8 variables – 4 for the asphalt between the theoretical (exact) and calculated asphalt concrete
concrete layer, and 2 for each of the aggregate base and subgrade maximum modulus (10δ+α) dropped below 1% or plateaued.
layers. The exact and seed (starting) values for the eight variables The backcalculation test results are presented in Table 3 for
along with their absolute differences are shown in Table 2. The the 40 combinations. The frequency range (on a log scale), for
maximum difference between the exact and seed values is which the asphalt concrete moduli were determined to an accu-
shown to be as high as 50%, even though the PULSE application racy less than 1%, is shown for every combination. The upper
is very robust and is capable of converging to the exact values frequency is represented by >15 (or frequency > 1015 Hz) for
with larger differences. The percent difference for the asphalt the combinations where the maximum modulus was deter-
concrete master curve is illustrated in Table 2 for two points: mined to less than 1% error.
maximum modulus (equal to 10δ+α), and modulus at a frequency The maximum error in the asphalt concrete moduli at 17 Hz
of 17 Hz and temperature of 20°C. and the unbound layers variables is shown in brackets. Those
The PULSE application was used for the dynamic backcalcula- errors are significantly low for all combinations demonstrating
tion of the variables for each of the 40 combinations. The asphalt the robustness of the method.
concrete moduli at the FWD dominant frequencies, and the
unbound layer moduli and Rayleigh damping coefficients were Asphalt concrete master curve determination
efficiently determined to an accuracy less than 1% after 3–4 iter- The possibility of determining the asphalt concrete master
ations and less than 0.5% after 6–8 iterations. The speed of con- curve is demonstrated in this study. It is shown that the master
vergence can be further improved by starting with closer seed curve can be reliably determined for most of the combinations
values, a process that will be evaluated in future research. (about 60–85%) starting at frequencies around 10−1 or 10−2 Hz
The asphalt concrete moduli were the fastest to convergence (log frequency = −1 or −2). The lowest frequency range was
at frequencies around 17 Hz, indicating that this frequency is also reported by Kutay et al. (2011). It is important to note

Table 2. Actual and seed values of variables.

Layer Variable Exact Value Seed Value Absolute Difference
Asphalt Concrete Sigmoidal coefficient δ Mix A: −0.127 Mix A: −0.161 Maximum E = 10δ±α
Mix B: 0.705 Mix B: 0.539 Mix A: 24%
Mix B: 21%
E at 17 Hz & 20°C
Mix A: 18%
Mix B: 6%
Sigmoidal coefficient α Mix A: 4.523 Mix A: 4.65
Mix B: 3.703 Mix B: 3.95
Sigmoidal coefficient β’ b′ = b + g × log[a(T)] Exact β’ + 0.2
Mix A: −1.188-0.494×log[a(T)]
Mix B: −1.417-0.548×log[a(T)]
Sigmoidal coefficient γ Mix A: −0.494 −0.65
Mix B: −0.548
Aggregate Base Modulus E2 275 MPa 345 MPa 25%
Rayleigh damping coefficient β 0.002 sec. 0.003 sec. 50%
Subgrade Modulus E3 35 MPa 50 MPa 40%
Rayleigh damping coefficient β 0.002 sec. 0.001 sec. 50%
 INTERNATIONAL JOURNAL OF PAVEMENT ENGINEERING 9

