Construction and Building Materials 76 (2015) 221–231

Contents lists available at ScienceDirect

Construction and Building Materials

journal homepage: www.elsevier.com/locate/conbuildmat

New predictive models for the dynamic modulus of hot mix asphalt
Maryam S. Sakhaeifar a,⇑, Y. Richard Kim b,1, Pooyan Kabir a,2
a
Department of Civil Engineering, Texas A&M University, 3136 TAMU, College Station 77843-3136, United States
b
Department of Civil, Construction & Environmental Engineering, North Carolina State University, Campus Box 7908, Raleigh, NC 27695-7908, United States

h i g h l i g h t s

⁄
 Developing two models for predicting the |E | of hot mix asphalt mixtures.
 Modeling is based on viscoelastic and time–temperature superposition concepts.
⁄
 The models use different mixture and binder properties to predict |E | values.
 Witczak and Hirsch models were evaluated using an exhaustive database.
 New models can be used in performance analyses of asphalt concrete pavements.

a r t i c l e i n f o a b s t r a c t

Article history: This paper presents a fundamental modeling framework for prediction of dynamic modulus of hot mix
Received 30 May 2013 asphalt mixtures based on viscoelastic principles. The outcomes are two closed-form models that can
Received in revised form 10 September be used to predict the mixture dynamic modulus for a wide range of temperatures (10°, 4.4°, 37.8°,
2014
and 54.4 °C) recommended in the American Association of State Highway and Transportation Officials
Accepted 9 November 2014
Available online 15 December 2014
(AASHTO) TP62-03 test protocol. To develop and verify the models a large database that covers the com-
plete range of potential input conditions was assembled. In general, the proposed models predict the
dynamic modulus with a very good level of accuracy.
Keywords:
Hot mix asphalt concrete
Ó 2014 Elsevier Ltd. All rights reserved.
Dynamic modulus
Binder shear modulus
Time–temperature superposition
Mastercurve
Predictive model
Mechanistic-Empirical Pavement Design
Guide (MEPDG)

1. Introduction ranging from laboratory tests on HMA mixtures and binders to

The dynamic modulus, |E⁄|, is a fundamental property that
defines the stiffness characteristics of hot mix asphalt (HMA) mix-
tures as a function of loading rate and temperature. The signifi-
cance of this material property is threefold. First, it is one of the
primary material property inputs in the Mechanistic-Empirical
Pavement Design (MEPDG) and software developed by NCHRP Pro-
ject 1-37A [1]. The MEPDG uses a mastercurve and time–tempera-
ture shift factors in its internal computations. In the MEPDG, the
mastercurve is constructed using a hierarchical structure of inputs
⇑ Corresponding author. Tel.: +1 (979) 845 9961; fax: +1 (979) 458 0780.
E-mail addresses: msakhaeifar@tamu.edu (M.S. Sakhaeifar), kim@ncsu.edu (Y.
Richard Kim), pk6472@tamu.edu (P. Kabir).
1 der as their primary inputs. In addition to these predictive models,
Tel.: +1 (919) 515 7758; fax: +1 (919) 515 7908.
2
Tel.: +1 (469) 258 8867.
estimates based on properties of the HMA mixtures. Second, the
|E⁄| is one of the primary HMA properties measured in the Super-
pave simple performance test protocol that complements the vol-
umetric mix design. Third, the |E⁄| is one of the fundamental linear
viscoelastic (LVE) material properties that can be used in advanced
HMA and pavement models that are based on viscoelasticity.
Researchers developing the MEPDG implemented a hierarchical
input structure in recognition of the fact that the |E⁄| values for
materials used in a particular design might not be available when
the analysis is performed. In the lowest level of this structure, users
may choose to utilize a predictive equation based on mixture vol-
umetric and binder properties to predict the mixture modulus [1].
For this purpose two different versions of Witczak predictive mod-
els have been used which use viscosity and shear modulus of bin-
Witczak models, others have developed predictive models, and

