Download as pdf or txt
Download as pdf or txt
You are on page 1of 18

Construction and Building Materials 354 (2022) 129162

Contents lists available at ScienceDirect

Construction and Building Materials


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

Improving asphalt mix design by predicting alligator cracking and


longitudinal cracking based on machine learning and dimensionality
reduction techniques
Jian Liu a, Fangyu Liu a, Hongren Gong b, Ebenezer O. Fanijo a, Linbing Wang a, c, *
a
Department of Civil and Environmental Engineering, Virginia Polytechnic Institute and State University, Blacksburg, VA 24061, USA
b
Key Laboratory of Road and Traffic Engineering of Ministry of Education, Tongji University, Shanghai, China
c
School of Environmental, Civil, Agricultural and Mechanical Engineering, University of Georgia, Athens, Georgia 30602, USA

A R T I C L E I N F O A B S T R A C T

Keywords: The asphalt mix design based on traditional laboratory fatigue cracking tests of asphalt mixture is not reasonable
Alligator cracking due to the difficulty of simulating the circumstance where asphalt mixture experiences in the pavement structure
Longitudinal cracking and under realistic climate and traffic conditions. The study aims to mitigate the cracking problems during the
Asphalt mix design
pavement design life, by improving the asphalt mix design process by introducing machine learning (ML)
Machine learning
Dimensionality reduction
models, which are used to predict alligator cracking (AC) and longitudinal cracking (LC), and their criteria
MEPDG (required ranges). The data containing AC and LC were extracted from the NCHRP 1-37A report and the long-
term pavement performance (LTPP) program. A total of 33 input features about climate condition, traffic,
pavement structure, and materials properties of each pavement layer were selected. Support Vector Regression
(SVR), Artificial Neural Networks (ANN), Kernel ridge regression (KRR), Gradient boosting (GB), Extra-trees, and
eXtreme Gradient Boosting (XGBoost) were used. Meanwhile, three dimensionality reduction approaches,
including Auto-Encoder (AE), Principal Component Analysis (PCA), and Recursive feature elimination with
Random Forest, were combined with the six ML algorithms (18 hybrid models produced) in order to decrease the
computation complexity and search for the optimum model. The 24 models (6 basic ML models plus 18 hybrid
models) were trained, and their hyperparameters were tuned using Bayesian optimization. Different evaluation
criteria R2, RMSE, MAE, SMAPE, and SI are calculated to evaluate the performance of these models. The results of
the study showed that except for ANN and SVR, the basic ML models can get better for predicting AC and LC
when they are combined with PCA or AE. Based on the performance indices, the optimum among these devel­
oped models proved to be PCA-ANN for predicting AC (R2 = 0.84) and LC (R2 = 0.83). Compared to other
pavement layers, the asphalt surface course is the highest contributor to AC and LC. The improved mix design
process was implemented to determine the mix proportion for the asphalt surface course of a construction project
and Mix-2 with 5.1 % asphalt content was recommended.

1. Introduction Mechanistic-Empirical Pavement Design Guide (MEPDG), AC and LC are


load-related cracking [1] which results from asphalt concrete fatigue
Pavement cracking is one of the main distresses of asphalt pavements under repetitive loads and environmental effects [2]. In addition to
and is considered the most critical indicator of asphalt pavement con­ traffic and environmental factors, many studies have demonstrated that
dition. Specifically, alligator cracking (AC) and longitudinal cracking properties of asphalt mixture have important effects on pavement
(LC) are the primary forms of asphalt pavement cracking. In addition, cracking [3,6], such as fatigue life [7] and stiffness [4]. Superior prop­
AC usually initiates at the bottom of the asphalt concrete layer and then erties of asphalt mixture are not only affected by the quality of aggre­
progresses up to the pavement surface. Top-down cracking, which is one gates and binders but also rely on scientific asphalt mix design
of the main types of LC, appears first at the pavement surface. In the approaches. Therefore, a reasonable asphalt mixture design can mitigate

* Corresponding author.
E-mail addresses: jian19@vt.edu (J. Liu), lfangyu@vt.edu (F. Liu), hgong@tongji.edu.cn (H. Gong), ebenfanijo@vt.edu (E.O. Fanijo), Linbing.Wang@uga.edu
(L. Wang).

https://doi.org/10.1016/j.conbuildmat.2022.129162
Received 3 June 2022; Received in revised form 29 August 2022; Accepted 11 September 2022
Available online 19 September 2022
0950-0618/© 2022 Elsevier Ltd. All rights reserved.
J. Liu et al. Construction and Building Materials 354 (2022) 129162

Table 1
Summary of studies on asphalt pavement performance prediction based on ML techniques.
Pavement Data size (source) Best model architecture No. of Goodness of fit statistics Studies
performance inputs

Rutting 440 (NCHRP Project 01-37A) Artificial neural networks (ANN) 20 (Coefficient of determination) Gong et al. [25]
(20–200-50–10-1) R2 = 0.87
6946 (LTPP) ANN (29,128,32,8,1) 29 R2 = 0.82 Haddad et al.[26]
5265 (LTPP) ANN (19, 50, 20, 10) 19 R2 = 0.93 Wang et al. [27]
IRI 1781 (LTPP) Gradient boosting 13 R2 = 0.90 Guo et al. [33]
1880 (Korea Highway data) Recurrent neural networks 9 R2 = 0.87 Choi and Do [34]
12,300 (LTPP) Random forest 19 R2 = 0.97 Gong et al.[35]
2439 (LTPP) ANN (5–10–10–10-1) 10 R2 = 0.75 Abdelaziz et al.
[36]
125 (Taiwan Highway data) ANN —— R2 = 0.95 Lin et al. [37]
Fatigue cracking 461,414 (NCHRP Project 01-37A) eXtreme Gradient Boosting 15 R2 = 0.58, 0.78 Gong et al. [30]
67 (LTPP) Random forest-ANN (12,10,4,1) 11 R2 = 0.91 Fathi et al. [28]
509 (NCHRP 01-37A) ANN (14,200,100,20,20,1) 14 R2 = 0.61 Gong et al. [29]
Dynamic module 4650 (Arizona State University Deep residual ANNs 13 R2 = 0.99 Moussa and Owais
(ASU) database) [41]
198 (Georgia HMA mixtures) Gradient boosting 19 R2 = 0.98 Worthey et al. [42]
1152 (Laboratory results) XGBoost 24 R2 = 0.96 Yasir et al. [43]
Cracking, rutting, 404, 460, 443 (LTPP) Random forest 17 —— Gong et al. [44]
IRI

cracking problems and extend pavement life. Carolina and the performances of the AC prediction model are R2 of 0.22
Traditional mix design approaches only consider volumetric and and 0.099, respectively. Local calibration [15] for Arizona shows R2 of
mechanical properties of asphalt mixture, which may not cover 0.226 and − 10.2 for AC and LC prediction, respectively. The predictions
comprehensive characteristics, for example, the fatigue resistance of of calibrated transfer functions seem to be biased. In addition to the
asphalt mixture. Therefore, researchers have introduced laboratory tests deterministic models of regression and calibrated MEPDG models,
(e.g., repeated flexural bending [5]), the corresponding evaluation pa­ another aspect is to use traditional probabilistic theory, such as Markov
rameters (e.g., fatigue life [5]), and the related requirements [6,7]. In process-based transition probability matrices (TPMs) [16], Miner’s law
addition, the common approaches to capture the fatigue properties also [17], Bayesian analysis [18], and survival model [19] to predict
include other laboratory tests, such as overlay tester [8] and cantilever cracking progression [20]. Besides, finite element methods also can be
bending tests [9], and fatigue life prediction models, such as exponen­ applied to simulate the cracking response of asphalt pavement [21–24],
tial, Weibull [10], and logit models [2]. However, laboratory fatigue Ambassa et al. [22] predicted fatigue lives of pavements by establishing
cracking tests are implemented on a specimen with a defined thickness viscoelastic finite element method (FEM) models considering load types,
and under assigned stress or strain, which cannot reflect the real speed, and asphalt mixture temperature. Kim [24] et al. and Norouzi
circumstance where asphalt mixture experiences because of the exis­ et al. [23] combined the Viscoelastic Continuum Damage model and
tence of various uncertain factors (structural parameters, environment FEM to simulate fatigue damage areas of asphalt pavement, further
effects, traffic loading, etc.). Although the Accelerated Loading Facility predicting cracking areas. However, complex climate conditions, which
(ALF) [11] accelerated loading test with full-scale construction can affect fatigue performance significantly, are hard to be simulated by
closely simulate the real case, the whole process costs much. Fatigue FEM.
cracking can be still estimated using fatigue cracking predictive models Recently, using machine learning (ML) and deep learning in pre­
with the inputs which contain asphalt concrete materials properties dicting asphalt pavement performance, such as rutting [25–27], fatigue
involved in the mix design process. The relationship between fatigue cracking [28–32], and international roughness index (IRI) [33–37], is a
cracking and asphalt mix proportion can be captured if the accurate hot topic. Table 1 summarized several past research on asphalt pave­
predictive model can be established. Serious cracking (AC and LC) ment performance prediction using ML techniques. Gong et al. [30]
means asphalt concrete layers have lost the ability to support repeated developed gradient boosted models to predict the AC and LC with the
loading. The criteria used to judge whether AC and LC of a pavement are inputs of climate variables, traffic variables (AADT), pavement thick­
so severe that the pavement fails can be adopted for controlling the fa­ ness, and asphalt mixture properties. However, the inputs related to
tigue resistance ability of asphalt mixture during asphalt mix design properties of asphalt mixture are limited and do not cover the adequate
phase. In this study, the inputs for AC and LC predictive models involve insights of asphalt concrete. Fathi et al. [38] also applied a hybrid ML
asphalt concrete characteristics of different pavement layer, but only the model with Random Forest and ANN to predict alligator deterioration
mix design process for the asphalt surface course of new flexible pave­ index using the LTPP dataset, and the inputs include deterioration index,
ments was analyzed. volumetric properties of asphalt mixture, and age. The same work was
At present, there already have been many studies focusing on also done by Inkoom et al. [39,40]. However, in these studies, the inputs
developing models to predict the cracking of asphalt pavement. One for alligator cracking models did not simultaneously contain the inte­
area is to fit a traditional linear or nonlinear regression predictive model grated information including climate, traffic, structures, and materials
[12]. Ker et al. [12] established a generalized linear model with the of each layer, especially for the variables involved in the mix design
assumption of Poisson distribution and the inputs of the KESALs, process, and the types of ML models used were limited. Therefore, this
pavement age, climate variables, and tensile strain under the surface study aims to establish ML models for predicting AC and LC to improve
layer, and the R2 for the model was 0.33 [12]. In the MEPDG, there are the asphalt mix design, and the inputs will contain as many variables
mechanistic-empirical models, called transfer functions, to predict AC involved in the mix design process as possible. In addition, the infor­
and LC progressions, but the transfer functions need to be calibrated mation on each pavement layer will be considered.
with nationwide or local data. The global calibrated transfer functions
for predicting AC and LC were verified on 405 and 312 observed data
1.1. Objective
with R2 of 0.275 and 0.544, respectively [1]. Hall et al. [13], Jadoun and
Kim [14] conducted calibration of MEPDG in Arkansas and North
Properties of asphalt mixture have important effects on pavement