Table 3. Dynamic backcalculation frequency range (log) and maximum error. that the FWD pulse does not contain very low frequencies that
Mix A B are adequate for determining the minimum modulus 10δ,
AC Thickness (cm) 7.5 12.5 18 18 unless very close seed values are used.
Temp. (°C) −1 −1.6 to >15 −2.1 to >15 −2.0 to >15 1.3 to 1.8 The potential of determining the master curve appears to be
[0.038%] [0.133%] [0.444%] [1.023%] dependent on the asphalt concrete mix (e.g. mix A vs. B), the
4.5 −3.9 to >15 1.1 to 1.4 1.0 to 2.6 0.3 to >15
[0.011%] [1.386%] [1.181%] [2.747%] asphalt concrete temperature (e.g. low vs. high), and the pave-
10 −3.1 to >15 −3.4 to >15 −3 to >15 −2.4 to >15 ment structure (e.g. h1 = 7.5, 12.5 or 18-cm).
[0.017%] [0.065%] [0.135%] [0.285%] Upon further investigation of the asphalt concrete mixes
15.5 −2.3 to >15 −2.1 to >15 −2.2 to >15 −3.6 to >15
[0.008%] [0.033%] [0.07%] [1.253%] and temperatures, it was found that the maximum modulus
18.5 −2.3 to >15 −1.9 to >15 −1.8 to >15 −1 to >15 can reliably be determined when the FWD dominant frequen-
[0.023%] [0.032%] [0.046%] [0.353%] cies are to the right of the minimum curvature as illustrated
21 −1.9 to >15 −1.5 to >15 −2.2 to 5.7 −0.9 to >15
[0.003%] [0.015%] [0.158%] [0.139%] in Figure 7 for mixes A and B. The points of minimum cur-
24 −1.7 to >15 −1.4 to 11.5 0.3 to 2.2 −0.8 to >15 vature are represented by solid dots on the graph. Note that
[0.001%] [0.005%] [1.27%] [0.052%] the master curves move to the left at lower temperatures
26.5 −1.8 to >15 −1.5 to 7.0 0.6 to 2.1 −0.7 to 6.7
[0.001%] [0.007%] [1.467%] [0.035%] and to the right at higher temperatures, when plotted vs.
32 −2.1 to 6.5 −2.1 to 4.3 0.5 to 2.5 −0.6 to 4.9 frequency.
[0.005%] [0.027%] [0.529%] [0.026%] Figure 7(a) shows the curves for mix A at temperatures of
37.5 −1.3 to 5.9 −1.2 to 5.1 −1.1 to 4.7 0.4 to 2.2
[0.021%] [0.013%] [0.011%] [1.364%] −1, 18.5 and 37.5°C. The master curve and the maximum
modulus were reliably determined through backcalculation

Figure 7. Asphalt concrete master curve prediction for (a) mix A and (b) mix B.
10 G. BAZI AND T. B. ASSI

Figure 8. Asphalt concrete master curve for mix A at reference temperature (20°C) determined using data from two temperatures (15.5 and 26.5°C).

to less than 1% error for the −1 and 18.5°C temperatures since where: r = Distance where the radius of curvature is calculated
the FWD dominant frequencies are to the right of the mini- = 20-cm; D0 = FWD centre deflection; D20 = FWD deflection at
mum curvatures, but this is not the case for the 37.5°C temp- 20-cm radial offset.
erature. Therefore, the backcalculation process has more The effect of the dominant frequencies, mixes and pavement
strength or power in determining the maximum modulus at structures requires further investigation, and a plan to tackle
the lower temperatures as opposed to the higher temperatures. this is recommended in the last section.
Whenever the FWD dominant frequencies are to the left of
the minimum curvature, the maximum modulus is typically Master curve determination at reference temperature
overestimated. The master curve at the reference temperature can be easily