http://dx.doi.org/10.1016/j.conbuildmat.2014.11.011
0950-0618/Ó 2014 Elsevier Ltd. All rights reserved.
222 M.S. Sakhaeifar et al. / Construction and Building Materials 76 (2015) 221–231
those that have gained some national interest are the Hirsch model
[2], a modified or latest form of Witczak’s original model [3]
referred to as the modified Witczak model, and a modified form of
the Hirsch model referred to as the Al-Khateeb model [4]. These
regression models also tend to have some limitations when applied
for a wide range of mixture properties and at high and low temper-
atures. Various researchers have reviewed these models using
independent data sets and have concluded that, while both have
advantages and disadvantages in terms of necessary inputs and
ease of use, the predictive capabilities of each could be improved
[5–8]. To capture the behavior of a HMA mixture fully, very low
and high temperature zones are critical and need to be taken into
account appropriately.
The use of Artificial Neural Network (ANN) techniques for pre-
dicting the mixture |E⁄| has also been undertaken by few research
efforts [9–11]. Ceylan et al. developed a series of ANN-based mod-
els and compared the predictions with the ones estimated from
few existing dynamic modulus predictive models [9]. The results
showed that ANN-based |E⁄| model showed a better performance
than Witczak model both in terms of fitness and bias [9]. Also,
the model demonstrated an improved sensitivity to mixture vari-
ables at different temperatures and frequencies of testing.
Singh et al. developed a model that utilizes aggregate shape
parameters (i.e., angularity, texture and form) in estimating the
dynamic modulus of asphalt mixes. The performance of this model
was compared with the widely accepted Witczak model that does
not use shape parameters of the aggregates. The results indicated
that the mean average relative error for the Witczak model was
estimated significantly higher than the developed model [12].
Dai developed a micromechanical finite-element (FE) model for
predicting the dynamic modulus and phase angle of asphalt mix-
tures. The simulation results of the asphalt mixture samples
showed good correlations with the numerical calibration of asphalt
mastic specimen. The results of this study also indicated that the
developed micromechanical FE model can provide a computational
tool for predicting the global viscoelastic properties of asphalt mix-
tures with captured microstructure and ingredient properties [13].
You et al. utilized a clustered distinct element method for model-
ing asphalt concrete microstructure to predict compressive
dynamic modulus of asphalt mixtures. This modeling approach
showed promising results in terms of estimating the dynamic
modulus and was implemented to portray bulk material behavior
in conjunction with fracture models to study crack behavior in
hot mix asphalt as well [14].
In this study, two close-form models are developed based on a
procedure that employs viscoelastic and time–temperature super-
position (t–TS) concepts, appropriate treatment of data censoring,
identification of dominant parameters, and efficient use of model-
ing techniques to thoroughly characterize the material behavior
under different testing conditions. These models use HMA mixture
gradation, volumetric and binder rheological properties, such as
the binder shear modulus (|G⁄|), and test conditions of |E⁄| mea-
surements. The laboratory test data used are a combination of mul-
tiple comprehensive databases based on the American Association
of State Highway and Transportation Officials (AASHTO) TP62-03
and asphalt mixture performance tester (AMPT) test protocols.
Also, the databases include test measurements from different aging
conditions and various types of modified and unmodified binders.
2. Objective
The objective of this research is to develop regression predictive
models for the dynamic modulus of HMA mixtures based on visco-
elastic concepts and using different physical and mechanical prop-
erties of mixture and binder.
3. Database
One of the most comprehensive database existing, for both bin-
der and mixture moduli, is the one used to develop the modified
Witczak model [3]. This database is extensive and covers a wide
range of material characteristics. At the outset of this study, it
was felt that such a database was sufficient to meet the research
goals. However, a careful, in-depth evaluation of the database
has identified three major issues that affect its usability:
 Different definitions of frequency for the binder |G⁄| and mixture
|E⁄| data are used;
 Estimated, rather than measured, |G⁄| and phase angle data are
used to populate the database; and
 The estimated |G⁄| data at temperatures lower than or equal to
4.4 °C are estimated in a manner that is inconsistent with other
temperatures; in conjunction with this inconsistency is that the
method of estimation is not satisfactorily explained.
When the modified Witczak model was developed and thus
when the Witczak database was populated with binder |G⁄| data,
the developers used what is referred to in this paper as an inconsis-
tent definition of frequency. In short, this means that if one wants to
predict the mixture |E⁄| using the modified Witczak model at, for
example, 25 Hz at 20 °C then that person would have to input into
the model the binder |G⁄| measured at 25 rad/s at 20 °C. The reason
for these differences appears rooted in the conventions of the time
definition typically applied in binder rheology research and in
asphalt mixture research, given in Eqs. (1) and (2) respectively.
The developers of the database required that the frequency used
for binder should be such that the time by the binder definition
was equal to the time defined for the mixture. The argument
regarding the appropriateness of these definitions is well beyond
the scope of this paper; however, it is important to understand that
this is state of data in the Witczak database.
1
tbinder ¼ ð1Þ
x
1
tmixture ¼ ð2Þ
f
where x is the angular velocity in radians per second, t is time and f
is the frequency in hertz.
A more logical approach, which incidentally is used in all of the
other |G⁄|-based models, is to use a consistent definition for the fre-
quencies. That is to say if one wants to predict the mixture |E⁄|
using a model at, for example, 25 Hz at 20 °C then that person
would input in the model the binder |G⁄| measured at 25 Hz at
20 °C. For this effort, the complete Witczak database was repopu-
lated which consumed substantial effort. It is critical that the
reader understands that all possible efforts have been taken to
ensure that the proper definitions of frequency (inconsistent or
consistent) are used for the respective models. This means that
all databases have been populated twice, once using the inconsis-
tent model for predictions with the modified Witczak model and
once using the consistent definition for new model development
and evaluation of Hirsch model.
Most of the database used to develop the new models contains a
comprehensive set of test data that was assembled and used as
part of the long-term pavement performance (LTPP) project [15],
which itself is based on 14,341 test data points from 801 HMA mix-
tures. This database later was expanded by assembling the
expanded North Carolina State University (NCSU) database. Now,
the database includes 20,209 data points from 1008 mixes tested
under different aging conditions and includes both modified and
unmodified binders measured based on two test protocols, i.e.,
 M.S. Sakhaeifar et al. / Construction and Building Materials 76 (2015) 221–231 223

the AMPT and TP-62 protocols. This expanded database is referred
to here as the ‘‘Master NCSU’’ database and consists of the
expanded NCSU, processed Witczak, Federal Highway Administra-
tion (FHWA) Mobile Trailer, WRI and Citgo databases. The Master
database was subdivided into two subsets that are used for cali-
brating and testing the performance of the new developed models.
A large portion of the database that contains the data tested at
NCSU covers a wide range of testing conditions and includes 223
test results and was used to develop the new models. The remain-
ing 10% of the expanded NCSU database, together with other avail-
able datasets, were used to verify the models. The NCSU database
data was gathered in a more recent project and then used for cal-
ibration of the model and a part of NCSU database and older dat-
abases used for checking the accuracy of the model. A summary
of the calibration NCSU database is presented in Table 1, and
details regarding the verification databases can be found elsewhere
[5].
4. Model formulations
4.1. Mixture and binder mastercurve construction
M
f R ¼ f  aM
T ð4Þ
logðaT Þ ¼ a1 T 2 þ a2 T þ a3 ð5Þ
aT = temperature-dependent shift factor,
a1, a2, a3 = polynomial coefficients,
f = frequency of loading at the desired temperature,
T = temperature of interest.
Also, the binder mastercurves were constructed using a full suite
of dynamic shear rheometer (DSR) test results. Significant literature
is also available on this concept [16,17]. In this study, Eq. (6) is used
for the mastercurve construction, and Eq. (8) is used for the t–T shift
factor function. Eq. (6) is known as the Christensen–Anderson–Mar-
asteanu (CAM) model [18], and Eq. (8) is the well-known Williams–
Landel–Ferry (WLF) and Arrhenius models. Rheological properties
of binders at specific temperature and frequency, similar to mixture
temperature and frequency, were obtained by using CAM model. In
this process typical recommended values for R and Ea,
8.314  103 kJ K1 mol1 and 189.879 kJ/mol respectively,
obtained from the original reference were used.

The mechanical behavior of HMA mixtures is governed basically Gg 109

jG j ¼  ¼ ð6Þ
by the viscoelastic nature of the asphalt binder. The specific char-  k mke   k mke
fc
acterization of asphalt concrete and its material properties, such as 1 þ fR 1 þ ff Bc
R
the dynamic modulus, is based on the dependence of its behavior
on both the time of loading and temperature, which can be merged B
into a single smooth sigmoidal function. The following sigmoidal f R ¼ f  aBT ð7Þ
function, Eq. (3), is considered for this purpose.
8   9
b < Ea 1 1
 273 þ CC 12 T R
T  0=
 2:303R 273þT T R
log jE j ¼ a þ 1
ð3Þ log aBT ¼ ð8Þ
1þ : C1 ðTT R Þ T > 0;
e ð Þ
dþg log f M
R C 2 þTT R

|E⁄| = dynamic modulus (psi), |G⁄| = dynamic shear modulus (Pa),

a = first sigmoidal function coefficient, minimum value of
log |E⁄|,
b = second sigmoidal function coefficient,
d, g = parameters describing the shape of the sigmoidal func-
tion, and
M
f R = mixture reduced frequency of loading at the reference tem-
perature (Hz).
The effects of time/frequency and temperature can be consid-
ered together in a single parameter referred to as reduced time or
reduced frequency. These parameters can be combined by using
the t–T shift factor defined in Eq. (5).
Gg = maximum shear modulus; glassy modulus (109 Pa),
T = test target temperature (°C),
TR = reference temperature (20 °C),
R = gas constant (8.314  103 kJ K1 mol1),
Ea = activation energy, determined through optimization
(189.879 kJ/mol),
B
f R = binder reduced frequency of loading at the reference tem-
perature (Hz), and fc, me, k, C1 and C2 = fitting coefficients.
Because the glassy modulus may be several orders of magnitude
gre0ater than the other coefficients, the value of this parameter is
normalized by taking its logarithm in order to fit the model [19].

