KSCE Journal of Civil Engineering (2023) 27(1):207-217 pISSN 1226-7988, eISSN 1976-3808

DOI 10.1007/s12205-022-0448-z www.springer.com/12205

Highway Engineering

A Machine Learning Tool for Pavement Design and Analysis

Guangwei Yang a, Kamyar C. Mahboub b

, Ryan L. Renfroc, Clark Graves d
,
and Kelvin C. P. Wang a
a
School of Civil and Environmental Engineering, Oklahoma State University, Stillwater, OK 74078, USA
b
Dept. of Civil Engineering, University of Kentucky, Lexington, KY 40506, USA
c
Gresham Smith, Nashville, TN 37201, USA
d
Kentucky Transportation Center, Lexington, KY 40506, USA

ARTICLE HISTORY ABSTRACT

Received 22 March 2022 The AASHTOWare Pavement ME Design program is a pavement analysis tool, which is
Accepted 13 September 2022 typically used for design purposes through an iterative trial-and-error process. To help the
Published Online 18 October 2022 designer with a reasonable starting point in this iterative process, this paper introduces a
machine learning method to embrace the recently updated models in AASHTOWare
KEYWORDS Pavement ME Design software for pavement design. A total number of 79,600 pavement
design scenarios (55,800 for flexible pavements and 23,800 for rigid pavements) were
AASHTOWare pavement ME design performed using the AASHTOWare Pavement ME Design software to consider various design
Pavement performance inputs, such as: design life, traffic volume, climate zone, thickness, and modulus of pavement
Pavement thickness design layers. The inputs and outputs of these design scenarios were used to develop the multi-
Multi target regression output Random Forests model to simultaneously predict multiple pavement distresses and
Machine learning thicknesses of pavement layers. The results indicate that the multi-output Random Forests
Multi-output random forests model can accurately predict pavement distresses and thicknesses for asphalt and concrete
pavements. This tool will simplify pavement design procedure based on the models in the
AASHTOWare Pavement ME Design software.

1. Introduction field testing on a test track in Ottawa, Illinois, in the 1950s with
later modifications (AASHTO, 1993). Unfortunately, this method
An over-designed pavement results in a thicker structure than has certain limitations, and it is not ideal for many pavement
necessary with higher associated costs, while an under-designed design situations. For example, 1) the original traffic, materials,
pavement results in a thinner structure than needed with increased and the climate conditions of the test track were significantly
maintenance cost due to premature failure (ODOT, 2015). Therefore, different from those of the present-day pavement designs and
a well-designed and robust pavement can have a substantially other regions; 2) the pavement performance was subjectively
lower life-cycle cost, which translates into a higher level of evaluated via the pavement serviceability index (PSI); 3) the
sustainability. To select the appropriate pavement type and methodology is incompatible with modern pavement rehabilitation
adequately design a pavement structure for a new or rehabilitated models and methods; and 4) it lacks mechanistic performance
project, the designer must follow a pavement design method to modelling (Zhang et al., 2000; ARA, 2004; Schwartz and Carvalho,
calculate specific pavement course thickness requirements by 2007; Li et al., 2011).
considering traffic, soil, climate, and other in-situ conditions. In 2004, the Mechanistic-Empirical Pavement Design Guide
Currently, a relatively large number of highway agencies (MEPDG) was developed under the National Cooperative Highway
primarily use the 1993 AASHTO Guide for Design of Pavement Research Program (NCHRP) Project 1-37A to address some of the
Structures (also known as 1993 AASHTO Guide), or some locally limitations in the 1993 AASHTO Guide (ARA, 2004). The MEPDG
modified version of that, to design pavement structures in the can mechanistically predict multiple pavement performance
United States (AASHTO, 1993). The empirical equations in the indicators based on engineering material properties, environmental
1993 AASHTO Guide were developed based on two years of conditions, vehicle class, and traffic loading characteristics for

CORRESPONDENCE Guangwei Yang [email protected] School of Civil & Environmental Engineering, Oklahoma State University, Stillwater, OK 74078 USA
ⓒ 2023 Korean Society of Civil Engineers
208 G. Yang et al.