The use of a stiffer mix (mix B in this case) has a positive determined using time-temperature relaxation moduli vs. fre-
effect on the master curve determination. Mix B, for quency relationships at a minimum of two temperatures,
example, has curves shifted upward to the left, as it is clearly which would require FWD testing at two different times. The
shown in Figure 7(b) for the 37.5°C temperature where the development of the asphalt concrete master curve for the thin
minimum curvature moved closer to the FWD dominant flexible structure (7.5-cm thick asphalt concrete) is illustrated
frequencies. This shift, among other factors, produces a wider in Figure 8. The moduli at two temperatures of 15.5 and
temperature range over which the maximum modulus can 26.5°C and at frequencies above 10−2 Hz were used along
reliably be determined. This is illustrated in Table 3 for with the minimum modulus 10δ to determine the master
mixes A and B using the pavement structure with 18-cm curve at a reference temperature of 20°C through data fitting
thick asphalt concrete layer, where the maximum modulus using the Microsoft Excel solver. The minimum modulus 10δ,
was determined to less than 1% error up to 18.5°C for mix A which is dependent on the aggregate gradation, can be deter-
and up to 24°C for mix B. mined from laboratory testing. This modulus is not expected
This study also showed that the temperature range over to change over time due to aging or damage, therefore it can
which the maximum modulus can reliably be determined is be reliably used in this process. On the other hand, the maxi-
extended for the pavement structures with thinner asphalt con- mum modulus, which is dependent on the limiting binder
crete layers. The maximum modulus was determined for mix A modulus (glassy modulus), is expected to vary over time and
to less than 1% error up to 18.5°C for the thick structure (18-cm it can be reliably determined using backcalculation. The error
AC) and up to 26.5°C for the thin structure (7.5-cm AC). This shown in Figure 8 is minimal, as expected, since the minimum
behaviour might be related to the larger deflections or curva- modulus was not varied during fitting and the moduli at fre-
tures for the thin structures, as more valuable information is quencies above 10−2 Hz were determined to less than 1% error.
extracted from the asphalt concrete layer for the backcalcula-
tion process. For example, the centre deflection is 100% larger
and the curvature is 430% larger for the thin structure when Conclusions and recommendations
compared to the thick structure. The curvature is the inverse PULSE , a dynamic backcalculation application, has been devel-
of the radius of curvature R20 (Leger and Autret 1972) calcu- oped and it uses ABAQUS for forward calculation and the
lated in accordance with equation 19. Newton’s root-solving algorithm for improving the variables
after every iteration. PULSE models the asphalt concrete layer
r2
R20 =
 (19) as a linear viscoelastic material and all other layers as linear
D0 elastic with damping. PULSE also has the potential of determin-
2×D0 × −1
D20 ing the asphalt concrete master curve.
 INTERNATIONAL JOURNAL OF PAVEMENT ENGINEERING 11