Table 1
Summary of |E⁄| data available in NCSU calibration mixture database.

Binder code No. of different mixes Total no. of tests Temp. (°C) Frequency (Hz) Aging condition
STOAa LTOAa or field
ALF-CRTB-76 1 5 10, 10, 35, 54 25, 10, 5, 1, 0.5, 0.1, 0.05, 0.01 5
ALF-SBSLG-70 1 3 3
ALF-Terpolymer-70 1 4 4
Citgo-Wilmington-64 42 118 111 7
MRL AAD-1–58 4 11 3 8
AA-Inman-64 3 9 9
Citgo-Wilmington -70 15 37 37
AA-Salisbury-76 1 3 3
El Paso-Charlotte-64 2 6 6
El Paso-Apex-64 1 3 3
AA-Salisbury-70 1 3 3
Patriot-AA-Greensboro-70 1 12 12
Citgo-Charlotte-70 1 3 3
Citgo-Charlotte-64 1 3 3
AA-Salisbury-64 1 3 3
a
STOA: Short term oven aging; LTOA: Long term oven aging.
224 M.S. Sakhaeifar et al. / Construction and Building Materials 76 (2015) 221–231

For all cases presented in this model development, 20 °C is taken as measured at a particular frequency (10 rad/s) and multiple tem-
the reference temperature. peratures, which may not be enough to create a full binder shear
modulus mastercurve. Therefore, because of this limitation, only
4.2. First modeling approach the exponential submodel was split from the global CAM–|E⁄| mul-
tivariate nonlinear function. This split function method increases
The formulation of the new dynamic modulus model is based the calculation speed in the analysis of nonlinear multivariate
on combining the relationships of the binder and mixture master- functions. In addition, this method provides high-order numerical
curves. In order to do so, the binder mastercurve is solved for solutions and error monitoring capabilities. In the next step, the
reduced frequency, which is a common t–T parameter used for effect of the mixture-to-binder shift factor ratio is parameterized
both binder and mixture mastercurves. It should be mentioned by the average shift factor developed for both mixtures and binders
that in all the following derivations, the subscripts ‘‘M’’ and ‘‘B’’ instead of by individual shift factors. The optimized coefficients for
refer to mixture and binder, respectively. The derived relationship the best fit average shift factor curves are shown in Fig. 1 and are
for the binder reduced frequency is presented in Eq. (9). presented in Eqs. (13) and (14) for the mixtures and binders from
calibration database, respectively.
B fc
fR ¼  1k ð9Þ
 mk log ðaT Þavg; Mixture ¼ 0:0007T 2  0:1592T þ 2:9177 ð13Þ
Gg e
jG j
1
log ðaT Þavg; Binder ¼ 0:0011T 2  0:1727T þ 3:018 ð14Þ
Using the same reference temperature, the amount of shifting at
⁄
each temperature required to form a single mastercurve is slightly The adoption of the average shift factor in the CAM–|E | model
different for the mixture and binder used in that specimen. has been evaluated and shows very good statistical fit for the pre-
Because of this difference, a correction factor has been incorpo- dicted dynamic modulus values obtained from the CAM–|E⁄| model
rated. This correction factor is introduced as the ratio of mixture- versus the measured values using calibration database. Best fit
to-binder shift factor and is presented in Eq. (10). For simplicity CAM mastercurves for the binders are particularly necessary for
the second-order polynomial, expressed in Eq. (5), is used in the cases such as specification binder data, where only limited binder
CAM–|E⁄| relationship. Thus, using the shift factor formulations shear modulus measurements are available. Using the calibration
presented in Eq. (5), the following expression in Eq. (10) can be database, the term expressed in Eq. (16) which is used in Eq. (15)
introduced as the correction factor. as the reduced frequency was found. This term is a function of
2 2 2 2
the binder shear modulus and is referred to as XG to distinguish
M
aM
T f f 10a1M ðT T R Þþa2M ðT T R Þ 2 2
the new mathematical sigmoidal formulation. It is important to
s¼ B
¼ RB ¼ ¼ 10ða1M a1B ÞðT T R Þþða2M a2B ÞðTT R Þ
aT f f R 10a1B ðT 2 T 2R Þþa2B ðT 2 T 2R Þ note that the fixed Gg value of 1 GPa was used also to process the
ð10Þ binder database; therefore, all the fitted binder shear modulus val-
ues are limited to 1 GPa (145,000 psi). The binder modulus are lim-
The combination of Eqs. (9) and (10) gives the reduced fre- ited to 1 GPa because of its rheological properties at different
quency of the mixture, the reduced frequency is used as the temperatures and this should be considered when using this
reduced frequency term in Eq. (3) and the combined relationship, model. In the next step, the new sigmoidal coefficients for con-
called the CAM–|E⁄| model, is derived and presented in Eq. (11): structing the mixture mastercurve are fitted again for the individ-
b ual mixtures. Fitting a new mastercurve for each mixture is
log jE j ¼ a þ 2 2 33 ð11Þ necessary once the average binder mastercurve is introduced in
6 6 77 the CAM–|E⁄| model. The modified sigmoidal function is now a
6 6 sfc 77 function of the average reduced frequency term and is introduced
1 þ exp 6d  g log 6  1 77
4 4  mke k 55
with new terms in Eq. (15).
Gg
jG j
1
b
log jE j ¼ a þ ð15Þ
The measured dynamic modulus values are then compared with 1 þ ecþdX G
the fitted values based on the CAM–|E⁄| model, Eq. (11). The CAM– 0 1
|E⁄| relationship was considered as multivariate nonlinear function.
B2:4392  100:0004T 2 þ0:0135T0:1003 C
Since, the search for extremes in a multivariate function can be B C
X G ¼ log B  0:12332 7:72273 C ð16Þ
very difficult, sometimes the function is split into parts; each hav- @ 145;000
A
ing a separate set of variables [20]. The prediction of the frequency jG j
1
in the sigmoidal function presented in Eq. (11) is referred to as the
exponential sub model, expressed in Eq. (12), and was considered
8
for the purpose of the proposed method. y = 0.0007x 2 - 0.1592x + 2.9177 Mixture Shift Factor
0 1 6 y = 0.0011x 2 - 0.1727x + 3.018 Binder Shift Factor
B C
B s  fc C 4
X G ¼ log B
B  k
C
1k C ð12Þ
@ Gg m e A
log (a T)