new and rehabilitated pavement structures. This work was based within the problem set, and by taking advantage of inter-correlated
upon the long-term pavement performance (LTPP) sites in the parameters (Aguiar et al., 2019; Santana et al., 2020).
US and Canada (ARA, 2004; Schwartz and Carvalho, 2007; FHWA, More specifically, the multi-output Random Forests (Multi RF)
2017). The overall objective of the new design guide was to has been explored with the goal of increasing the predictive
improve the reliability of the pavement design, to make the pavement performance of the MTR problems (Linusson, 2013; González et al.,
design process more scientific, and to better predict the in-service 2016; Hoffman et al., 2021). The Random Forest (RF) is a robust
performance of pavements (ARA, 2004). machine learning algorithm with demonstrated success in pavement
The AASHTOWare Pavement ME Design software was first engineering, such as prediction of pavement IRI (Gong et al., 2018),
introduced under version 0.7 in June 2004, and it has been and pavement friction (Yang et al., 2019), as well as investigation into
continually enhanced and updated (AASHTO, 2015). It is important the impacts of asphalt mixture properties on pavement performance
to note that the AASHTOWare Pavement ME Design software is (Gong et al., 2019). The Multi RF consists of fitting one RF regressor
a pavement analysis tool. Therefore, the pavement designer must per target. For instance, after training the Multi RF for pavement
assume a trial pavement structure at the outset of the design distress prediction, the model includes one RF regressor for rutting,
process, and then use the AASHTOWare Pavement ME Design another RF regressor for IRI, etc. Because multiple pavement
software to calculate the predicted performance under specific distresses are predicted based on a given pavement structure in the
design conditions. Without an initial reasonable estimate of the AASHTOWare Pavement ME Design, the Multi RF method was
pavement structure, the trial-and-error type of user design experience explored in this study to better correlate with the latest
with AASHTOWare Pavement ME Design may end up being AASHTOWare Pavement ME Design software.
very time-consuming. Further, the iterative process is not as This paper presents a machine learning approach, named as
straightforward as compared to a traditional pavement design KYPDT-M, for pavement design and analysis using the Multi RF to
method such as the 1993 AASHTO Guide (Schwartz and Carvalho, simultaneously predict pavement distresses and estimate the
2007; Schwartz etc., 2011; Graves and Mahboub, 2012; Iowa thicknesses of AC or PCC layer. A total number of 79,600 pavement
SUDAS, 2021). design scenarios (55,800 for flexible pavements and 23,800 for rigid
The Kentucky Pavement Design Tool (KYPDT) was developed pavements) were performed using the AASHTOWare Pavement
in 2010 as a powerful and user-friendly tool to perform pavement ME Design 2.5.5 to prepare a massive matrix of inputs and
analysis and design for both flexible and rigid pavements (Graves, outputs for training the Multi RF. Various design inputs were
2012). This tool minimizes the need for the designer to conduct considered, such as: design life, traffic volume, climate zone, and
many trial-and-error software runs, and it offers a reasonable first pavement structures with different thickness and modulus values
estimate for the pavement design. However, it used the multiple of AC/PCC, base (or subbase), and subgrade layers. Finally,
adaptive regression splines (MARS) technique to predict each Multi RF models were developed for pavement design (estimated
pavement distress and estimate the corresponding thickness layer thickness values) and pavement analysis (predicted pavement
of the AC or PCC layer. For instance, three MARS models distresses). This tool is expected to benefit transportation agencies to
were needed to predict pavement fatigue cracking, rutting, perform pavement design and analysis, especially local agencies
and international roughness index (IRI) for a given pavement with limited resources who might need a good first estimate for
structure. Additionally, to back-calculate the thickness of the AC their pavement design, which could be fine-tuned via Pavement
layer, three MARS models (fatigue cracking model, rutting model, ME Design -- if desired.
and IRI model) were developed in KYPDT, which generated
different thickness values of the AC layer for each distress model. 2. Features and Assumptions
For example, the predicted thickness of the AC layer may be 8.0 in.
(20.3 cm) generated by the fatigue cracking model, while it may be As shown in Fig. 1, pavement design and analysis are the two
11.5 in. (29.2 cm) generated by the rutting model, and 9.0 in. essential features of KYPDT-M:
(22.9 cm) due to the IRI model (Graves, 2012). Therefore, the user 1. Pavement Design –The Multi RF estimates thicknesses of
was left with the decision to choose a thickness corresponding to a AC/PCC layer, base layer, or subbase layer based upon the
distress mode that the user might think would be the most critical. traffic volume, climate zone, materials in each pavement
In recent years, the multi target regression (MTR) method that layer, and the desired pavement distresses threshold at the
simultaneously predicts multiple output/target variables using end of design life.
the same set of inputs has been reported to be advantageous over 2. Pavement Analysis – The Multi RF predicts pavement
the single-target regression models in various studies (Borchani distresses at the end of design life based upon the traffic
et al., 2015; Xi et al., 2018; Xu et al., 2019). The single-target volume, climate zone, materials in each pavement layer, and
regression trains a classifier to predict a single object, whereas thicknesses of AC/PCC layer, base layer, or subbase layer.
the MTR trains a classifier to predict a set of objects. Therefore, Further, two assumptions are considered in KYPDT-M:
multiple parameters can be predicted simultaneously using the 1. The default traffic distribution and material parameters of
MTR method to save the model training time and improve the AC/PCC layers in the AASHTOWare Pavement ME Design
overall predictive performance by exploring inter-target correlations were used to run these pavement scenarios.
 KSCE Journal of Civil Engineering 209