The capabilities of the PULSE application were demon- Disclosure statement

strated for three simulated flexible pavement structures and No potential conflict of interest was reported by the author(s).
two mixes using 40 combinations. The asphalt concrete
moduli at the FWD most dominant frequencies of around
17 Hz, and the variables of aggregate base and subgrade layers ORCID
(moduli and Rayleigh damping coefficients β), converged to
Gabriel Bazi http://orcid.org/0000-0002-9295-5079
the exact values in a few iterations for all combinations. The Tatiana Bou Assi http://orcid.org/0000-0001-8684-8958
master curves, and mainly the maximum asphalt concrete
moduli, were reliably determined to less than 1% and less
than 4% error for about 60% and 85% of the combinations, References
respectively. It was shown that the potential of determining A2F03 Committee on Management of Quality Assurance, 2002. Glossary
the master curve and maximum modulus is dependent on of Highway Quality Assurance terms. In: Transportation research
the mix, the asphalt concrete temperature and pavement board. Transportation Research Circular No. E-C037. Washington,
structure. It was also shown that the FWD is not sensitive DC: National Research Council.
to very low frequencies, hence the minimum asphalt concrete ABAQUS, 2019. Finite element computer program. Vélizy-Villacoublay:
Dassault Systèmes Simulia Corp.
modulus cannot be reliably determined. The minimum mod-
Al-Khoury, R., et al., 2001. Spectral element technique for efficient par-
ulus, representing the aggregate skeleton, can be determined ameter identification of layered media: Part II: Inverse calculation.
from laboratory testing since it is usually assumed to remain International Journal of Solids and Structures, 38 (48–49), 8753–8772.
unchanged as opposed to the maximum modulus that changes Al-Qadi, I., et al., 2004. The Virginia smart road: the impact of pavement
due to aging and damage. instrumentation on understanding pavement performance. Journal of
the Association of Asphalt Paving Technologists, 73, 427–465.
The PULSE application was used for the backcalculation of an Al-Qadi, I., et al., 2010. Runway instrumentation and response measure-
LTPP section with seasonal FWD testing. The asphalt concrete ments. Transportation Research Record: Journal of the Transportation
moduli at 17 Hz, in addition to the other layers’ variables, were Research Board, 2153, 162–169.
reliably determined. The application was able to capture small Bazi, G., et al., 2019a. Instrumented flexible pavement responses under air-
changes in the layers’ properties due to changes in the asphalt craft loading. International Journal of Pavement Engineering. doi:10.
1080/10298436.2019.1671589.
concrete temperature and unbound layers freezing and thawing Bazi, G., et al., 2019b. Effects of Rayleigh damping on the subgrade’s appar-
conditions. The unbound layer moduli were found to be signifi- ent non-linearity. ASCE Journal of Transportation Engineering, Part B
cantly affected by the asphalt concrete modulus due to changes Pavements, in press.
in confinement. Bazi, G., et al., 2020. Assessment of a low-volume flexible pavement
The Newton-Raphson’s method was shown to be very through dynamic backcalculation. In: 99th Annual meeting of the trans-
portation research board, Poster Session 1731, Washington, DC.
powerful and sensitive to small variations in the parameters.
Chapra, S.C, 2018. Applied numerical methods with MATLAB for engineers
This algorithm requires multiple calls to the ABAQUS software and scientists. New York, NY: McGraw-Hill Education.
to generate the Jacobian matrix. The Broyden’s method, a Chatti, K., et al., 2017. Enhanced analysis of falling weight deflectometer
quasi-Newton method, was briefly evaluated as it requires data for use with mechanistic-empirical flexible pavement design and
one call to ABAQUS to update the Jacobian matrix for every analysis and recommendations for improvements to falling weight
deflectometers. McLean, VA: Turner-Fairbank Highway Research
iteration. The Broyden’s method appears to be promising Center.
based on initial evaluations and will be further investigated. Chatti, K., Ji, Y., and Harichandran, R., 2004a. Dynamic time domain
Future research will focus on (i) quantifying the dominant backcalculation of layer moduli, damping, and thicknesses in flexible
frequencies of the FWD pulses, (ii) extracting more infor- pavements. Transportation Research Record, 1869 (1), 106–116.
mation from the FWD deflection time histories, (iii) under- Chatti, K., Ji, Y., and Harichandran, R.S., 2004b. Development of a
dynamic backcalculation computer program for flexible pavement
standing and forecasting the prediction strength of the layer parameters. Final report. Michigan Department of
master curve and maximum moduli using Taylor series, (iv) Transportation, Lansing.
applying the method for the master curve determination Chatti, K. and Lei, L., 2012. Forward calculation of subgrade modulus using
using field measured FWD data, and (v) testing the application falling weight deflectometer time histories and wave propagation theory.
over a wider range of flexible pavement structures and GeoCongress 2012: State of the Art and Practice in Geotechnical
Engineering (pp. 1400–1409).
conditions.
Chopra, A.K, 2012. Dynamics of structures. Upper Saddle River, NJ:
Previous research has compared the asphalt concrete master Prentice Hall.
curve determined from field measured data to the relaxation Garg, N., Pecht, F., and Jia, Q., 2010. Subgrade stress measurements under
modulus master curve determined from laboratory testing heavy aircraft gear loading at FAA national airport pavement test facil-
through creep compliance or complex modulus data. It is ity. In: B. Huang, E. Tutumluer, I. L. Al-Qadi, J. Prozzi, and X. Shu, eds.
Paving materials and pavement analysis: Proceedings of the
important to note that differences are expected when compar- GeoShanghai 2010 international conference, Reston, VA, pp. 484–491.
ing field- to lab-measured data or laboratory data among each Gopalakrishnan, K., et al., 2014. Development of asphalt dynamic modulus
other. For example, Kim (2008, p. 113 figure 4–16) reported master curve using falling weight deflectometer measurements, TR-659.
that significant differences exist in asphalt concrete complex Gopalakrishnan, K. and Thompson, M., 2006a. Assessing damage to airport
modulus when testing is performed under confined and pavement structure due to complex gear loads. Journal of Transportation
Engineering, 132 (11), 888–897.
unconfined conditions. Therefore, alternative methods shall Gopalakrishnan, K. and Thompson, M., 2006b. Evaluation of traffic
be evaluated to validate the field-determined asphalt concrete induced subgrade stresses in flexible airfield pavements. Road
master curve. Materials and Pavement Design, 7 (4), 457–476.
12 G. BAZI AND T. B. ASSI