2
jG j
1
0
Also, it is critical to consider that a minimization algorithm is as -10 0 10 20 30 40 50 60
-2
likely to converge to a local minimum as to a global one [20].
Therefore, in order to avoid multiple local minima and the problem -4
these local minima may cause, which would make the search for
the roots of the target functions practically useless, only the essen- -6
Temperature (°C)
tial part of the function is split. Practically, many of the users of the
final predictive models have only limited specification binder shear Fig. 1. Comparison of individual and average t–T shift factor functions for the
modulus data from the laboratory. In general, these test results are mixtures and binders using the NCSU calibration database.
 M.S. Sakhaeifar et al. / Construction and Building Materials 76 (2015) 221–231 225

a = location parameter for the |E⁄| that indicates the minimum
⁄
value of the log |E |,
b = range of possible values to be added to a, to give the maxi-
mum value of log |E⁄| (a + b),
c, d = shape coefficients in the sigmoidal function, and
XG = predictor variable that corresponds to the average binder
mastercurve using the calibration binder database.
The performance of the average binder mastercurve using the
calibration binder database is shown versus reduced frequency for
both the calibration (Fig. 2a) and verification datasets (Fig. 2b–d).
It is observed that the optimized average binder mastercurve is
within the range of all the individual binder mastercurves.
Therefore, the CAM–|E⁄| relationship presented in Eq. (15) was
considered to develop the new closed-form models and the upper
and lower asymptotes of the sigmoidal terms were expressed by
appropriate mixture properties. The new closed-form model gen-
erated based on this procedure is called the global model, details
of which are described after the section related to second modeling
approach.
4.3. Second modeling approach
Based on the findings from the first modeling approach, a new
modeling procedure was considered for developing the second
closed-form model. From the first modeling approach it was found
that the independent variable in the sigmoidal function (XG) is a
function of the logarithmic transformation of the binder shear
modulus and can be substituted with log |G⁄|, indicated in Eqs.
(17) and (18). Also, as in the first modeling approach, once the
independent variable is substituted with the new term of log |G⁄|,
the new fitted sigmoidal terms were found for all the individual
mixtures.
b sieves: p3/4 (%), p3/8 (%), and p4 (%) respectively,
log jE j ¼ a þ
1 þ ecþdX G
X G ¼ logðjG jÞ ð18Þ
The new fitted low and high asymptotes can be expressed by
different sets of material properties for predicting the dynamic
modulus values in the second proposed model, simplified global
model, described in the following sections.
5. Submodels
The mathematical sigmoidal function adopted for developing
new models consist of four distinct parts, a, b, c and d, where the
two variables, c and d, are considered as the exponent terms. The
logarithmic scale of the dynamic modulus, log |E⁄|, is considered
to be a dependent variable and two different functions of binder
shear modulus are found to be the appropriate independent vari-
ables in these two modeling approaches. The two parts, a and b,
are then modeled with a dominant set of predictor variables that
constitute a partial combination of gradation and volumetric
parameters. These two parts of the sigmoidal function are referred
to here as the ‘‘a’’ and ‘‘b’’ submodels, and as mentioned earlier, the
other two terms, c and d, are referred to as the exponent terms.
5.1. Predictor variables used in the submodels
In order to develop a robust model, it is necessary to investigate
and evaluate a set of appropriate predictor variables that enable
the model to be comprehensive and capable of estimating the
dynamic modulus values under a wide range of test conditions
and mixture and binder properties. Therefore, the available predic-
tor variables considered for model development include:
 Binder shear modulus, |G⁄|b (psi)
 Binder phase angle
 Cumulative percentage retained on the 3/400 , 3/800 , and #4
ð17Þ
 Aggregate passing the #200 sieve (0.075 mm), p200 (%)

1.E+05 1.E+05
Individual Binder Mastercurve Individual Binder Mastercurve
1.E+04 Average Binder Mastercurve 1.E+04 Average Binder Mastercurve

1.E+03 1.E+03
|G*| (psi)
|G*| (psi)

1.E+02 1.E+02

1.E+01 1.E+01

1.E+00 1.E+00

1.E-01 1.E-01
(a) Calibration Data (b) Verification Data
1.E-02 1.E-02
1.E-05 1.E-03 1.E-01 1.E+01 1.E+03 1.E+05 1.E+07 1.E-05 1.E-02 1.E+01 1.E+04 1.E+07 1.E+10
Reduced Frequency (Hz) Reduced Frequency (Hz)

1.E+06 1.E+05
Individual Binder Mastercurve Individual Binder Mastercurve
1.E+05 Average Binder Mastercurve 1.E+04 Average Binder Mastercurve

1.E+04 1.E+03
|G*| (psi)
|G*| (psi)

1.E+03
1.E+02
1.E+02
1.E+01
1.E+01
1.E+00
1.E+00
1.E-01 1.E-01
(c) Verification Data (d) Verification Data
1.E-02 1.E-02
1.E-05 1.E-02 1.E+01 1.E+04 1.E+07 1.E+10 1.E-05 1.E-03 1.E-01 1.E+01 1.E+03 1.E+05
Reduced Frequency (Hz) Reduced Frequency (Hz)

Fig. 2. Comparison of individual and average |G⁄| mastercurves for: (a) NCSU calibration and verification databases; (b) 10% NCSU, (c) processed Witczak and (d) AMPT
databases.
226 M.S. Sakhaeifar et al. / Construction and Building Materials 76 (2015) 221–231