Fig. 1. Summary of Inputs and Outputs of KYPDT-M: (a) Inputs and Outputs of Pavement Design, (b) Inputs and Outputs of Pavement Analysis

2. For the case of concrete pavements with doweled transverse 3. Database of KYPDT-M
joints, a dowel bar spacing of 12 inches was used. And the
size of dowel bars was held constant according to the 3.1 Climate Zones
thickness of the concrete pavement, as shown in Table 1. To ensure that a wide range of climate conditions was covered in
KYPDT-M, five regions from four climatic zones were selected
Table 1. Pavement Thicknesses and Corresponding Dowel Bar Sizes in accordance with the LTPP’s climatic zone (Wu et al., 2010):
for Rigid Pavements (PGI, 2003) Denver, CO from the dry-freeze zone, Phoenix, AZ from the
Pavement Thickness Dowel Bar Diameter
dry-no freeze zone, Atlanta, GA from the wet-no freeze zone,
(inch (cm)) (inch (cm)) and Boston MA as well as Lexington, KY from the wet-freeze
zone. For each city, multiple weather stations embedded in the
6 (15.2) 1.25 (3.2)
AASHTOWare Pavement ME Design 2.5.5 software were selected
8 (20.3) 1.25 (3.2)
for the analysis, as listed in Table 2. The AASHTOWare Pavement
9 (22.9) 1.25 (3.2)
ME Design 2.5.5 program uses the Modern-Era Retrospective
10 (25.4) 1.25 (3.2)
Analysis for Research and Applications (MERRA) climate stations
12 (30.5) 1.50 (3.8)
for flexible pavement design, and North American Regional
14 (35.6) 1.75 (4.4)
Reanalysis (NARR) climate stations for rigid pavement design.

Table 2. List of Cities and Asphalt Binder PG Grades for Each Climatic Zone
Flexible Pavement Rigid Pavement
Climatic Zone City
PG Grade MERRA Climate Stations NARR Climate Stations
Dry-Freeze Denver, CO 76-28 144121, 143545, 144122. 03065, 3017d, 23036, 93067, 3040a.
Dry-No Freeze Phoenix, AZ 76-10 136621, 136622, 136046. 03184, 03192, 23183.
Wet-No Freeze Atlanta, GA 76-22 136666, 136667, 137242. 13874, 03888, 53863.
Wet-Freeze Boston, MA 64-28 147055, 147056, 146479. 14739, 14702, 54704, 54733.
Wet-Freeze Lexington, KY 64-22 141849, 141850, 141851. 93820.
210 G. Yang et al.