Harichandran, R.S., et al., 1993. Modified Newton algorithm for backcal- MATLAB version, 2018. The MathWorks, Inc. Natick, MA.
culation of pavement layer properties. Transportation Research Matsui, K., et al., 2000. Comparative studies of backcalculated results from
Record, 1384, 15. FWDs with different loading duration. In: Nondestructive testing of
Kang, Y.V, 1998. Multifrequency back-calculation of pavement-layer pavements and backcalculation of moduli: Third Volume. West
moduli. Journal of Transportation Engineering, 124 (1), 73–81. Conshohocken, PA: ASTM International.
Kim, Y.R, 2008. Modeling of asphalt concrete. New York, NY: ASCE Press, Sauer, T, 2012. Numerical analysis. Boston, MA: Pearson.
McGraw Hill. Sebaaly, P.E., et al., 2002. Effects of off-road tires on flexible and
Kim, Y.R., Xu, B., and Kim, Y, 2000. A new backcalculation procedure granular pavements. Pierre, SD: South Dakota Department of
based on dispersion analysis of FWD time-history deflections and sur- Transportation.
face wave measurements using artificial neural networks. In: Sebaaly, B., Davis, T.G., and Mamlouk, M.S., 1985. Dynamics of falling
Nondestructive testing of pavements and backcalculation of moduli: weight deflectometer. Journal of Transportation Engineering, 111 (6),
third volume. West Conshohocken, PA: ASTM International. 618–632.
Kutay, M.E., Chatti, K., and Lei, L, 2011. Backcalculation of dynamic mod- Sebaaly, B.E., Mamlouk, M.S., and Davies, T.G. 1986. Dynamic analysis of
ulus mastercurve from falling weight deflectometer surface deflections. falling weight deflectometer data. Transportation Research Record,
Transportation Research Record, 2227 (1), 87–96. 1070, 63–68.
Lasdon, L.S., et al., 1976. Design and testing of a generalized reduced gra- Turkiyyah, G.M, 2005. Feasibility of backcalculation procedures based on
dient code for nonlinear programming. Technical Report SOL 76-3, dynamic FWD response data. Olympia, WA: Washington State
Systems Optimization Laboratory, Stanford University, CA. Department of Transportation.
Leger, P. and Autret, P., 1972. The use of deflection measurements for the Ullidtz, P., 1998. Modelling flexible pavement response and performance.
structural design and supervision of pavements. In: Proceedings of the Lyngby, DK: Tech Univ. of Denmark Polytekn.
Third International Conference on the structural Design of asphalt Uzan, J., 1994a. Advanced backcalculation techniques. In: ASTM, ed.
pavements, London, Vol. 1. Nondestructive testing of pavements and backcalculation of moduli:
Ling, C., David, E.N., and Siddharthan, R., 1991. Comparison of dynamic second Volume. Philadelphia, PA: ASTM International, 3–37.
and static backcalculation moduli for three-layer pavements. Uzan, J., 1994b. Dynamic linear back calculation of pavement
Transportation Research Record, 1293, 86–92. material parameters. Journal of Transportation Engineering, 120 (1),
Losa, M., 2002. The influence of asphalt pavement layer properties on 109–126.
vibration transmission. International Journal of Pavements, 1 (1), 67– Wang, H., et al., 2013. Three-dimensional finite element modelling of
76. Sao Paulo, Brazil. instrumented airport runway pavement responses. Transportation
Luo, F., 2008. Dynamic backcalculation of pavement properties using optim- Research Record: Journal of the Transportation Research Board, 2367
ization in nondestructive evaluation. PhD Dissertation, The Catholic (1), 76–83.
University of America. Zaabar, I., et al., 2014. Backcalculation of asphalt concrete modulus master
Magnuson, A.H., Lytton, R.L., and Briggs, R.C, 1991. Comparison of com- curve from field-measured falling weight deflectometer data: using a
puter predictions and field data for dynamic analysis of falling weight new time domain viscoelastic dynamic solution and genetic algorithm.
deflectometer data. Transportation Research Record, (1293), 61–71. Transportation Research Record, 2457 (1), 80–92.