2
J. Liu et al. Construction and Building Materials 354 (2022) 129162

Fig. 1. Typical pavement structure analyzed in this study.

Table 2
Summary statistics of input and output variables of ML models.
Type of variables Variables indicators Min. Max. Avg. SD.

Input —— Age (month) Age 1.0 321.0 127.0 70.87


Climate Mean ESALs (k/year) KESAL 2.9 23.9 14.9 5.62
Average annual freeze index Freeze_index 0.0 1087.8 203.0 279.21
Average monthly temperature difference (℃)/ year Temp_max_diff 10.3 35.2 23.1 4.76
Average annual precipitation (cm) /year Ann_precip 7.6 1708.2 820.3 249.78
Asphalt Surface Course Thickness (inch) Surface_AC_h 0.9 10.1 3.1 2.11
Specific gravity of binder Surface_AC Gb 0.985 1.045 1.023 0.012
Binder penetration at 77 ◦ F (0.1 mm) Surface_AC pen_77 ◦ F 33.0 225.0 87.2 31.65
Percent passing sieve size (%) 19 mm Surface_AC P_19 mm 81.0 100 98.6 3.47
9.5 mm Surface_AC P_9.5 mm 57.0 100.0 83.4 10.79
4.75 mm Surface_AC P_4.75 mm 41.5 80.0 60.1 7.34
0.075 mm Surface_AC P 0.075 mm 0.1 50.0 6.18 4.83
Bulk specific gravity of aggregate Surface_AC Gsb 2.376 2.972 2.606 0.102
Asphalt content (%) Surface_AC Pb 2.5 7.0 5.3 0.73
Bulk specific gravity of asphalt mixture Surface_AC Gmb 1.914 2.583 2.263 0.105
Air void (%) Surface_AC VV 2.4 13.5 6.8 3.20
Asphalt Binder Course Thickness (inch) BC_h 1.4 14.5 5.3 2.65
Percent passing sieve size (%) 19 mm BC P_19 mm 65.0 100.0 87.5 10.72
9.5 mm BC P_9.5 mm 36.0 93.0 66.7 14.3
4.75 mm BC P_4.75 mm 23.0 72.5 49.8 12.05
0.075 mm BC P_0.075 mm 2.0 10.1 5.3 2.25
Binder penetration at 77 ◦ F (0.1 mm) BC pen_77 ◦ F 41.7 132.0 82.1 19.41
Asphalt content (%) BC Pb 3.4 6.3 4.8 0.75
Air void (%) BC VV 4.6 13.5 8.5 2.21
Base layer Asphalt Treated Base Thickness (inch) ATB_h 3.0 12.7 7.8 2.87
Granular Base Thickness (inch) GB_h 2.9 25.8 10.8 5.76
Granular Base Mr (MPa) GB Mr 55.1 267.5 166.9 42.92
Subbase layer Thickness SB_h 3.6 23.5 9.8 5.39
Materials type SB Mater_type —— —— —— ——
Mr (MPa) SB Mr 78.5 192.1 122.8 34.84
Subgrade Materials type SG Mater_type —— —— —— ——
Mr (MPa) SG Mr 34.1 182.2 75.1 21.98
Output —— Alligator cracking (%) AC 0 84.1 7.2 12.56
Longitudinal cracking (ft./mi) LC 13.5 8007 1131 1484.7

cracking, and AC and LC are the main types of cracking. Scientific 2. Data preparation
asphalt mix design can mitigate the cracking of pavement. The objective
of this study is to improve the asphalt mix design by introducing ML 2.1. Collection of dataset
models to predict AC and LC and their corresponding criteria. The whole
dataset was collected from the report NCHRP 1-37A and the LTPP The MEPDG introduces a state-of-the-practice tool to design and
dataset. six ML models and 18 hybrid models made by the combination analyze new and rehabilitated pavement structures. The NCHRP 1-37A
of ML algorithm with three dimensionality reduction techniques would report [45] illustrates the specific guides for the Mechanistic-Empirical
be optimized (hyperparameter tunning) and trained with the input design of new and rehabilitated pavement. In the NCHRP 1-37A report,
variables relating to climate, traffic, pavement structure, and pavement the distress transfer functions were calibrated using data from the LTPP
materials. The prediction performance of the best model was compared database over representative pavement sections throughout the entire
with that of global calibrated transfer functions. Finally, the mix pro­ U.S. The LTPP database records the information on the environment,
portion of the asphalt surface course of a construction project was traffic, inventory, and materials in more than 2500 sections. To conve­
determined based on the improved asphalt mix design approach. niently compare the performance of ML models and the calibrated
transfer functions for predicting AC and LC, the data which were used to
train and test the ML models were the data for the calibration of the
design of new flexible pavement. Meanwhile, the observed pavement

3
J. Liu et al. Construction and Building Materials 354 (2022) 129162

Table 3 ATB layers have little difference. Therefore, the three features (the type
The number of samples in each data analysis step. of materials for ATB and GB, the Mr for GB) were deleted. Finally, the 33
AC LC input parameters were selected and their names are displayed in Table 2.
If an item lacks one layer, zeros were assigned to the attributes of this
Original data sample from NCHRP 1-37A 634 532
Data samples after removing abnormal zero targets 579 474 layer. The missing Mr of subgrade and subbase/base layers can be filled
Number of training data samples 463 379 with the sections that are the same type of layers and have the same
Number of testing data samples 116 95 materials. Missing values on material characteristics of the surface
Number of data samples for verification of transfer functions 513 381 asphalt concrete layer and binder course, such as VV and Gmb of asphalt
mixture, were inferred by the same feature from the nearest sections or
distress data of pavement sections from various states are beneficial to calculated by other features. In addition, categorical variables, which
achieving better generalization abilities of ML models. The data for are materials type of subgrade and subbase layers, were encoded with
globally calibrating AC and LC prediction models can be extracted from the Ordinal encode method. This process transforms each feature of
Appendix-EE-1 of the NCHRP 1-37A report [45]. Additionally, the categorical variables into a unique integer number.
length and the area of each section in the LTPP are 152 m (500 feet) and Finally, a total of 579 items (rows) from 82 sections for AC were
557.4 m2 (6,000 ft2), and the units of each observed AC and LC data reserved. As regards the LC, 83 sections with 474 items (rows) were
were transformed to the percent of lane area and ft./mi, respectively. kept. The summary statistics of these input variables of ML models and
Since the factors referring to traffic conditions, climate conditions, two targets are presented in Table 2. Table 3 shows the number of
pavement materials properties, and pavement structures are supposed to samples in each data analysis step.
considered for ML prediction models, these attributes of each item (row) Fig. 2 shows the AC or LC versus age and their histogram distribu­
were extracted from the LTPP dataset. Additionally, the data for veri­ tions for the database. As can be seen in Fig. 2, the histograms of AC and
fying global calibrated distress transfer functions were extracted from LC suggest that most AC samples are within 0–5 %, and the majority of
Appendix-II-1 of NCHRP 1-37A. LC samples are in the range of 0 to 500 ft./mi. The MEPDG [1] recom­
mends three performance thresholds (Interstate, primary, and other) for
pavement design. The criteria are used to judge [46] whether a pave­
2.2. Data pre-processing and feature selection ment fails at the end of the design life in this study. Asphalt pavements
are generally designed for 20 years. The AC criteria [47] for Primary
In general, before fed into ML models, the original dataset should be roads and Interstate roads are 10 % and 20 %, respectively. The longi­
processed to deal with outliers, missing values, and categorical vari­ tudinal cracking for both Interstate and Primary roads is 2000 ft./mi
ables. Although the original dataset is extracted from the NCHRP 1-37A [30]. The dotted lines in Fig. 2 represent these criteria. According to
report, outliers for AC and LC also exist. The original data were intended Fig. 2, all the failure points (sections) of both AC and LC which are above
for calibration of the transfer functions, so there should be no inter­ the criteria before 20 years nearly belong to Primary roads. For the AC
ruption of maintenance or rehabilitation on the corresponding sections. dataset, at least 15.8 % of sections fail in their design service life, and for
However, as regards the change of AC or LC over time in one section, the LC dataset, at least 20.4 %. According to the design principle of
some observed AC or LC in the same section suddenly become zero MEPDG, these sections’ actual performances fail to satisfy the require­
although the last observed AC or LC in this section are large enough, ment of reliability when the level of the design reliability is 90 %.
which is unrealistic. Therefore, the items (rows) corresponding to these
observed AC or LC with abnormal zero values were removed directly. 3. Methodology
Fig. 1 displays the typical pavement structure of the sections from the
studied dataset. The purpose of this study is to improve asphalt mix 3.1. Machine learning
design by predicting AC and LC. More characteristics of surface asphalt
concrete (AC) layers were considered. Additionally, each base or sub­ Machine learning (ML) is a subfield of computational intelligence
base layer is supposed to have three attributes: materials type, thickness, which concentrates on the use of data and algorithm to teach computers
and resilient modulus (Mr). However, the materials of asphalt treated and make predictions. Machine learning has widely been used in various
base (ATB) are all asphalt treated mixture, and the materials of granular fields, such as marketing analytics [48], medical diagnosis [49], self-
base (GB) layers are all granular types of materials. Besides, all the Mr of

Fig. 2. Scatter plots and histograms of alligator cracking or longitudinal cracking and age for the database.