Take Denver, CO as an example, the climate data for pavement A-5. The modulus values range from 103,421 kPa (15,000
analysis was from MERRA, and it included the following stations: psi) to 275,790 kPa (40,000 psi) for base layers, 55,158 kPa
144121, 143545, and 144122. (8,000 psi) to 124,106 kPa (18,000 psi) for subgrade.
Further, a proper asphalt binder grade had to be selected for 6. Base thickness (inch (cm)) with subbase: 6 (15.2), 8 (20.3),
flexible pavements in KYPDT-M to reflect the local climate. 10 (25.4), 12 (30.5).
Therefore, the research team chose a typical asphalt binder for 7. Subbase thickness (inch (cm)): 8 (20.3), 10 (25.4), 12 (30.5),
each city per “US State Binder Specifications” (AI, 2020): PG 14 (35.6), 16 (40.6).
76-28 for Denver, CO; PG 76-10 for Phoenix, AZ; PG 76-22 for 8. Base, subbase, and subgrade combinations with default
Atlanta GA; PG 64-28 for Boston, MA; and PG 64-22 for modulus values in AASHTOWare Pavement ME Design
Lexington, KY, as summarized in Table 2. software representing AASHTO A-1-a & A-3 & A-1-a, A-
1-a & A-3 & A-7-6, A-1-a & River-run Gravel & A-7-6,
3.2 Considered Scenarios A-1-a & River-run Gravel & A-5, A-3 & River-run Gravel
This study considers four types of pavement design options for & A-7-6, and A-3 & River-run Gravel & A-5. The modulus
each climate zone using the AASHTOWare Pavement ME Design values range from 199,948 kPa (29,000 psi) to 275,790 kPa
software: 1) flexible pavement with base layer; 2) flexible pavement (40,000 psi) for base layers, 103,421 kPa (15,000 psi) to
with base and subbase layers; 3) rigid pavement without dowel 199,948 kPa (29,000 psi) for subbase layers, and 55,158 kPa
bars; 4) rigid pavement with dowel bars. Each type of pavement (8,000 psi) to 124,106 kPa (18,000 psi) for subgrade.
design considers a vast array of input variables, including different Therefore, there are 1,260, 4,320, 1,120, and 1,260 pavement
design lives, traffic volumes, AC (PCC) thickness, base (subbase), designs for scenarios in climatic zones 1, 2, 3, and 4, respectively.
and subgrade combinations. The range of input variables are as Considering the wide range of input values for each parameter,
follows: the total number of AASHTOWare Pavement ME Design runs
1. Average Annual Daily Truck Traffic (AADTT): 1,000, was equal to 55,800 for flexible pavements and 23,800 for rigid
4,000, 8,000, 12,000, 16,000, 20,000. pavements. These scenarios were run by creating projects for
2. AC thickness (inch (cm)): 4 (10.2), 6 (15.2), 8 (20.3), 10 different climatic zones, base layer, subbase layer, and subgrade
(25.4), 12 (30.5), 14 (35.6), 16 (40.6). types using the AASHTOWare Pavement ME Design software.
3. JPCP thickness (inch (cm)): 6 (15.2), 8 (20.3), 9 (22.9), 10 An example of project data scenario matrix is given in Fig. 2(a).
(25.4), 12 (30.5), 14 (35.6). Obviously, this was a lengthy process, and it took about six
4. Base thickness (inch (cm)) without subbase: 6 (15.2), 8 months to execute all matrix scenarios.
(20.3), 10 (25.4), 12 (30.5), 14 (35.6), 16 (40.6), 18 (45.7). After analysing a large matrix of pavement design/analysis
5. Base and subgrade combinations with default modulus scenarios based on AASHTOWare Pavement ME Design input/
values in AASHTOWare Pavement ME Design software output files, the results were exported into a Microsoft Excel
representing AASHTO A-1-a & A-1-a, A-3 & A-1-a, A-3 worksheet using a software developed by the research team -- an
& A-7-6, River-run Gravel & A-7-6, and River-run Gravel & example of this process is shown in Fig. 2(b). This data export

Fig. 2. Preparing Database for the KYPDT-M: (a) Example of Scenario Analysis in AASHTOWare 2.5.5, (b) Program to Export AASHTOWare Results into
Excel File
 KSCE Journal of Civil Engineering 211