 Air voids (by Vmix), Va (%)  Term a: gradation: p3/4, p3/8, p4, p200, volumetrics: Va, Vbeff and
 Effective asphalt content (by Vtotal), Vbeff (%) VMA.
 Voids in mineral aggregate, VMA (%)  Term b: gradation: p3/4, p3/8, p200, volumetrics: Va, Vbeff and VMA.
 Voids filled with asphalt, VFA (%)
5.2. Optimization of the a and b submodels
Explanatory analysis of the independent variables is performed
using Pearson multivariate correlation analysis to determine which A forward stepwise regression process was run to find a func-
parameters correlate most strongly with the terms ‘‘a’’ and ‘‘b’’ in tion with suitable variable terms. This stepwise regression analysis
the sigmoidal function. The correlation analysis indicates that was conducted for the calibration database using all the variables
some variables, such as Va and Vbeff, have the poorest correlation in terms of linear and second order polynomial regression func-
with each other among all the other volumetric parameters. In tions and also the exponential function for air voids. It is found that
addition, knowing that by having specific gravity, two volumetric other mathematical function such as the exponential function for
parameters are sufficient for calculating the other volumetrics, the variables like percentage of air voids, exp(Va), does not improve
only Va and Vbeff were considered for the model development. In the accuracy of the developed model. This set of trials resulted in
developing the predictive models, some of the predictor variables low or high limiting values for predicting the dynamic modulus
are used interchangeably to capture the interactions amongst the values. Therefore, other mathematical functions rather than linear
variables more effectively. For example, in the modified Witczak and second order polynomial regression functions were ignored for
model, the term Vbeff/(Va + Vbeff) is used to present the volume con- further investigation and for development of subsequent pilot
centration of the aggregate and binder [3]. This term was also con- models. The interaction terms for presenting the full quadratic
sidered for the same purpose as in the modified Witczak model. function were also found not to contribute significantly to the
The results of the correlation for a and b versus the predictor vari- accuracy of the predictive model. The two terms, ‘‘c’’ and ‘‘d’’, in
ables are shown in Fig. 3a and b, respectively. The results indicate the sigmoidal function describe the shape of the function and are
that the two terms, a and b, in the sigmoidal function correlate at a found through the optimization process. The results of all investi-
significant 90% confident interval with the following variables. For gations for developing these pilot models are presented in Table 2.
the global model shown in Fig. 3(I) the dominant parameters in
each terms are as follows: 6. Model selection

 Term a: gradation: p3/4, p3/8, p200, volumetrics: Va, Vbeff and VMA. In order to find suitable models, a critical step was to run for-
 Term b: gradation: p200, volumetrics: Va, Vbeff and VMA. ward stepwise regression analysis for the various terms of inde-
pendent variables using different transform functions. Therefore,
For the simplified global model shown in Fig. 3(II) the dominant a number of pilot models were developed and evaluated by consid-
parameters in each terms are as follows: ering the different sets of variables in order to assess the impor-

Vbeff/(Va+Vbeff)
Vbeff/(Va+Vbeff)

Pearson Correlation Pearson Correlation

95 % Confidence Interval 95 % Confidence Interval
90% Confidence Interval 90% Confidence Interval
Vbeff

Vbeff
VMA

VMA
ρ200
ρ200

VFA

VFA
ρ34

ρ38

ρ34

ρ38
Va

Va
ρ4
ρ4

0.5 0.5
Correlation Coefficient

Correlation Coefficient

0.3 0.3

0.1 0.1

-0.1 -0.1

-0.3 -0.3
(a) (b)
-0.5 -0.5
(I)
Correlation Coefficient
Correlation Coefficient

Fig. 3. Correction coefficients versus independent variables: (a) a and (b) b submodels of the (I) global model and (II) simplified global models using NCSU calibration
database.
 M.S. Sakhaeifar et al. / Construction and Building Materials 76 (2015) 221–231 227

Table 2
Regression model coefficients and fitness measurements for pilot closed-form models.

Submodel Variable Global model Simplified global model

1 2 3 4 1 2 3
a Intercept 5.3470 0.75943 6.14890 6.17160 5.91950 6.42230 6.41970
q34 0.00269 0.00263 0.00269 0.00279
(q34)2 0.00015 0.00014
q38 0.0023 0.00127 0.00135 0.00137 0.00564 0.00547 0.00556
(q38)2
q200 0.13259 0.14345 0.09668 0.10641 0.11547 0.11786 0.11787
(q200)2 0.003384 0.00364 0.00103
Va 0.05284 0.05248 0.05705 0.05532 0.05528
(Va)2 0.00331
k1exp(k2/va) k1 5.31540
k2 0.06797
Vbeff 0.05153 0.20936 0.17725 0.17740 0.06244 0.16333 0.16266
(Vbeff)2 0.00780 0.00619 0.00618 0.00491 0.00487
b Intercept 1.7977 0.76002 0.98760 1.0154 1.31340 0.57023 0.57677
q34 0.00021 0.00045
(q34)2
q38 0.00719 0.00697 0.00713
(q38)2
q200 0.15783 0.16022 0.10111 0.08395 0.10857 0.16280 0.16167
(q200)2 0.00861 0.00779 0.00186 0.00531 0.00520
Va 0.01424 0.01420 0.02194 0.01893 0.01889
(Va)2
k1exp(k2/va) k1 1.702
k2 18.621
Vbeff 0.01598 0.21059 0.16876 0.17103 0.03714 0.16122 0.16031
(Vbeff)2 0.00953 0.00745 0.00757 0.00596 0.00592
Sigmoidal const. d 0.86126 0.81552 0.81189 0.81189 1.86802 1.86450 1.86450
g 0.52723 0.54152 0.54698 0.54698 0.96086 0.95991 0.95991
Arithmetic R2 0.95 0.95 0.95 0.95 0.96 0.96 0.96
Se/Sy 0.22 0.22 0.22 0.22 0.21 0.21 0.21
Log R2 0.98 0.98 0.98 0.98 0.98 0.98 0.98
Se/Sy 0.14 0.14 0.14 0.14 0.14 0.14 0.14

Table 3
p-Values from statistical analysis of parameters used in the pilot closed-form models.

Submodel Variable Global model Simplified global model

1 2 3 4 1 2 3
a Intercept 0 0.9941 0 0 0 0 0
q34 <0.00001 <0.00001 <0.00001 0.00069
(q34)2 <0.00001 <0.00001
q38 <0.00001 <0.00001 <0.00001 <0.00001 <0.00001 <0.00001 <0.00001
(q38)2
q200 <0.0001 <0.00001 <0.00001 <0.00001 <0.00001 <0.00001 <0.00001
(q200)2 0.13718 0.10316 0.63553
Va <0.00001 <0.00001 <0.00001 <0.00001 <0.00001
(Va)2 0
k1exp(k2/va) k1 0.95932
k2 0.95925
Vbeff <0.00001 <0.00001 <0.00001 <0.00001 <0.00001 <0.00001 <0.00001
(Vbeff)2 <0.00001 <0.00001 <0.00001 <0.00001 <0.00001
b Intercept <0.00001 <0.00001 <0.00001 <0.00001 <0.00001 0.00012 0.00008
q34 0.86056 0.65649
(q34)2
q38 <0.00001 <0.00001 <0.0001
(q38)2
q200 <0.00001 <0.00001 0.00032 <0.00001 <0.00001 <0.00001 <0.0001
(q200)2 0.00571 0.01150 0.53624 0.00002 0.00002
Va 0.00035 0.00030 <0.00001 <0.00001 <0.00001
(Va)2
k1exp(k2/va) k1 <0.00001
k2 <0.00001
Vbeff <0.00001 <0.00001 <0.00001 <0.00001 <0.00001 <0.00001 <0.0001
(Vbeff)2 <0.00001 <0.00001 <0.00001 <0.00001 <0.00001
Sigmoidal const. d, g 0 0 0 0 0 0 0
228 M.S. Sakhaeifar et al. / Construction and Building Materials 76 (2015) 221–231