4
J. Liu et al. Construction and Building Materials 354 (2022) 129162

driving cars [50], civil engineering [51–54], etc. over the past years. The √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
√∑
ML models for establishing prediction models in this study include √n
√ ( ŷi − yi )2
Support Vector Regression (SVR), Artificial Neural Networks (ANN), √
RMSE = i=1 (2)
Kernel ridge regression (KRR), Gradient boosting (GB), Extra-trees, and n
eXtreme Gradient Boosting (XGBoost).
Support vector machine (SVM) was first put forth by Vapnik [55]. 1∑ n
MAE = |yi − ŷi | (3)
The SVR incorporates the principle of SVM to achieve the regression n i=1
objective. SVR can fit a continuous-valued function by minimizing the
width of the region in which most training samples lie. Additionally, 100% ∑
n
|yi − ŷi |
SVR allows us to set the tolerance of error (an error margin and tolerance SMAPE = (4)
n i=1 (yi + ŷi )/2
of falling outside the region). In general, to solve the nonlinear prob­
lems, the input features are usually transformed into higher dimensional √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
√∑
√ n [(y − y) − ( ŷ − ̂
features to obtain better fitting results. ANN originates from biological √ i i y ) ]2
√i=1
systems and is used to capture more complex patterns from data. An SI = √
∑ n (5)

ANN model generally has three parts: an input layer, hidden layers, and yi 2
an output layer. Specifically, the network of an ANN model consists of
i=1

connections, each connection transfers the output neuron of this layer to


Where n is the number of samples in a dataset; yi is a measured target; ŷi
the input neuron of the next layer and is assigned a weight at the same
is a predicted target; y is the mean value of all the measured targets in a
time. The activation function determines what the output of a node
should be when given inputs. An ANN model is trained by minimizing dataset; ̂
y is the mean value of all the predicted targets in a dataset. if the
the error between actual targets and predicted ones. KRR just combines R2 is larger or the RMSE, MAE, MAPE, and SI are smaller, the model
the ridge regression (least-squares with l2-norm regularization) and performs well. In this study, the mean MSE is used to evaluate the per­
kernel trick. The form of the KRR is similar to SVR: the model is fitted as formance of an ML model with a combination of hyperparameters dur­
a linear function in the high dimensional space transformed by a kernel ing the Bayesian optimization process. The final test prediction accuracy
and the low dimensional data. In general, regarding nonlinear problems, of each ML model would be captured by all the aforementioned per­
training a KRR model is faster than other ML models. formance metrics.
Ensemble methods are the techniques that create multiple learning
algorithms and then combine them to provide better predictive perfor­ 3.3. Bayesian optimization
mance. Bagging and boosting are the main classes of ensemble learning
methods. Bagging involves establishing many decision trees on different The traditional methods for hyperparameter tunning include grid
samples which are extracted randomly from the same dataset and then search and random search. The grid search algorithm exhaustively trains
averaging the total predictions. Both Extra-trees and Random Forest a model with each possible combination of hyperparameters of a given
(RF) follow the bagging method but with fewer variables trained for space, which cost much time. In a random search, a model is trained
each tree. On the other hand, the cut points for splitting samples of one with the limited number of hyperparameter combinations from the
node in the Extra-trees [56] are selected by random instead of optimal randomized search on a given distribution of each hyperparameter.
selection for RF. Boosting involves adding basic estimators sequentially However, random search is appropriate when they are a smaller number
that make up for the prediction error made by prior estimators and of hyperparameters. Bayesian optimization is a sequential model-based
outputting a weights average of the results of all the learners. Both GB optimization that obtains information from the previous results to up­
and XGBoost are the typical Boosting methods. A GB model is trained by date the optimization function [72].
optimizing the arbitrary differentiable loss functions. Specifically, each In Bayesian optimization, a function is evaluated using a surrogate
tree in a GB model is fit by calculating the negative gradient of the given model (usually a Gaussian process). In a Gaussian process model, the 5-
loss function. Compared to GB, XGBoost adds advanced regularization fold cross-validated (CV) ML performance is regarded as the dependent
(L1 & L2) by adjusting the minimum loss reduction for splitting samples, variable (function f(x)), and the variable x is a combination of hyper­
which can decrease the possibility of getting overfitting. In general, parameters. The ML performance functions can be thought to be a
XGBoost can achieve higher performance than GB. Gaussian process and follow an N-dimensional multivariate normal
These ML models mentioned above have been used to predict distribution.
properties of asphalt mixture and asphalt pavement, such as asphalt
content [53], dynamic modulus [57–64], indirect tensile strength (6)

f (x)∽GP(m(x), k(x, x ))
[65–68], rut depth [35,44], and international roughness index (IRI)
The mean function m(x) is usually zero. The covariance function
[69–71].
k(x, x ), which determines a covariance matrix, is given by a kernel (k)

between the hyperparameters, the Matern kernel was used in this study.
3.2. Evaluation metrics
First, some f(x) points (random generative hyperparameters resulted
in MSE) are sampled, and a gaussian process is fitted based on these
To compare the performance of different predictive models, the
points. Then the prediction model for MSE is established as a function of
metrics to evaluate the performance of prediction models should be
hyperparameters. The next combination of hyperparameters used to
introduced. Five statistical metrics including R2 (coefficient of deter­
establish the model is found based on the acquisition function. The
mination), RMSE (root mean square error), MAE (mean absolute error),
acquisition function is a function that can be evaluated at x and then
SMAPE (symmetric mean absolute percentage), and SI (scatter index)
achieves the balance between exploitation and exploration. The ex­
are defined as follows:
pected improvement, which was used as the acquisition function of the
∑n
(yi − ŷi )2 study, is defined as follows:
R2 = 1 − ∑in 2
(1)
i (yi − y) EI(x)= E[max(0, f (x) − f (x* )] (7)

where f(x* ) is the best model performance (MSE) with a combination of


hyperparameters x which has been achieved so far. The trained Gaussian
process can predict the probability of the model’s performance

5
J. Liu et al. Construction and Building Materials 354 (2022) 129162

Fig. 3. Schematic of autoencoder model used in this study.

Fig. 4. Mean MSE (AE performance) vs the number of iterations during the Bayesian optimization process.

(P(f (x)|x )) when given the input x, and (f(x)|x ) follows a normal dis­ model performance f(x* ). The next hyperparameters xt+1 are the values
tribution. Therefore, the equation to explain the expected improvement which result in the best-expected improvement [73]:
can be rewritten below:
∫∞ xt+1 = argmaxEI(xt+1 ) (10)
xt+1
EI(x) = P(f (x)|x )(f (x) − f (x* ) )df (x)
f (x* ) Re-built the Gaussian process model based on the previous points
∫ ∞
and the new sample and search the next sample again. Repeat the pro­
= *
N(f (x)|μ(x), σ (x) )(f (x) − f (x ) )df (x) (8)
f (x* ) cess until the number of samples meets the set requirement. Finally,
search the history of sampled points to find the point with the highest
where μ(x), σ (x) is the mean and standard deviation of the normal dis­ model performance. In this study, the hyperparameters of ML and
tribution for (f(x)|x ), and they can be evaluated by the Gaussian process. autoencoder models would be tuned using Bayesian optimization.
Eq. (8) can transform into an analytical form through applying tedious
integration by parts [73] as follows:
3.4. Dimensions reduction methods
EI(x) = (μ(x) − f (x* ) )φ(z) + σ(x)ϕ(z) (9)
Dimensionality reduction approaches are to transform features into a
where φ(z) is the cumulative distribution function of the normal dis­ low dimension, and are classified into two types: linear and nonlinear
tribution N(0, 1), ϕ(z) is the probability density function of the normal approaches. Dimensionality reduction methods can not only decrease
distribution N(0, 1), z is the scaled difference between the μ(x) and best the complexity of ML models but also reduce the noise of the data, which
can improve the performance of ML models. However, using fewer

6
J. Liu et al. Construction and Building Materials 354 (2022) 129162

Fig. 5. The cumulative variance of different numbers of principal components.

Fig. 6. The mean MSE of CV results vs the number of iterations (different combinations of hyperparameters) when applying Bayesian optimization on ML models.