software automatically gathers inputs and outputs from different varies from 0.07 inches (0.2 cm) to 0.15 inches (0.4 cm) for rigid
AASHTOWare Pavement ME Design runs into a worksheet, (Fig. 4).
which significantly reduced the needed time to prepare the Further, as shown in the top three plots in Fig. 3, the predicted
massive database for the development of the KYPDT-M program. pavement distresses of flexible pavement, as reported by the
AASHTOWare Pavement ME Design runs, decrease as the AC
4. Data Observations and Analysis thickness increases for a given traffic volume. However, as observed
from plots related to base or subbase layers in Fig. 3, the
Once the massive database was prepared, a preliminary analysis predicted pavement distresses of flexible pavement show very
was performed to better understand the relationship between little variation as the thickness of the base or subbase layer is
inputs and outputs of the AASHTOWare Pavement ME Design changed for a given traffic volume. For example, as the AC layer
runs for flexible and rigid pavements. Fig. 3 shows example thickness increases from 6 inches (15.2 cm) to 12 inches (30.5
inputs (AADTT, thickness of AC, base, and subbase layers) and cm) for AADTT 20000, the pavement total rut varies from 0.62
outputs (rut of pavement structure, top-down fatigue cracking, inches (1.6 cm) to 0.43 inches (1.1 cm). However, when the
and IRI) from the AASHTOWare Pavement ME Design runs for subbase layer thickness increases from 6 inches (15.2 cm) to 12
20 years of flexible pavement design life with base and subbase inches (30.5 cm) for AADTT 20000, the pavement total rut only
layers in Atlanta, GA. Fig. 4 shows example inputs (AADTT, changes from 0.56 inches (1.4 cm) to 0.55 inches (1.4 cm). This
thickness of PCC and base layers) and outputs (pavement faulting, indicates that the distress models in the AASHTOWare Pavement
cracking, and IRI) from the AASHTOWare Pavement ME Design ME Design for flexible pavement are sensitive to thickness changes
runs for 30 years of rigid pavement design life with dowel bar in in the AC layer; however, they appear to be less sensitive to
Atlanta, GA. As shown in Figs. 3 and 4, the AASHTOWare thickness changes in the base or subbase layers.
Pavement ME Design predicts higher distresses when AADTT Similarly, when studying the plots for the PCC and base layer
increases for flexible pavement as well as rigid pavement for a given thicknesses in Fig. 4, the predicted pavement distresses for rigid
layer thickness. For example, as AADTT increases from 4,000 to pavement show a clear decline as the PCC slab thickness
20,000, the pavement total rut predicted by the AASHTOWare increases. However, very minor changes in distress occur as the
Pavement ME Design changes from 0.54 inches (1.4 cm) to 0.71 base layer thickness changes for a given traffic volume. For
inches (1.8 cm) for flexible pavement (Fig. 3), and the pavement example, for AADTT 20000, the pavement faulting changes
faulting predicted by the AASHTOWare Pavement ME Design from 0.15 inches (0.38 cm) to 0.08 inches (0.2 cm) when PCC

Fig. 3. Example Inputs and Outputs from AASHTOWare Runs for Flexible Pavements
212 G. Yang et al.

Fig. 4. Example Inputs and Outputs from AASHTOWare Runs for Rigid Pavements

slab thickness changes from 6 inches (15.2 cm) to 14 inches which resulted in the development of the Multi RF models.
(35.6 cm). On the other hand, the pavement faulting changes When training each Multi RF model, the objective is to generate
from 0.27 inches (0.69 cm) to 0.24 inches (0.61 cm) as base a model to predict multiple outputs as follows
thickness increases from 6 inches (15.2 cm) to 18 inches (45.7
Y = { yˆ1 ∈ Y1, yˆ2 ∈ Y2, …, yˆn ∈ Y1n } (1)
cm). Hence, the distress models in AASHTOWare Pavement
ME Design 2.5.5 for rigid pavement are sensitive to thickness where Y is the final Multi RF model, yˆn is the prediction values
changes in the PCC slab layer; however, they appear to be less from the nth RF model, and Yn is the nth RF model.
sensitive to thickness changes in the base layer. Therefore, the developed Multi RF model contains an RF
model for each output. An RF model builds various decision
5. Multi-Output Random Forests Models trees by randomly selecting different data from the prepared
dataset and combines the results of these different trees to
Eight Multi RF models were developed to consider different achieve an accurate and stable prediction. Each tree contains root
scenarios, including 20- or 30-years design life, base or subbase node, internal nodes, and leaf nodes. For each RF model in this
layers for AC, with or without dowel bar for PCC, for pavement study, the number of trees is 100 without a depth limit so that the
design (estimating pavement layer thicknesses) and pavement nodes in the tree are expanded until all leaves contain less than 2
analysis (predicting pavement distresses) in the KYPDT-M: samples. Further, the mean squared error (MSE) of the out-of-
1. Model 1: 20 years AC with a base layer. bag (OOB) predictions is calculated as follows to evaluate the
2. Model 2: 20 years AC with base and subbase layers. accuracy performance of the RF model (González et al., 2016).
3. Model 3: 20 years PCC with a base layer.
1 n 2
4. Model 4: 20 years PCC including dowel bars and a base MSE = --- ∑ i = 1 ( yi – ŷiOOB ) (2)
n
layer.
5. Model 5: 30 years AC with a base layer. where n is the sample size, yi is the actual value of the observation,
6. Model 6: 30 years AC with base and subbase layers. and ŷiOOB is the average prediction for the ith observation from all
7. Model 7: 30 years PCC with a base layer. trees for which this observation has been OOB.
8. Model 8: 30 years PCC including dowel bars and a base Tables 3 and 4 summarize the verification results of different
layer. Multi RF models for pavement design and analysis in the KYPDT-
In total, there were 12,600, 43,200, 11,200, and 12,600 datasets M. Each Multi RF model consists of fitting one RF regressor per
prepared for the development of Model 1&5, Model 2&6, Model target, and the R-squared values of the Multi RF model as well as
3&7, and Model 4&8, respectively. Approximately 80% and 20% each individual RF regressor are listed in Tables 3 and 4. Obviously,
of the data were randomly selected for model training and testing, the higher the R-squared value, the better the Multi RF model or
 KSCE Journal of Civil Engineering 213