8.E+06 1.E+08
Global Model Global Model
LOE LOE
1.E+07
Predicted |E*| (psi)

Predicted |E*| (psi)

6.E+06

1.E+06
4.E+06
1.E+05

2.E+06
1.E+04
(a) R2=0.95, Se/Sy=0.22, n=6336 (b) R2=0.98, Se/Sy=0.14, n=6336
0.E+00 1.E+03
0.E+00 2.E+06 4.E+06 6.E+06 8.E+06 1.E+03 1.E+04 1.E+05 1.E+06 1.E+07 1.E+08
Measured |E*| (psi) (I) Measured |E*| (psi)

8.E+06 1.E+08
Simp. Global Model Simp. Global Model
LOE LOE
Predicted |E*| (psi)

1.E+07

Predicted |E*| (psi)

6.E+06

1.E+06
4.E+06
1.E+05

2.E+06
1.E+04

(a) R2=0.96, Se/Sy=0.21, n=6336 (b) R2=0.98, Se/Sy=0.14, n=6336

0.E+00 1.E+03
0.E+00 2.E+06 4.E+06 6.E+06 8.E+06 1.E+03 1.E+04 1.E+05 1.E+06 1.E+07 1.E+08
Measured |E*| (psi) (II) Measured |E*| (psi)

1.0E+07 1.E+08
Modified Witczak Modified Witczak
LOE LOE
1.E+07
Predicted |E*| (psi)
Predicted |E*| (psi)

7.5E+06

1.E+06
5.0E+06
1.E+05

2.5E+06
1.E+04
(a) R2=0.89, Se/Sy=0.70, n=6336 (b) R2=0.96, Se/Sy=0.23, n=6336
0.0E+00 1.E+03
0.0E+00 2.5E+06 5.0E+06 7.5E+06 1.0E+07 1.E+03 1.E+04 1.E+05 1.E+06 1.E+07 1.E+08
Measured |E*| (psi) (III) Measured |E*| (psi)

8.E+06 1.E+08
Hirsch Model Hirsch Model
LOE LOE
Predicted |E*| (psi)
Predicted |E*| (psi)

1.E+07
6.E+06

1.E+06
4.E+06
1.E+05

2.E+06
1.E+04

(a) R2=0.94, Se/Sy=0.37, n=6336 (b) R2=0.97, Se/Sy=0.18, n=6336

0.E+00 1.E+03
0.E+00 2.E+06 4.E+06 6.E+06 8.E+06 1.E+03 1.E+04 1.E+05 1.E+06 1.E+07 1.E+08
Measured |E*| (psi)
(IV) Measured |E*| (psi)

Fig. 4. Predicted modulus values using (I) global, (II) simplified global, (III) modified Witczak and (IV) Hirsch models for NCSU database (calibration set) in (a) arithmetic and
(b) logarithmic scales.

tance of individual variables as well as the interactions amongst In addition, the statistical analysis was conducted to identify
them. The regression model coefficients and fitness measurements the dominant parameters (p-value < 0.01) for each investigated
for all the pilot models required for selecting the finalized models pilot models. The results of this analysis are summarized Table 3.
are summarized in Table 2. Based on the results of this analysis the fourth and third pilot
 M.S. Sakhaeifar et al. / Construction and Building Materials 76 (2015) 221–231 229

models were selected as the finalized predictive models based
on the first and second proposed modeling approaches
respectively.
6.1. Final selected global model
After careful evaluation of all the candidate models for both the
calibration and verification datasets, the best model based on the
first modeling approach was selected. The selected model is called
Global Model and is presented in Eq. (19).
log jE j ¼ 6:1716  0:00269q34  0:00137q38  0:10641q200
 0:05248V a  0:1774V beff þ 0:00618V 2beff
1:0154 þ 0:08395q200 þ 0:0142V a þ 0:17103V beff  0:00757V 2beff
þ |E⁄| database. The new developed model is called the simplified glo-
1 þ expð0:81189  0:54698X G Þ
ð19Þ
predictions would be unrealistic and the model would be unable
to predict the dynamic modulus. Comparisons of predicted versus
measured dynamic modulus values for the calibration database are
presented in Fig. 4(I). The statistical measurements for the final
global model indicate an excellent fit with high correlation coeffi-
cients and error terms of R2 = 0.95 and Se/Sy = 0.22, and R2 = 0.98
and Se/Sy = 0.14 in arithmetic and logarithmic scales, respectively.
6.2. Final simplified global model
The final predictive model based on the second modeling
approach was selected after a thorough evaluation of all the candi-
date pilot models and was calibrated using 90% of the Master NCSU
bal model and is expressed as follows:

log jE j ¼ 6:4197  0:00014q234  0:00547q38  0:11786q200

 0:05528V a  0:16266V beff þ 0:00487V 2beff
0:57677 þ 0:00713q38 þ 0:16167q200  0:0052q2200 þ 0:01889V a þ 0:16031V beff  0:00592V 2beff
þ  ð21Þ
1 þ e1:86450:95991 log jG j

0 1
The statistical measurements for the final simplified global
B2:4392  100:0004T 2
þ0:0135T0:1003 C
model indicate an excellent fit with high correlation coefficients
B C
X G ¼ log B  0:12332 7:72273 C ð20Þ and error terms of R2 = 0.96 and Se/Sy = 0.21, and R2 = 0.98 and Se/
@ 145;000
A
jG j
1 Sy = 0.14 in arithmetic and logarithmic scales, respectively. The
comparisons of the predicted versus measured dynamic modulus
|E⁄|=dynamic modulus (psi), values for two new models are presented in both arithmetic and log-
|G⁄|b = dynamic shear modulus of binder (psi), and arithmic scales in Fig. 4(I) and (II) and are compared with predictions
T = test temperature (°C). from Modified Witczak and Hirsch models in Fig. 4(III) and (IV),
respectively.
It is important to mention that the database used to calibrate
the global model is processed by fixing an upper limit for the bin- 7. Model performance
der shear modulus that is equal to the glassy modulus, which is
recommended as 1 GPa (145,000 psi) [18]. For this reason, any bin- The final models are applicable to a wide range of mix types and
der shear modulus values used in the global model for predicting test conditions. The performance of these models was evaluated on
the dynamic modulus should be less than 1 GPa; otherwise, the a large and extensive set of verification databases, which includes

Table 4
Description of the developed closed-form models and their validation statistics.