7
J. Liu et al. Construction and Building Materials 354 (2022) 129162

Fig. 7. The mean MSE of CV results vs the number of iterations (different combinations of hyperparameters) when applying Bayesian optimization on hybrid models
made by each ML model with AE.

method and a neural network-based model. The autoencoder is to


Table 4 explore the relationship between actual data and the data in the desired
Time taken for hyperparameter tunning of ML models using Bayesian
dimension. The architecture of an autoencoder model consists of three
optimization.
parts: encoder, coder (Bottleneck), and decoder. In this study, a two-
ML Basic ML Combined with Combined with Combined layer autoencoder was used. Fig. 3 shows the architecture of the
models model RFE PCA with AE
autoencoder model used in this study. As can be seen in Fig. 3, the
Time taken for hyperparameter tunning (s) neurons of the Bottleneck layer are the new features after the dimen­
SVR 185.5 164.7 173.53 123.3 sionality of reduction. The loss function for the autoencoder model is to
KRR 55.9 33.0 45.2 46.2 calculate the error between the reconstruction of the inputs and the
ANN 1575.15 422.1 674.3 783.4 actual inputs.
GB 331.0 237.5 301.8 246.5
How does it find the best neurons of the hidden layer and the
Extra- 218 178.1 192.3 195.4
trees bottleneck layer? Autoencoder is one of the ANN models, so the number
XGBoost 252 115.2 117.8 178.2 of neurons in the hidden layer is also a hyperparametes. In this study, the
neurons of the hidden layers, bottleneck layers, and learning rate (LR)
are tuned using Bayesian optimization. The search bounds for the neu­
features means that only the part of the information from the original rons of the two layers and LR are (6, 16), (17, 28), and (10− 6, 10− 1).
dataset will be fed into models, which may reduce the accuracy of Fig. 4 shows the change of mean MSE (CV results) over the number of
models. calls to each autoencoder model of AC and LC during the Bayesian
This study explores the combination of dimensionality reduction optimization process. The optimal combination of hyperparameters
methods and ML models to establish a better prediction model. The happened in the minimum value of mean MSE. As a result, the optimal
name of a hybrid model produced by the combination of ML and networks for the AC and LC dataset are the same: 28 neurons in the
dimensionality reduction methods is abbreviated to (dimensionality hidden layer, and 16 neurons in the bottleneck layer. However, the LR
reduction method)-(ML). for training the AC and the LC datasets are 0.018 and 0.020, respec­
tively. The other hyperparameters were determined based on
3.4.1. Autoencoder experience.
Autoencoder (AE) [74] is a nonlinear dimensionality reduction

8
J. Liu et al. Construction and Building Materials 354 (2022) 129162

Table 5
The optimal hyperparameters of ML models and hybrid models based on Bayesian optimization.
ML models Name of Hyperparameters Basic ML model Combined with RFE Combined with PCA Combined with AE

Values MSE Values MSE Values MSE Values MSE

SVR C 0.08 33.20 11.92 34.60 1000 27.15 999.9 36.18


Gamma 1.531 3.406 4.508 0.010
Kernel function Poly Poly RBF RBF
Degree 5 5 3 5
KRR Alpha 0.002 28.03 0.117 36.39 0.018 26.75 0.001 27.66
Gamma 1.187 0.866 10.00 10.00
Kernel function RBF Poly RBF RBF
Degree 3 4 5 1
ANN LR 0.0003 31.26 0.001 55.24 0.0118 25.01 0.005 50.88
Batch size 34 16 256 16
Hidden layers 5 3 3 7
Neurons of each hidden layer 512 257 512 512
GB N_estimators 1000 27.46 932 28.22 977 25.55 1000 24.12
Max_depth 39 14 25 41
Min_samples_split 2 2 3 2
Min_samples_leaf 21 19 44 34
Extra-trees N_estimators 1000 27.82 1000 28.65 1000 23.55 682 24.62
Max_depth 8 9 45 32
Min_samples_split 2 2 2 2
Min_samples_leaf 1 1 1 1
XGBoost N_estimators 1000 28.94 538 29.75 1000 25.50 974 23.74
Max_depth 25 15 25 25
Gamma 1.000 0.100 0.100 0.810
Subsample 0.10 0.10 1.00 0.29
Colsample_bytree 0.90 0.90 0.50 0.51
Min_child_weight 5 4 32 5

Table 6
Statistical metrics of ML and hybrid models made by ML models and hybrid models with dimensionality reduction methods on training and testing data samples for AC.
Method Datasets Metrics SVR KRR ANN GB Extra-trees XGBoost

Basic ML model Training R2 0.8709 0.9356 0.9274 0.9857 0.9810 0.9853


Testing R2 0.6478 0.7975 0.7904 0.7465 0.7519 0.7733
MSE 41.97 24.13 24.98 30.21 29.56 29.39
MAE 2.32 1.95 1.78 2.43 2.15 2.35
SMAPE 149.04 142.47 151.17 148.90 144.63 150.67
SI 56.19 % 42.54 % 43.23 % 47.67 % 47.17 % 47.02 %
Combined with RFE Training R2 0.8339 0.8746 0.8243 0.9737 0.9689 0.9809
Testing R2 0.6134 0.7040 0.5266 0.7180 0.7410 0.7519
MSE 46.07 35.27 56.41 28.84 25.85 27.17
MAE 3.09 2.79 3.04 2.35 1.98 2.00
SMAPE 152.05 151.38 153.87 152.11 143.42 142.30
SI 58.87 % 51.43 % 64.56 % 46.58 % 44.08 % 45.21 %
Combined with PCA Training R2 0.9354 0.9558 0.9243 0.9883 0.9994 0.9953
Testing R2 0.7787 0.8344 0.8400 0.7635 0.8210 0.7927
MSE 26.48 19.73 19.06 28.18 21.33 24.70
MAE 2.02 1.80 1.74 2.4091 1.62 2.16
SMAPE 148.01 148.14 148.06 148.93 125.22 143.13
SI 44.63 % 38.52 % 37.69 % 46.04 % 40.04 % 43.11 %
Combined with AE Training R2 0.8683 0.9375 0.7875 0.9927 0.9994 0.9978
Testing R2 0.7105 0.8326 0.6002 0.8214 0.8109 0.7810
MSE 34.49 19.95 47.65 21.27 22.53 26.09
MAE 2.52 1.98 2.77 1.73 1.64 1.92
SMAPE 150.49 148.80 156.55 142.63 48.97 145.93
SI 50.93 % 38.67 % 59.86 % 40.01 % 41.14 % 44.30 %

3.4.2. Principal component analysis (PCA) predicting AC and LC, respectively. Fig. 5 represents how cumulative
PCA uses an orthogonal transformation to linearly transform original variance changes with the increase in the number of components. Ac­
features into a new set of linearly unrelated variables. PCA is to handle cording to Fig. 5, the first component can explain the largest variance,
Eigen-decomposition on the covariance matrix of original features. The and the last 10 components can only explain a little variance. In this
eigenvectors of the covariance matrix are Principal Components and study, the selected components were supposed to explain at least 95 %
describe the direction of the new feature space. The eigenvalues explain variance of the original dataset. A threshold of 95 % was used to select
the variance of the original data along the axes of the new coordination. the number of suitable PCAs. Fig. 5. shows that the cumulative variances
Therefore, the new set of variables is linearly unrelated. The explained for the first 15 and 16 components are 0.951 and 0.956. Since the fea­
variance ratio of a principal component is the variance of the principal tures after transformation for autoencoder models are 16, the first 16
component divided by the total variance. The cumulative explained PCAs fed into ML models were selected.
variance ratio refers to the sum of explained variance ratios of given
principal components. PCA was implemented on the input features for

9
J. Liu et al. Construction and Building Materials 354 (2022) 129162

Table 7
Statistical metrics of ML and hybrid models made by ML model with between observed and predicted data for LC.
Method Dataset Metrics SVR KRR ANN GB Extra-trees XGBoost
2
Basic ML model Training R 0.8556 0.9055 0.8791 0.9245 0.9153 0.9873
Testing R2 0.7162 0.7597 0.7334 0.7429 0.7775 0.7637
MSE 507,960 430,095 477,244 460,123 398,323 448,044
MAE 364.57 328.05 330.91 412.70 308.12 359.83
SMAPE 113.27 106.14 106.56 103.87 105.15 112.52
SI 46.92 % 43.18 % 45.47 % 42.90 % 41.54 % 43.94 %
Combined with RFE Training R2 0.8372 0.9124 0.8418 0.9273 0.9643 0.9827
Testing R2 0.6924 0.7200 0.6120 0.7597 0.7737 0.7518
MSE 550,661 501,275 694,510 430,122 387,247 431,815
MAE 337.91 385.00 441.09 380.87 306.43 367.94
SMAPE 115.92 114.34 121.34 123.76 100.21 121.05
SI 48.85 % 46.33 % 54.83 % 43.20 % 40.99 % 43.21 %
Combined with PCA Training R2 0.8383 0.9102 0.9340 0.9951 0.9992 0.9998
Testing R2 0.7778 0.7612 0.8367 0.7503 0.8091 0.7775
MSE 375,815 427,414 292,208 447,023 341,682 398,270
MAE 308.46 331.54 272.50 352.67 269.20 315.37
SMAPE 109.95 107.34 112.85 109.44 102.77 108.52
SI 40.37 % 43.04 % 35.50 % 44.04 % 38.50 % 41.54 %
Combined with AE Training R2 0.8057 0.9633 0.8471 0.9901 0.9999 0.999
Testing R2 0.6773 0.7740 0.6661 0.7653 0.8041 0.7922
MSE 577,661 404,452 597,740 420,192 350,599 371,907
MAE 386.53 286.61 391.39 353.33 263.43 291.56
SMAPE 113.69 108.19 117.62 111.91 94.97 103.96
SI 50.05 % 41.55 % 50.73 % 42.68 % 38.99 % 39.91 %