Table 3. Validation Testing Results for Multi RF Models and Each Layer for Pavement ME Design for base or subbase layer only for Model
Pavement Design 2, the R-squared value is 0.69 or 0.30. It means that the Multi RF
R-Squared Values model has better predictive performance when estimating the
thickness for AC layer as compared to the thickness for base or
Multi RF Each RF Model
Models Multi RF subbase layers. This is because distress models in the AASHTOWare
Model AC/PCC Base Subbase Pavement ME Design 2.5.5 for flexible pavement are more
Thickness Thickness Thickness
sensitive to thickness changes in the AC layer; however, they
1 0.76 0.99 0.59 N/A appear to be less sensitive to thickness changes in the base or subbase
2 0.66 0.99 0.69 0.30 layer, as shown in Fig. 3. Further, similar results are observed from
3 0.63 1.00 0.27 N/A other models in Table 3: the R-squared value of RF regressor for
4 0.66 1.00 0.37 N/A AC/PCC layer thickness estimation is always larger than 0.9,
5 0.73 0.97 0.55 N/A whereas those for base and subbase layers are relatively low (less
6 0.64 0.99 0.71 0.24 than 0.75). This indicates the Multi RF is very effective and
7 0.62 1.00 0.26 N/A accurate when predicting the AC/PCC layer thickness, but not as
8 0.67 1.00 0.40 N/A effective for base or subbase layers. Obviously, this is a carryover
effect from the AASHTOWare Pavement ME Design models
with the same trend tendency.
individual RF regressor. Figure 5 displays thickness prediction of Model 2 and Model 8 as
examples to demonstrate the capability of the Multi RF model
5.1 Pavement Design Verification for thickness prediction of pavement layers. In each figure, the
For pavement design, each Multi RF model was trained to estimate predicted thickness from the Multi RF model is compared against
the thicknesses of the AC/PCC layer, base layer, or subbase layer the thickness values used in the AASHTOWare Pavement ME
based on the traffic volume, climate zones, materials properties Design analysis runs. The more data points fall along the best
of each pavement layer, and the threshold pavement distress fitting line in each plot, the better the Multi RF predicts the layer
levels at the end of the design life. There were six distresses for thickness. Fig. 5 indicates that the Multi RF offers very accurate
AC pavements and three distresses for PCC pavements. Take the predictions for the AC/PCC layer thickness as compared to the
results of Model 2 in Table 3 as an example: the R-squared value predictions for the base or subbase layers. Table 5 summarizes
of the Multi RF model is 0.66 while the R-squared values of RF the coefficients of best fitting lines in Fig. 5 for AC and PCC
regressor are 0.99, 0.69, and 0.30 for AC, base, and subbase pavement thickness prediction. Again, this is a carryover effect
layer thickness predictions, respectively. This finding has the from the AASHTOWare Pavement ME Design models. The
following two implications: authors would like to note that the need for base/subbase in a
1. The R-squared value is 0.66 when comparing the predicted pavement goes beyond structural issues. In addition to the need
thicknesses from the Multi RF model to the thicknesses for adequate and uniform base/subbase support, the pavement
used as inputs in AASHTOWare Pavement ME Design for needs for proper drainage and frost protection.
AC, base, and subbase layers using Model 2;
2. The R-squared value is 0.99 when comparing the predicted 5.2 Pavement Analysis Verification
thicknesses from the Multi RF model to the thicknesses For pavement analysis, each Multi RF model was trained to
used as inputs in AASHTOWare Pavement ME Design predict pavement distresses at the end of design life based on the
analysis for AC layer only using Model 2. traffic volume, weather condition, materials properties of each
When comparing the predicted thicknesses from the Multi RF pavement layer, and thicknesses of AC/PCC layer, base layer, or
model to the thicknesses used as inputs in AASHTOWare subbase layer. Take the results of Model 2 in Table 4 as an example:

Table 4. Validation Testing Results for Multi RF Models and Each Distress for Pavement Analysis
Multi RF Models 1 2 3 4 5 6 7 8
R-Squared Values Multi RF Model 0.99 0.98 1.00 1.00 0.99 0.98 1.00 1.00
IRI 0.99 0.99 1.00 1.00 0.99 0.99 1.00 1.00
Rut-AC 0.99 0.98 N/A N/A 1.00 0.94 N/A N/A
Rut-Total 1.00 0.99 N/A N/A 0.99 0.98 N/A N/A
Bottom-Up Fatigue Cracking 0.99 1.00 N/A N/A 0.98 1.00 N/A N/A
Thermal Cracking 1.00 1.00 N/A N/A 1.00 1.00 N/A N/A
Top-Down Fatigue Cracking 0.97 0.95 N/A N/A 0.98 0.97 N/A N/A
Faulting N/A N/A 1.00 1.00 N/A N/A 1.00 1.00
Cracking N/A N/A 1.00 1.00 N/A N/A 1.00 1.00
214 G. Yang et al.

Fig. 5. Thickness Prediction of Models 2 and 8: (a) Model 2: AC Thickness (in.), (b) Model 8: PCC Thickness (in.), (c) Model 2: Base Thickness
(in.), (d) Model 8: Base Thickness (in.), (e) Model 2: Subbase Thickness (in.)

the R-squared value of the Multi RF model is 0.98 while the R- AASHTOWare Pavement ME Design runs. As displayed in Fig. 6,
squared values of RF regressor are 0.99, 0.98, 0.99, 1.00, 1.00, most of the data points located along the best fitting line, which
and 0.95 for IRI, rut depth of AC layer, rut depth of pavement indicates that the predicted distresses from the Multi RF model
structure, bottom-up fatigue cracking, thermal cracking, and top- are very similar to those from the AASHTOWare Pavement ME
down fatigue cracking, respectively. Further, similar results are Design runs. Table 5 summarizes the coefficients of best fitting
observed from other models in Table 4: the R-squared values of lines in Fig. 6 for AC and PCC pavement distress prediction.
the Multi RF model and each RF regressor for pavement distresses Therefore, the Multi RF models offer good prediction performance
are always larger than 0.9. Thus, it implies that the Multi RF for pavement distresses.
model accurately predicts pavement distresses based on the
given inputs. 6. Conclusions
Figure 6 displays the distress prediction of Model 2 and Model 8
as examples. In each figure, the estimated distress from the Multi This paper presents the machine learning based tool to
RF model is compared against the distress predictions from the simultaneously predict multiple pavement distresses (for pavement
 KSCE Journal of Civil Engineering 215

Fig. 6. Distress Prediction of Models 2 and 8: (a) IRI (in./mi.), (b) IRI (in./mi.), (c) AC Rut (in.), (d) Faulting (in.), (e) Bottom-Up Cracking (% lane
area.), (f) Cracking (% slabs)