Model Reference scale Statistical parameters for calibration and verification data
90% NCSU 10% NCSU Processed Witczak AMPT
Global modela Arithmetic Se/Sy = 0.22 Se/Sy = 0.27 Se/Sy = 0.51 Se/Sy = 0.71
R2 = 0.95 R2 = 0.93 R2 = 0.75 R2 = 0.92
Log Se/Sy = 0.14 Se/Sy = 0.13 Se/Sy = 0.30 Se/Sy = 0.42
R2 = 0.98 R2 = 0.99 R2 = 0.92 R2 = 0.94
Simplified global modela Arithmetic Se/Sy = 0.21 Se/Sy = 0.29 Se/Sy = 0.52 Se/Sy = 0.68
R2 = 0.96 R2 = 0.91 R2 = 0.76 R2 = 0.92
Log Se/Sy = 0.14 Se/Sy = 0.14 Se/Sy = 0.32 Se/Sy = 0.41
R2 = 0.98 R2 = 0.98 R2 = 0.92 R2 = 0.94
Modified Witczakb Arithmetic Se/Sy = 0.70 Se/Sy = 0.93 Se/Sy = 1.49 Se/Sy = 1.00
R2 = 0.89 R2 = 0.87 R2 = 0.79 R2 = 0.87
Log Se/Sy = 0.23 Se/Sy = 0.30 Se/Sy = 0.41 Se/Sy = 0.49
R2 = 0.96 R2 = 0.96 R2 = 0.92 R2 = 0.93
Hirsch model Arithmetic Se/Sy = 0.37 Se/Sy = 0.47 Se/Sy = 0.50 Se/Sy = 0.29
R2 = 0.94 R2 = 0.97 R2 = 0.79 R2 = 0.92
Log Se/Sy = 0.18 Se/Sy = 0.21 Se/Sy = 0.29 Se/Sy = 0.34
R2 = 0.97 R2 = 0.99 R2 = 0.92 R2 = 0.94
a
90% of NCSU database was used to calibrate this model.
b
The Witczak database was used to calibrate this model.
230 M.S. Sakhaeifar et al. / Construction and Building Materials 76 (2015) 221–231

the test measurements from both the TP-62 test protocol (Witczak datasets; more details are available elsewhere [15]. In order to
and 10% NCSU databases) and the AMPT test protocol, which itself assess the performance of the modified Witczak model fairly, this
includes a large set of data including FHWA Mobile Trailer and WRI model is applied to a large portion of the database that was origi-
databases. Table 4 summarizes the relevant goodness-of-fit statis- nally used to generate the Witczak model (about one-third of the
tics for predictions from the new developed models and compares Witczak database, known here as processed Witczak), as shown in
them with the modified Witczak and Hirsch model for all available Fig. 5(III) and is compared against Hirsch model predictions in

9.E+06 1.E+08
Global Model Global Model
LOE LOE
1.E+07

Predicted |E*| (psi)

Predicted |E*| (psi)

6.E+06
1.E+06

1.E+05
3.E+06

1.E+04

(a) R2=0.75, Se/Sy=0.51, n=3180 (b) R2=0.92, Se/Sy=0.30, n=3180

0.E+00 1.E+03
0.E+00 3.E+06 6.E+06 9.E+06 1.E+03 1.E+04 1.E+05 1.E+06 1.E+07 1.E+08
Measured |E*| (psi) Measured |E*| (psi)
(I)
9.E+06 1.E+08
Simp. Global Model Simp. Global Model
LOE LOE
1.E+07
Predicted |E*| (psi)

Predicted |E*| (psi)

6.E+06
1.E+06

1.E+05
3.E+06
1.E+04

(a) R2=0.76, Se/Sy=0.52, n=3180 (b) R2=0.92, Se/Sy=0.32, n=3180

0.E+00 1.E+03
0.E+00 3.E+06 6.E+06 9.E+06 1.E+03 1.E+04 1.E+05 1.E+06 1.E+07 1.E+08
Measured |E*| (psi) (II) Measured |E*| (psi)

1.2E+07 1.E+08
Modified Witczak Modified Witczak
LOE LOE
1.E+07
Predicted |E*| (psi)
Predicted |E*| (psi)

9.0E+06

1.E+06
6.0E+06
1.E+05

3.0E+06
1.E+04

(a) R2=0.79, Se/Sy=1.49, n=3180 (b) R2=0.92, Se/Sy=0.41, n=3180

0.0E+00 1.E+03
0.0E+00 3.0E+06 6.0E+06 9.0E+06 1.2E+07 1.E+03 1.E+04 1.E+05 1.E+06 1.E+07 1.E+08
Measured |E*| (psi) (III) Measured |E*| (psi)

8.E+06 1.E+08
Hirsch Model Hirsch Model
LOE LOE
1.E+07
Predicted |E*| (psi)
Predicted |E*| (psi)

6.E+06

1.E+06
4.E+06
1.E+05

2.E+06
1.E+04
(a) R2=0.79, Se/Sy=0.50, n=3180 (b) R2=0.92, Se/Sy=0.29, n=3180
0.E+00 1.E+03
0.E+00 2.E+06 4.E+06 6.E+06 8.E+06 1.E+03 1.E+04 1.E+05 1.E+06 1.E+07 1.E+08
Measured |E*| (psi) (IV) Measured |E*| (psi)