3.4.3. Recursive feature elimination polynomial kernel. C can determine the weight of the overall error re­
RFE achieves reducing the dimension of a dataset by selecting the sults from samples inside the region. The gamma is the coefficient of
features that are more relevant in predicting the target variable. In the gaussian RBF and polynomial kernels and determines the variance of the
RFE algorithm, all the features of a dataset are fed into an ML model two functions. The degree of polynomial (Poly) kernel affects how
first, and the importance analysis of each feature is implemented. The flexible the margins of the region are. The search bounds for C, the
feature that has the least importance on the target variable is deleted. kernel function, the gamma, and the degree are (10− 3, 103), (RBF kernel,
Next, the remaining features are used to fit the ML model and remove Poly kernel, linear kernel), (10− 3, 101), and (1, 5), respectively. Since
the poorest feature again. Repeat the process until the number of the basic form of a KRR model is identical to an SVR model, the
eventual remaining features is the number you set. Two parameters hyperparameters and their search bounds for a KRR model are the same
should be determined when implementing RFE: the number of features as that for an SVR model.
to select, and the algorithm used to select the features. In this study, The key hyperparameters for ANN are learning rate, batch size,
Random Forest (RF) was adopted to help choose features. Like autoen­ number of hidden layers, and neurons. LR is used to describe how
coder and PCA, the number of the select feature was 16. The feature quickly the parameters updates during gradient descent. The number of
selection results by the RFE method for the AC and LC dataset are listed hidden layers and neuron determine the complexity of an ANN model.
below. Batch size means the number of samples fed into an ANN model in one
AC dataset: Age, Ann_temp, Temp_max_diff, Ann_precip, KESAL, iteration. To search for these hyperparameters easily, the number of
Surface_AC_h, Surface_AC pen_77◦ F, Surface_AC P_4.75 mm, Surface_AC neurons in each hidden layer is the same. The search bounds for LR,
Pb, Surface_AC Gmb, Surface_AC Gsb, Surface_AC VV, SG Mr, SB Mr, GB_h, batch size, and the number of hidden layers and neurons are (10− 6,
and GB Mr. 10− 1), (16, 256), (1, 7), and (5, 512), respectively.
LC dataset: Age, Freeze_index, Ann_precip, Surface_AC Gb, Surfa­ The basic estimator of Extra-trees and GB is the decision tree. Like
ce_AC pen_77◦ F, Surface_AC P_9.5 mm, Surface_AC P_4.75 mm, Surfa­ the decision tree, the hyperparameters, including maximum depth
ce_AC P 0.075 mm, Surface_AC Pb, Surface_AC Gmb, Surface_AC Gsb, BC (max_depth), minimum data point needed to split at nodes (min_sam­
9.5 mm sieve, BC VV, SG Mr, SB Mr, and GB Mr. ples_split), and the minimum number of samples required at a leaf
(Min_samples_leaf), are also critical for GB and Extra-trees. Additionally,
4. Development of Machine learning the number of trees (n_estimators) has a direct impact on the prediction
result. The search bounds of max_depth, n_estimators, min_samples_s­
Hyperparameters are crucial because they can control the overall plit, and min_samples_leaf are (1, 50), (1, 1000), (2, 100), and (1, 100),
behavior of ML models. More precisely, choosing suitable hyper­ respectively. As for XGboost, besides max_depth and n_estimators,
parameters is the part process of training an ML model. Hyperparameter Gamma which is like a regularization parameter, subsample which de­
tunning works by training multiple trials with combinations of hyper­ termines how many samples are taken to build a tree, colsample_bytree
parameters to determine the optimal ones. In this study, the hyper­ which describes how many features are used to create each tree, and
parameters of the ML models were going to be tuned using Bayesian min_child_weight means the minimum weight required to create a new
optimization with the 5-fold CV. Each ML model may contain many node were also tuned. Their search bounds are (0.1, 1), (0.1, 1), (0.5,
hyperparameters, but only several key hyperparameters of each ML 0.9), and (5, 300).
model were tuned, other hyperparameters were determined according The hyperparameters of 6 ML models and 18 hybrid models made by
to experience. the combination of each ML model with each dimensionality reduction
The major hyperparameters for SVR include the regularization method were tuned for the AC dataset and the LC dataset. The Bayesian
parameter (C), the kernel function, the gamma, and the degree for the optimization was implemented for all the ML models with 50 iterations.

10
J. Liu et al. Construction and Building Materials 354 (2022) 129162

Fig. 8. Scatter diagrams of the predicted values by ML models and their corresponding hybrid models vs the observed values of AC.

The convergence analysis for ML models and hybrid models made by ML basic ML models, which suggests that dimensionality reduction methods
models with AE (AE-ML) during the Bayesian optimization process is decrease the complexity of ML models and can help reduce models’
shown in Fig. 6 and Fig. 7. The time taken for hyperparameter tunning of training time. According to Fig. 6 and Fig. 7, no matter which the target
each ML model is presented in Table 4. The optimal hyperparameters of is, the training results of CV for ML models and AE-ML models tend to be
ML models and hybrid models are presented in Table 5. As can be seen in stable after 40 iterations. SVR and ANN have the higher mean MSE than
Table 4, the time taken for the hybrid ML models is less than that for other ML models for both AC and LC datasets. When combined with AE,

11
J. Liu et al. Construction and Building Materials 354 (2022) 129162

Fig. 9. Scatter diagrams of the predicted values by ML models and their corresponding hybrid models vs the observed values of LC.

SVR and ANN also have the higher mean MSE than others. models for the AC and LC dataset are shown in Table 6 and Table 7.
Fig. 8 and Fig. 9 present the distribution of the predicted values by basic
5. Results and discussions models and their corresponding hybrid models vs. the observed values of
AC and LC. According to Table 6 and Table 7, no matter which the target
5.1. Performance comparison of models is, R2 in the training dataset for these models is obviously larger than the
test results, which suggests that the majority of models are overfitting,
The train and test results (prediction performance) of ML and hybrid for example, the training R2 is close to 1, but test R2 is less than 0.8 for

12
J. Liu et al. Construction and Building Materials 354 (2022) 129162

Fig. 10. The predicted values by the MEPDG transfer functions vs the observed values of AC and LC.

Fig. 11. Comparison of performance of the MEPDG transfer functions and PCA-ANN for AC and LC.

XGBoost based models. The main reason for getting overfitting is that (
C3
)
compared to the number of features, the number of the train data sample LCTop = 10.56 (12)
1 + eC1 − C2 Log(DI)
is limited. The performance of basic ML models is greater than hybrid
models of ML and RFE for the AC and LC dataset, which implies that the where ACBottom is the area of alligator cracking (% lane area); LCTop is
combination of the RFE method decreases the prediction accuracy, and the length of longitudinal cracks (ft/mi); DI is cumulative damage index;
one possible reason is that the RFE method discards significant parts of
C*1 = − C*2 ; C*2 = − 2.40874 − 39.748(1 + HHMA )− 2.856 ; HHMA is the total
original features. As can be seen in Fig. 8, Fig. 9, Table 6, and Table 7,
HMA thickness (inch); C1,2,3 are the transfer function regression coef­
except for ANN and SVR, the basic ML models can get better for pre­
ficients:C1 = 7.00, C2 = 3.5, C3 = 1.00. DI and HHMA and can be
dicting AC and LC when they are combined with PCA or AE. However, as
extracted from Appendix-II [45].
for ANN-based models, the test R2 of PCA-ANN is higher than that of
The inputs for Eq. (11) and Eq. (12) are aligned with the measured
basic ANN, but the R2 of AE-ANN is smaller than that of basic ANN.
AC and LC from Table 3. Fig. 10 presents the predicted values by the
Among all the models, PCA-ANN and AE-ANN have the largest (0.8400)
transfer functions and their corresponding measured values of AC and
and smallest (0.6002) R2 for prediction AC, respectively. As for pre­
LC. As shown in Fig. 10, the AC transfer function overestimates the AC
dicting LC, PCA-ANN and RFE-ANN have the lowest (272.50) and
when the measured AC are close to 0, but the LC transfer function
highest (441.09) MAE, respectively. Based on the analysis above, PCA-
underestimated the LC when the majority of samples are along the x-
ANN presents the best performance in predicting AC and LC.
axis. The possible reason for underestimation of LC is that MEPDG as­
sumes LC results from top-down cracking, but are actually caused by
5.2. Comparison of MEPDG transfer functions and ML models many reasons, for example, poor joint construction. Fig. 11 shows the
performance comparison of PCA-ANN and the transfer functions on the
To compare the performance of ML models and the calibrated pre­ common part of test samples. In Fig. 11, the PCA-ANN performs rela­
diction functions of MEPDG for AC and LC, the AC and LC were pre­ tively better than MEPDG prediction models for both AC and LC.
dicted based on the transfer functions below [1]:
( )
1 6000 5.3. Features importance analysis: The SHAP values
ACBottom = * +(Log(DI*100))C*
(11)
60 1 + e 1
C 2
The analysis of the SHAP values is a promising method to evaluate

13
J. Liu et al. Construction and Building Materials 354 (2022) 129162

Fig. 12. Feature importance analysis based on SHAP values on the XGBoost model with the original features.

the influence of each factor on a model [75] (XGBoost with the original each group of variables is presented in Fig. 13. Fig. 13 shows that
features in this study). The SHAP value is computed by taking the compared to other pavement layers, surface asphalt concrete (AC) layer
average marginal contribution of a feature. the SHAP values break down is more considerable for evaluation of AC and LC. Subbase/base is in the
a prediction into a baseline value and SHAP values of features, so the middle rank for AC prediction, but it drops to the last rank for LC pre­
model prediction output f(x) for each sample can be also reproduced as diction, it suggests that the cause of AC is more attributed to Subbbase/
the sum of SHAP values and the baseline constant values as follows: base, while Subbase/base is less influential on the cause of LC.
f (x) = baselinevalue + sum(SHAPvaluse) (13)
6. Improvement of asphalt mix design and case study
Where baselinevalue is the mean of all the prediction outputs for
regression models. Every feature of each sample has its corresponding To control cracking problems during the design life period for new
SHAP value. sum(SHAPvaluse) is the sum of the SHAP value of each asphalt pavements, the best LC and AC predictions models (PCA-ANN)
feature. According to Eq. (13), a feature with a positive SHAP value were added to the asphalt mix design process. The AC and LC criteria of
means that this feature of a sample contributes to the increase of the the MEPDG were adopted to be the AC and LC requirements during the
output, and a negative SHAP value means that the feature leads to a mix design process. The procedure of the improved asphalt mix design is
contrary effect on the output. For the whole train dataset, the distribu­ presented in Fig. 14.
tion of the SHAP value of all the samples on all the features is displayed First of all, the initial parameters of pavement, such as design life,
in Fig. 12. If one point of a feature is red, it means the feature value in climate, and properties of pavement except for the surface AC layer,
this sample is high. Points in blue represent lower values. A red point should be determined. Next, asphalt binders and aggregates are selected
(high feature value) with positive SHAP values means that the feature in following the requirements of the Superpave. Subsequently, prepare
high value can make the output increase. The importance of a feature is three trial aggregate blends at the same initial Pb and estimate the
proportional to the absolute mean SHAP value of the feature in all the volumetric properties at 4 % VV by calculating the measured volumetric
samples. In Fig. 12, one feature ranking at the top of the listed feature properties of these trial blends. The proper gradation is then selected by
means the absolute mean SHAP value of this feature is high. No matter checking the AC and LC at the end of pavement life predicted by ML
which target is predicted, the top four most important factors are Age, models and estimated volumetric properties (Voids in the mineral
Surface_AC_Gmb, Ann_precip, and Surface_AC P_4.75 mm. On the other aggregate (VMA), (Voids Filled with Asphalt) VFA, and Dust to binder
hand, compared to the other variables of subgrade/subbase/base layers, ratio (D.P.)). The actual optimal Pb and the corresponding volumetric
GB Mr is found to contribute to the targets most for AC and LC datasets. properties of the selected gradation are determined through establishing
To identify the importance of each group of variables (age, traffic, relationships between different Pb and volumetric properties of labora­
climate, type of pavement layers) on the targets, the average of absolute tory samples in this gradation. Lastly, whether the AC and LC by the ML
mean SHAP values of features from a group was used to capture the models and volumetric properties of the final mixture can satisfy the
importance of that group. The average of absolute mean SHAP values for Superpave specification and AC and LC criteria are verified. The blend
gradations should be adjusted until all the properties of the

14
J. Liu et al. Construction and Building Materials 354 (2022) 129162

Fig. 13. Feature importance analysis of each group of features on the XGBoost model.

corresponding mixture follow the requirements above. traditional Superpave laboratory tests. The actual optimal Pb and the
The mixture of the asphalt surface course for a construction project corresponding volumetric properties, which all meet the requirements
would be determined based on the improved mix design method. The and are not quite close to the boundary of criteria, of Mix- 2 are shown in
construction pavement is a section of a primary highway. The estimated Table 10. Thus, the Mix-2 with 5.1 % asphalt content was recommended
18K-lb. ESAL rate in the study lane is 166,000 ESAL/yr, and the design to be the material of the asphalt surface course of this project.
period is 20 years. Except for the asphalt surface layer, other pavement
layers and their materials’ properties had been determined. The pave­ 7. Summary and conclusions
ment is composed of a poorly fine-grained soils subgrade, a 203 mm (8-
inch) lime treated subbase, an 89 mm (3.5-inch) asphalt treated base, a Properties of asphalt mixture have significant influences on pave­
76 mm (2.9-inch) asphalt binder course, and a 64 mm (2.5-inch) of ment cracking, and properties of asphalt mixture are affected by the
asphalt surface course. The same asphalt binder was used for all the asphalt mix design process. To control pavement cracking during the
asphalt concrete layers. According to Fig. 2, the AC and LC requirements pavement’s service life, the study improved the asphalt mix design
for a primary highway are 20 % lane area and 2000 (ft/mi), respectively. process by establishing ML models to predict AC and LC and introducing
The specific information on climate, estimated traffic, pavement struc­ their criteria. The dataset was collected from the NCHRP 1-37A and the
ture, and pavement materials are displayed in Table 8. LTPP. SVR, ANN, KRR, GB, Extra-trees, XGBoost, and their combination
There were three gradations shown in Table 9 after the initial se­ models with AE, PCA, and RFE were optimized, trained, and compared
lection. Three trial blends samples were first compacted with the same based on the input variables relating to climate, traffic, pavement
initial Pb, and then volumetric properties (Pb, VMA, VFA, and D. P.) of structure, and materials properties of each pavement layer. The per­
each mixture at 4 % VV were estimated based on the measured volu­ formance of the best models was compared to that of global calibrated
metric properties of initial compaction samples. Meanwhile, the AC and functions of predicting AC and LC. Lastly, the best AC and LC prediction
LC of each mixture at 4 % VV at the end of pavement life could be models were applied to design an asphalt mixture of the asphalt surface
predicted by the trained PCA-ANN model. The estimated volumetric course for a pavement project, the following conclusions were drawn:
properties and the predicted AC and LC of the three blends are presented
in Table 9. As can be seen from Table 9, the predicted AC and LC, VFA, ▪ The results on the selection of the number of reduced features
VMA, and D.P. of all the three mixes can meet the requirements show that 28 neurons in the hidden layer and 16 neurons in the
(AASHTO M323 and MEPDG). However, the mixes whose properties are bottleneck layer were determined for encoder sections of the
very close to the boundary of criteria were going to be excluded. Mix-1 AE models. The First 15 and 16 PCAs can explain close to 95 %
and the Mix-3 were eliminated because the D.P. of Mix-1 is 1.16 (close to variance of the original dataset.
1.2) and the VMA of Mix-3 is 14.0 (equal to 14.0). Therefore, the Mix-2 ▪ The Bayesian optimization process for tuning hyperparameters
was selected to obtain the actual optimal Pb by implementing the shows that no matter which the target is, the training results of

15
J. Liu et al. Construction and Building Materials 354 (2022) 129162

Fig. 14. The procedure of the improved asphalt mix design with AC and LC prediction ML models.

Table 8 Table 9
The basic information about climate, traffic, pavement materials, and pavement A summary of three asphalt mixtures determined by the Superpave mix design.
structure except for the surface AC layer for the pavement project. Variables Mix-1 Mix-2 Mix-3 Requirement
Type of variables Variables Values (NMPS (NMPS = (NMPS = (AASHTO M323,
= 9.5 12.5 mm) 12.5 mm) MEPDG)
Design life Age (year) 20 mm)
Traffic Estimated ESALs rate (k/year) 166
Climate Average annual air temperature (℃) 15 Surface AC thickness 2.5
Annual freeze index 40 Penetration at 77 ◦ F 69
Average monthly temperature maximum 22 (0.1 mm)
difference (℃) Gb 1.0197
Average annual precipitation (cm) /year 1144 Percent 19 mm 100 100 99
Subgrade Mr (MPa) 72.6 Passing 9.5 95 84 62
Materials type Fine-grained soils sieve size mm
Subbase Layer Thickness (inch) 8.0 (%) 4.75 65 43 35
Materials type Lime treated soil mm
Mr (MPa) 113.2 0.075 5.8 4.4 4.2
Base Layer Granular Base thickness (inch) —— mm
Granular Base Mr (MPa) —— Gsb 2.624 2.631 2.639
Asphalt Treated base thickness (inch) 3.5 Estimated Gmb@Ndes 2.343 2.365 2.384
Binder course Penetration at 77 ◦ F (0.1 mm) 69 Air void (%) 4 4 4
Percent passing sieve size 19 mm 99 Estimated Optimal Pb 5.31 5.18 4.70
9.5 mm 62 (%)
4.75 mm 35 Estimated VMA (%) 15.4 14.8 14.0 Min./NMPS
0.075 mm 4.2 15.0/9.5 mm
Pb (%) 4.7 14.0/12.5 mm
VV (%) 6.35 Estimated VFA (%) 74.1 72.9 70 65–75
Thickness (inch) 2.9 Estimated Dust to 1.16 0.94 0.98 0.6–1.2
binder ratio (D.P.)
Predicted alligator 6.02 6.09 7.12 Max. 20
CV for all the ML models and all the AE-ML models tend to be cracking (% lane
area)
stable after 40 iterations.
Predicted longitudinal 963.38 1053.41 1116.36 Max. 2000
▪ The performance of basic ML models is greater than hybrid cracking (ft. /mi)
models of RFE-ML for the AC and LC datasets. Except for ANN ⨯ ✓ ⨯
and SVR, the basic ML models can get better for predicting AC
and LC when they are combined with PCA or AE. Compared to

16
J. Liu et al. Construction and Building Materials 354 (2022) 129162

Table 10
The real optimal Pb and the corresponding volumetric properties for Mix-3.
Mixtures Actual Opt. Pb Actual Actual VMA Actual Actual D. Predicted alligator cracking (% Predicted longitudinal cracking
(%) Gmb (%) VFA P. lane area) (ft. /mi)
(%)