Fig. 5. Predicted modulus values using (I) global, (II) simplified global, (III) modified Witczak and (IV) Hirsch models for the processed Witczak database (verification set) in
(a) arithmetic and (b) logarithmic scales.
 M.S. Sakhaeifar et al. / Construction and Building Materials 76 (2015) 221–231
Fig. 5(IV). The evaluations show that the modified Witczak model
tends to have larger scatter and error in its predictions than both
new models. The first observation to be made is that the modified
Witczak model, for all the databases presented in Table 4 and
Fig. 5(III), tends to overestimate the dynamic modulus values over
the entire range, but particularly at the extreme modulus values.
Based on previous investigations [10], it is believed that this effect
is due to the aforementioned use of inappropriate binder shear
modulus values at temperatures equal to or below 4.4 °C (overes-
timation at high modulus values) and the use of dynamic modulus
values at higher than recommended strain levels (overestimation
at low modulus values). The modified Witczak predictions also
tend to show bias at very high temperatures, which is due to the
Witczak model’s insensitivity to volumetric changes and its inabil-
ity to capture these differences clearly [6]. It was observed that the
Hirsch model behaves in a reasonable fashion, although it exhibits
undesirable behavior and a horizontal pattern, amplified in at the
low and high |E⁄| values in the line of equality (LOE) graphs shown
in Figs. 4(IV) and 5(IV). This undesirable pattern in the Hirsch
model prediction is related to the insensitivity of the model and
its inability to distinguish the performance differences among dif-
ferent mixtures for a given set of environmental and other condi-
tions [6]. The developers of the Hirsch model note that during its
initial development, substantial errors occurred when predictions
were made at the extreme high and low modulus values [2]. For
an independent and expanded database as shown in Figs. 4(IV)
and 5(IV), it appears that the Hirsch model developers, with their
available data set, were not able to completely address the original
shortcomings of their model, as plateau appear at the high and low
modulus values. Errors at the very low temperature of 10 °C (high
modulus values) are caused by the model having an apparently
limited modulus of approximately 3,500,00 psi, even though values
higher than this are often measured in the laboratory at 10 °C.
Comparing new models prediction with Hirsch model reveals the
importance of the gradation of the mixture for predicting mechan-
ical behavior of mixture. This observation shows that insensitivity
of the Hirsch model in some loading and frequencies may happen
because lack of gradation parameters in the model. The new pro-
posed models overcome this issue by reducing the bias and the
use of gradation and volumetric parameters and relating them to
mechanical behavior of mixture at different temperatures and
frequencies.
8. Summary and recommendations
Two regression models (global and simplified global) were
developed in this study for estimating the dynamic modulus of
HMA mixtures and were calibrated for a wide range of typical test-
ing conditions, from 10° to 54.4 °C, as recommended in the
AASHTO TP62-03 test protocol. The developed models are new
rational models developed based on viscoelastic principles and
using the t–TS concept. The newly developed dynamic modulus
models were tested for accuracy and rationality, and are found to
estimate the dynamic modulus of HMA mixtures with a very good
level of accuracy and rationality over the full range of potential con-
ditions. The set of models developed in this research study has prac-
tical implications beyond the current study. The most direct use of
these models is the prediction of one of the primary inputs, |E⁄| val-
ues, for MEPDG or other structural/performance analyses of asphalt
concrete pavements. These models are shown to estimate more
accurate and less biased dynamic modulus values than those pre-
dicted by the |E⁄| predictive model used in the current MEPDG
231
and Hirsch model over a wider range of testing conditions which
include critical zones of very low and high temperatures. Therefore,
these models are recommended to be considered for the future
practical implementation of pavement performance analysis.
Acknowledgements
The authors gratefully acknowledge the support from the
Southern Transportation Center, help from Dr. Ramon Bonaquist
at AAT, Dr. Michael Harnsberger at WRI, and Dr. Raj Dongre at
the FHWA-TFHRC in providing data for the database developed in
this study.
References
[1] Advanced Research Associates. 2002 Design guide: design of new and
rehabilitated pavement structures. NCHRP 1-37A project, National
Cooperative Highway Research Program, Washington, D.C.: National
Research Council; 2004.
[2] Christensen Jr DW, Pellinen TK, Bonaquist RF. Hirsch model for estimating the
modulus of asphalt concrete. J Assoc Asphalt Paving Technol (AAPT) 2003;72.
[3] Bari J. Development of a new revised version of the Witczak E⁄ predictive
models for hot mix asphalt mixtures. Ph.D. dissertation, Arizona State
University; 2005.
[4] Al-Khateeb G, Shenoy A, Gibson N, Harman T. A new simplistic model for
dynamic modulus predictions of asphalt paving mixtures. Association of
asphalt paving technologists annual meeting, Paper Preprint CD; 2006.
[5] Sakhaei Far MS. Development of new dynamic modulus (|E⁄|) predictive
models for hot mix asphalt mixtures. Ph.D. dissertation, North Carolina State
University; 2011.
[6] Schwartz CW. Evaluation of the Witczak dynamic modulus prediction model.
In: Proceedings of the 84th annual meeting of the Transportation Research
Board, TRB, Washington, D.C.; 2005.
[7] Ceylan H, Schwartz CW, Kim S, Gopalakrishnan K. Accuracy of predictive
models for dynamic modulus of hot-mix asphalt. J Mater Civ Eng
2009;21(6):286–93.
[8] Yu J, Williams RC, Cascione A. Evaluation of dynamic modulus predictive
models for asphalt mixtures containing recycled asphalt shingles.
Transportation research board 92nd annual meeting; 2013. p. 16.
[9] Ceylan H, Gopalakrishnan K, Kim S. Looking to the future: the next-generation
hot mix asphalt dynamic modulus prediction models. Int J Pavement Eng
2009;10(5):341–52.
[10] Sakhaei Far MS, Underwood BS, Kim YR, Jackson N. The application of artificial
neural networks for estimating dynamic modulus of asphalt concrete. Journal
of transportation research board, No. 2127, Washington, D.C.: TRB, National
Research Council; 2009. p. 173–86.
[11] Sakhaeifar MS, Underwood BS, Kim YR, Puccinelli J, Jackson N. Development of
artificial neural network predictive models for populating the dynamic
modulus of LTPP sections. Journal of transportation research board, No.
2181, Washington, D.C.: National Research Council; 2010. p. 88–97.
[12] Singh D, Zaman M, Commuri S. Inclusion of aggregate angularity, texture, and
form in estimating dynamic modulus of asphalt mixes. Road Mater Pavement
Des 2012;13:327–44.
[13] Dai Q. Prediction of dynamic modulus and phase angle of stone-based
composites using a micromechanical finite-element approach. ASCE J Mater
Civ Eng 2010;22:618–27.
[14] You Z, Buttlar WG. Micromechanical modeling approach to predict
compressive dynamic moduli of asphalt mixtures using the distinct element
method. J Transport Res Rec 2006;1970:73–83.
[15] Kim YR, Underwood BS, Sakhaei Far M, Jackson N, Puccinelli J. LTPP computed
parameter: dynamic modulus. Final report for project: DTFH61-02-D-00139,
Washington, D.C.: Federal Highway Administration; 2009.
[16] Attoh-Okine NO. Predicting roughness progression in flexible pavements using
artificial neural networks. In: Proceedings of the third international conference
managing pavements, vol. 1. San Antonio (TX); 1994. p. 55–62.
[17] Kim YR, Seo Y, King M, Momen M. Dynamic modulus testing of asphalt
concrete in indirect tension mode. Transportation research record: journal of
the transportation research board, No. 1891, Washington, D.C.: TRB, National
Research Council; 2004. p. 163–73.
[18] Zeng M, Bahia HU, Zhai H, Anderson MR, Turner P. Rheological modeling of
modified asphalt binders and mixtures. Asphalt Paving Technol., AAPT
2002;70:403–41.
[19] Underwood BS, Sakhaei Far MS, Kim YR. Using limited purchase specification
tests to perform full linear viscoelastic characterization of asphalt binder.
ASTM J Test Eval 2010;38(5):558–67.
[20] Foxes Team. Optimization, nonlinear fitting and equations solving: tutorial on
numerical analysis with Optimiz.xla, vol. 1. Italy: Foxes Team; 2006.