Mix-2 5.1 2.356 15.0 73 0.96 6.13 1075.34


Requirements (AASHTO M323, Min. 14 65–75 0.6–1.2 Max. 20 Max. 2000
MEPDG)

other models trained in this study, ANN-PCA has the best per­ Data availability
formance for predicting AC and LC.
▪ Feature importance analysis based on the SHAP values shows The authors do not have permission to share data.
that no matter which target is predicted, the top four most
important factors are Age, Surface_AC_Gmb, Ann_precip, and Acknowledgments
Surface_AC P_4.75 mm. Compared to other pavement layers,
Asphalt surface course is more considerable for the evaluation This work is sponsored by a grant from the Center for Integrated
of AC and LC. The cause of AC is more attributed to Subbbase/ Asset Management for Multimodal Transportation Infrastructure Sys­
base, but Subbbase/base is less influential on the cause of LC. tems (CIAMTIS), a US Department of Transportation University Trans­
▪ With the improved mix design process by introducing AC and portation Center, United States under federal grant number
LC prediction models, Mix-2 with 5.1 % asphalt content was 69A3551847103. The authors are grateful for the support.
determined for the asphalt surface course of the construction
pavement. References:

Compared to traditional laboratory tests (e.g., repeated flexural [1] T. Officials, Mechanistic-Empirical Pavement Design Guide: A Manual of Practice,
AASHTO, Washington, DC, USA, 2008.
bending) to evaluate the fatigue life of mixture during asphalt mix [2] A. Mateos, J.A. Gómez, R. Hernández, Y. Tan, L.G.L. Salazar, A. Vargas-Nordcbeck,
design phase, the method to evaluate the crack resistance (AC and LC) of Application of the logit model for the analysis of asphalt fatigue tests results,
asphalt concrete in the circumstance, where the asphalt concrete is Constr. Build. Mater. 82 (2015) 53–60.
[3] L.A. Myers, R. Roque, B.E. Ruth, Mechanisms of surface-initiated longitudinal
assumed to be paved on actual pavements, with ML models can save wheel path cracks in high-type bituminous pavements, J. Assoc. Asphalt Paving
much time and labor. Professional engineers can evaluate the pavement Technol. 67 (1998).
severe fatigue cracks potential of a designed asphalt mixture in advance [4] A. Collop, D. Cebon, Stiffness reductions of flexible pavements due to cumulative
fatigue damage, J. Transp. Eng. 122 (2) (1996) 131–139.
and adjust binder, aggregates, gradation, or asphalt content in time. [5] F. Zhou, S. Hu, T. Scullion, Integrated asphalt (overlay) mixture design, balancing
However, the proposed method is only applicable to asphalt surface rutting and cracking requirements, Texas Transportation Institute, Texas A & M
course of new flexible pavements but does not involve the other asphalt University System, 2006.
[6] A.A. Tayebali, G.M. Rowe, J.B. Sousa, Fatigue response of asphalt-aggregate
concrete layers and overlay of old pavement. Sometimes, maintenance
mixtures (with discussion), J. Assoc. Asphalt Paving Technol. 61 (1992).
or rehabilitation activities will be implemented to repair cracks before [7] F. Zhou, S. Im, S. Hu, D. Newcomb, T. Scullion, Selection and preliminary
the end of pavement life, but asphalt mix is selected in the method only evaluation of laboratory cracking tests for routine asphalt mix designs, Road Mater.
by predicting AC and LC at the end of pavement life. Life cycle perfor­ Pavement Des. 18 (sup1) (2017) 62–86.
[8] F. Zhou, T. Scullion, Upgraded overlay tester and its application to characterize
mance and cost should be considered. The design of asphalt concrete reflection cracking resistance of asphalt mixtures, Texas Transportation Institute,
should not only consider the future cracking progression, but also still Texas A & M University System, 2003.
need to consider future possible maintenance or rehabilitation activities. [9] D. Bodin, C. de La Roche, G. Pijaudier-Cabot, Size effect regarding fatigue
evaluation of asphalt mixtures: laboratory cantilever bending tests, Road Mater.
Besides, the generalization of ML models is limited because the AC and Pavement Des. 7 (sup1) (2006) 181–200.
LC prediction models established by this study only work for the tradi­ [10] B.-W. Tsai, J.T. Harvey, C.L. Monismith, Application of Weibull theory in
tional hot mix asphalt. Therefore, the diversity of the training data will prediction of asphalt concrete fatigue performance, Transp. Res. Rec. 1832 (1)
(2003) 121–130.
be enhanced in the future. Specifically, the type of asphalt mixture in the [11] R. Yeo, Fatigue performance of cemented materials under accelerated loading:
dataset is supposed to be more varied and include recycled asphalt influence of vertical loading on the performance of unbound and cemented
concrete, stone matrix asphalt (SMA), open graded friction course materials, 2008.
[12] H.-W. Ker, Y.-H. Lee, P.-H. Wu, Development of fatigue cracking prediction models
(OGCF), and warm mix asphalt. using long-term pavement performance database, J. Transp. Eng. 134 (11) (2008)
477–482.
CRediT authorship contribution statement [13] K.D. Hall, D.X. Xiao, K.C. Wang, Calibration of the mechanistic–empirical
pavement design guidefor flexible pavement design in Arkansas, Transp. Res. Rec.
2226 (1) (2011) 135–141.
Jian Liu: Methodology, Supervision, Data curation, Formal analysis, [14] Y.R. Kim, F.M. Jadoun, T. Hou, N. Muthadi, Local calibration of the MEPDG for
Visualization, Writing – original draft, Writing – review & editing. flexible pavement design, North Carolina State University. Dept. of Civil,
Fangyu Liu: Validation, Formal analysis, Investigation. Hongren Gong: Construction, and …, 2011.
[15] M.I. Souliman, M.S. Mamlouk, M.M. El-Basyouny, C.E. Zapata, Calibration of the
Resources, Supervision. Ebenezer O. Fanijo: Investigation. Linbing AASHTO MEPDG for flexible pavement for arizona conditions, Proceedings of the
Wang: Conceptualization, Resources, Supervision, Project administra­ Transportation Research Board 89th Annual Meeting, Transportation Research
tion, Funding acquisition. Board Washington, DC, USA, 2010, pp. 243-286.
[16] K.C. Wang, J. Zaniewski, G. Way, Probabilistic behavior of pavements, J. Transp.
Eng. 120 (3) (1994) 358–375.
Declaration of Competing Interest [17] L. Sun, W.R. Hudson, Probabilistic approaches for pavement fatigue cracking
prediction based on cumulative damage using Miner’s law, J. Eng. Mech. 131 (5)
(2005) 546–549.
The authors declare that they have no known competing financial [18] L. Gao, J.P. Aguiar-Moya, Z. Zhang, Bayesian analysis of heterogeneity in modeling
interests or personal relationships that could have appeared to influence of pavement fatigue cracking, J. Comput. Civil Eng. 26 (1) (2012) 37–43.
the work reported in this paper. [19] Q. Dong, B. Huang, Evaluation of influence factors on crack initiation of LTPP
resurfaced-asphalt pavements using parametric survival analysis, J. Perform.
Constr. Facil 28 (2) (2014) 412–421.
[20] A. Abed, N. Thom, L. Neves, Probabilistic prediction of asphalt pavement
performance, Road Mater. Pavement Des. 20 (sup1) (2019) S247–S264.

17
J. Liu et al. Construction and Building Materials 354 (2022) 129162

[21] Y. Wang, A. Norouzi, Y.R. Kim, Comparison of fatigue cracking performance of [48] M. Wedel, P. Kannan, Marketing analytics for data-rich environments, J. Mark. 80
asphalt pavements predicted by pavement ME and LVECD programs, Transp. Res. (6) (2016) 97–121.
Rec. 2590 (1) (2016) 44–55. [49] I. Kononenko, Machine learning for medical diagnosis: history, state of the art and
[22] Z. Ambassa, F. Allou, C. Petit, R.M. Eko, Fatigue life prediction of an asphalt perspective, Artif. Intell. Med. 23 (1) (2001) 89–109.
pavement subjected to multiple axle loadings with viscoelastic FEM, Constr. Build. [50] J. Stilgoe, Machine learning, social learning and the governance of self-driving
Mater. 43 (2013) 443–452. cars, Soc. Stud. Sci. 48 (1) (2018) 25–56.
[23] A. Norouzi, Y. Richard Kim, Mechanistic evaluation of fatigue cracking in asphalt [51] Y. Reich, Machine learning techniques for civil engineering problems, Comput.-
pavements, Int. J. Pavement Eng. 18 (6) (2017) 530–546. Aided Civ. Infrastruct. Eng. 12 (4) (1997) 295–310.
[24] Y.R. Kim, C. Baek, B.S. Underwood, V. Subramanian, M.N. Guddati, K. Lee, [52] Z. Li, J. Liu, An evaluation method for the skeleton structure of cement-stabilized
Application of viscoelastic continuum damage model based finite element analysis crushed rock material using X-Ray CT images, J. Test. Eval. 48 (5) (2019)
to predict the fatigue performance of asphalt pavements, KSCE J. Civ. Eng. 12 (2) 3341–3359.
(2008) 109–120. [53] J. Liu, F. Liu, C. Zheng, D. Zhou, L. Wang, Optimizing asphalt mix design through
[25] H. Gong, Y. Sun, Z. Mei, B. Huang, Improving accuracy of rutting prediction for predicting effective asphalt content and absorbed asphalt content using machine
mechanistic-empirical pavement design guide with deep neural networks, Constr. learning, Constr. Build. Mater. 325 (2022), 126607.
Build. Mater. 190 (2018) 710–718. [54] F. Liu, W. Ding, Y. Qiao, L. Wang, An artificial neural network model on tensile
[26] A.J. Haddad, G.R. Chehab, G.A. Saad, The use of deep neural networks for behavior of hybrid steel-PVA fiber reinforced concrete containing fly ash and slag
developing generic pavement rutting predictive models, Int. J. Pavement Eng. power, Front. Struct. Civ. Eng. 14 (6) (2020) 1299–1315.
(2021) 1–17. [55] V. Vapnik, The nature of statistical learning theory, Springer science & business
[27] C. Wang, S. Xu, J. Liu, J. Yang, C. Liu, Building an improved artificial neural media2013.
network model based on deeply optimizing the input variables to enhance rutting [56] P. Geurts, D. Ernst, L. Wehenkel, Extremely randomized trees, Machine learning 63
prediction, Constr. Build. Mater. 348 (2022), 128658. (1) (2006) 3–42.
[28] A. Fathi, M. Mazari, M. Saghafi, A. Hosseini, S. Kumar, Parametric study of [57] K. Gopalakrishnan, S. Kim, Support vector machines approach to HMA stiffness
pavement deterioration using machine learning algorithms, Airfield Highway prediction, J. Eng. Mech. 137 (2) (2011) 138–146.
Pavements 2019 (2019) 31–41. [58] W. Xu, X. Huang, Z. Yang, M. Zhou, J. Huang, Developing hybrid machine learning
[29] H. Gong, Y. Sun, W. Hu, B. Huang, Neural networks for fatigue cracking prediction models to determine the dynamic modulus (E*) of asphalt mixtures using
using outputs from pavement mechanistic-empirical design, Int. J. Pavement Eng. parameters in Witczak 1–40D model: a comparative study, Materials 15 (5) (2022)
22 (2) (2021) 162–172. 1791.
[30] H. Gong, Y. Sun, B. Huang, Gradient boosted models for enhancing fatigue cracking [59] X. Wang, J. Zhao, Q. Li, N. Fang, P. Wang, L. Ding, S. Li, A hybrid model for
prediction in mechanistic-empirical pavement design guide, J. Transp. Eng. Part B: prediction in asphalt pavement performance based on support vector machine and
Pavements 145 (2) (2019) 04019014. grey relation analysis, J. Adv. Transp. 2020 (2020).
[31] F. Liu, J. Liu, L. Wang, Deep learning and infrared thermography for asphalt [60] A. Eleyedath, A.K. Swamy, Prediction of dynamic modulus of asphalt concrete
pavement crack severity classification, Autom. Constr. 140 (2022), 104383. using hybrid machine learning technique, Int. J. Pavement Eng. (2020) 1–16.
[32] F. Liu, J. Liu, L. Wang, Asphalt pavement crack detection based on convolutional [61] D. Singh, M. Zaman, S. Commuri, Artificial neural network modeling for dynamic
neural network and infrared thermography, IEEE Trans. Intell. Transp. Syst. modulus of hot mix asphalt using aggregate shape properties, J. Mater. Civ. Eng. 25
(2022). (1) (2013) 54–62.
[33] R. Guo, D. Fu, G. Sollazzo, An ensemble learning model for asphalt pavement [62] H. Gong, Y. Sun, Y. Dong, B. Han, P. Polaczyk, W. Hu, B. Huang, Improved
performance prediction based on gradient boosting decision tree, Int. J. Pavement estimation of dynamic modulus for hot mix asphalt using deep learning, Constr.
Eng. (2021) 1–14. Build. Mater. 263 (2020), 119912.
[34] S. Choi, M. Do, Development of the road pavement deterioration model based on [63] H. Ceylan, C.W. Schwartz, S. Kim, K. Gopalakrishnan, Accuracy of predictive
the deep learning method, Electronics 9 (1) (2019) 3. models for dynamic modulus of hot-mix asphalt, J. Mater. Civ. Eng. 21 (6) (2009)
[35] H. Gong, Y. Sun, X. Shu, B. Huang, Use of random forests regression for predicting 286–293.
IRI of asphalt pavements, Constr. Build. Mater. 189 (2018) 890–897. [64] G.S. Moussa, M. Owais, Pre-trained deep learning for hot-mix asphalt dynamic
[36] N. Abdelaziz, R.T. Abd El-Hakim, S.M. El-Badawy, H.A. Afify, International modulus prediction with laboratory effort reduction, Constr. Build. Mater. 265
Roughness Index prediction model for flexible pavements, Int. J. Pavement Eng. 21 (2020), 120239.
(1) (2020) 88–99. [65] M. Nazemi, A. Heidaripanah, Support vector machine to predict the indirect tensile
[37] J.-D. Lin, J.-T. Yau, L.-H. Hsiao, Correlation analysis between international strength of foamed bitumen-stabilised base course materials, Road Mater.
roughness index (IRI) and pavement distress by neural network, 82nd Annual Pavement Des. 17 (3) (2016) 768–778.
Meeting of the Transportation Research Board, 2003, pp. 1-21. [66] M. Maalouf, N. Khoury, T.B. Trafalis, Support vector regression to predict asphalt
[38] A. Fathi, M. Mazari, M. Saghafi, A. Hosseini, S. Kumar, Parametric study of mix performance, Int. J. Numerical Analyt. Methods Geomech. 32 (16) (2008)
pavement deterioration using machine learning algorithms, Airfield and highway 1989–1996.
pavements 2019: Innovation and sustainability in highway and airfield pavement [67] S. Rahman, A. Bhasin, A. Smit, Exploring the use of machine learning to predict
technology, American Society of Civil Engineers Reston, VA2019, pp. 31-41. metrics related to asphalt mixture performance, Constr. Build. Mater. 295 (2021),
[39] S. Inkoom, J. Sobanjo, A. Barbu, X. Niu, Prediction of the crack condition of 123585.
highway pavements using machine learning models, Struct. Infrastruct. Eng. 15 (7) [68] F.-Y. Liu, W.-Q. Ding, Y.-F. Qiao, L.-B. Wang, Q.-Y. Chen, Compressive behavior of
(2019) 940–953. hybrid steel-polyvinyl alcohol fiber-reinforced concrete containing fly ash and slag
[40] S. Inkoom, J. Sobanjo, A. Barbu, X. Niu, Pavement crack rating using machine powder: experiments and an artificial neural network model, J. Zhejiang Univ.-Sci.
learning frameworks: partitioning, bootstrap forest, boosted trees, Naïve bayes, A 22 (9) (2021) 721–735.
and K-Nearest neighbors, J. Transp. Eng. Part B: Pavements 145 (3) (2019) [69] J.H. Choi, T.M. Adams, H.U. Bahia, Pavement roughness modeling using back-
04019031. propagation neural networks, Comput.-Aided Civ. Infrastruct. Eng. 19 (4) (2004)
[41] G.S. Moussa, M. Owais, Modeling hot-mix asphalt dynamic modulus using deep 295–303.
residual neural networks: parametric and sensitivity analysis study, Constr. Build. [70] M. Hossain, L. Gopisetti, M. Miah, Prediction of international roughness index of
Mater. 294 (2021), 123589. flexible pavements from climate and traffic data using artificial neural network
[42] H. Worthey, J.J. Yang, S.S. Kim, Tree-based ensemble methods: predicting asphalt modeling, Airfield Highway Pavements 2017 2017 256-267.
mixture dynamic modulus for flexible pavement design, KSCE J. Civ. Eng. 25 (11) [71] M. Hossain, L. Gopisetti, M. Miah, International roughness index prediction of
(2021) 4231–4239. flexible pavements using neural networks, J. Transp. Eng. Part B: Pavements 145
[43] Y. Ali, F. Hussain, M. Irfan, A.S. Buller, An eXtreme Gradient Boosting model for (1) (2019) 04018058.
predicting dynamic modulus of asphalt concrete mixtures, Constr. Build. Mater. [72] J. Wu, X.-Y. Chen, H. Zhang, L.-D. Xiong, H. Lei, S.-H. Deng, Hyperparameter
295 (2021), 123642. optimization for machine learning models based on Bayesian optimization,
[44] H. Gong, Y. Sun, W. Hu, P.A. Polaczyk, B. Huang, Investigating impacts of asphalt J. Electron. Sci. Technol. 17 (1) (2019) 26–40.
mixture properties on pavement performance using LTPP data through random [73] E. Brochu, V.M. Cora, N. De Freitas, A tutorial on Bayesian optimization of
forests, Constr. Build. Mater. 204 (2019) 203–212. expensive cost functions, with application to active user modeling and hierarchical
[45] A.R. Associates, Guide for mechanisticempirical design of new and rehabilitated reinforcement learning, arXiv preprint arXiv:1012.2599 (2010).
pavement structures, Final Rep. No. NCHRP 1-37A, Transportation Research Board [74] P. Baldi, Autoencoders, unsupervised learning, and deep architectures, Proceedings
of the National Academies, Washington, DC, 2004. of ICML workshop on unsupervised and transfer learning, JMLR Workshop and
[46] Q. Dong, B. Huang, Failure probability of resurfaced preventive maintenance Conference Proceedings, 2012, pp. 37-49.
treatments: investigation into long-term pavement performance program, Transp. [75] H. Kaur, H. Nori, S. Jenkins, R. Caruana, H. Wallach, J. Wortman Vaughan,
Res. Rec. 2481 (1) (2015) 65–74. Interpreting interpretability: understanding data scientists’ use of interpretability
[47] T. Officials, Mechanistic-empirical pavement design guide: a manual of practice, tools for machine learning, in: Proceedings of the 2020 CHI Conference on Human
AASHTO2008. Factors in Computing Systems, 2020, pp. 1–14.

18

You might also like