analysis) and estimate thicknesses of AC or PCC layer (for pavement conduct many trial-and-error runs.
design). The inputs and outputs of 79,600 pavement design scenarios It is worth mentioning that the KYPDT-M is a user-friendly
(55,800 for flexible pavement and 23,800 for rigid pavement) via pavement design tool that is fully compatible with the AASHTOWare
the AASHTOWare Pavement ME Design 2.5.5 software considering Pavement ME Design and it plays a support role. Furthermore, the
various climate regions, traffic volume, and pavement structures KYPDT-M in no way should be viewed as a replacement for the
were used to develop the Multi RF models. The analysis results AASHTOWare Pavement ME Design software. Additionally, if
demonstrated that Multi RF models offer good performance for possible, the solutions proposed by the KYPDT-M should be
predicting pavement distresses and estimating thickness for AC further verified and fine-tuned by employing the AASHTOWare
or PCC layers. It was demonstrated that the machine learning Pavement ME Design software, as well as using local calibration
based KYPDT-M would be an effective tool for pavement engineers, data. Moreover, as a free online tool, the KYPDT-M will benefit
particularly those engineers working for organizations with transportation agencies, especially local agencies with limited
limited resources. The KYPDT-M works efficiently and accurately resources, to arrive at a reasonable estimate of pavement design and
to produce a pavement design estimate without having to analysis in an efficient and accurate manner before they decide to
216 G. Yang et al.

Table 5. Coefficients of Best Fitting Lines for Pavement Thickness Prediction multi-output regression. Wiley Interdisciplinary Review: Data Mining
(Fig. 5) and Distress Prediction (Fig. 6) and Knowledge Discovery.5(5):216-233, DOI: 10.1002/widm.1157
Coefficients Federal Highway Administration (FHWA) (2017) Highway materials
Pavement
Prediction Indicators engineering course. Model E - Mechanistic-empirical design.
Type a b Washington, DC, USA
Distress AC IRI (in./mi) 0.99 1.96 Gong H, Sun Y, Hu W, Polaczyk P, Huang B (2019) Investigating
AC-Rut (in.) 0.98 0.01 impacts of asphalt mixture properties on pavement performance
using LTPP data through random forests. Construction and Building
Bottom-Up Cracking 1 0.11
Materials 204:203-212
(% lane area)
Gong H, Sun Y, Shu X, Huang B (2018) Use of random forests regression
PCC IRI (in./mi) 1 -0.01 for predicting IRI of asphalt pavements. Construction and Building
Faulting (in.) 1 0 Materials 189:890-897
Cracking (% slabs) 1 -0.01 González C, Mira J, Ojeda JA (2016) Applying multi-output random
Thickness AC AC Layer (in.) 0.99 0.09 forest models to electricity price forecast. Preprints, DOI: 10.20944/
PREPRINTS201609.0053.V1
Base Layer (in.) 0.57 3.94
Graves RC (2012) An Innovative approach to mechanistic empirical
Subbase Layer (in.) 0.27 8.81 pavement design. PhD Thesis. University of Kentucky, Lexington, KY,
PCC PCC Layer (in.) 1 0.01 USA
Base Layer (in.) 0.23 8.89 Graves RC, Mahboub KC (2012) Streamlining use of mechanistic-
Note: the equation of the best fitting line in Figs. 5 and 6 is empirical pavement design guide. Transportation Research Record:
Y ( Multi RF Model ) = a*Y ( M – E PDT ) + b Journal of the Transportation Research Board 2305(1):170-176, DOI:
10.3141/2305-18
Hoffman K, Sung JY, Zazzera A (2021) Multi-output random forest
regression to emulate the earliest stages of planet formation.
fine-tune their decisions via the AASHTOWare Pavement ME arXiv:2104.12845v1, DOI: 10.48550/arXiv.2104.12845
Design software. Iowa Statewide Urban Design and Specifications (Iowa SUDAS)
(2021) Design manual. Iowa State University, Ames, IA, USA
Acknowledgments Li J, Uhlmeyer JS, Mahoney JP, Muench ST (2011) Use of the 1993
AASHTO guide, MEPDG and Historical performance to update the
Not Applicable WSDOT pavement design catalog. WA-RD 779.1, Washington State
Department of Transportation, Tumwater, Washington, Washington
DC, USA
ORCID Linusson H (2013) Multi-output random forests. University of Boras
(Högskolan I Borás), VT 2013: MAGI04
Guangwei Yang https://orcid.org/0000-0002-0870-2440 Ohio Department of Transportation (ODOT) (2015) Pavement design
Kamyar C. Mahboub https://orcid.org/0000-0003-1668-4585 manual. Columbus, OH, USA
R. Clark Graves https://orcid.org/0000-0002-2366-7815 Pavement Guide Interactive (PGI) (2003) Pavement tools consortium
Kelvin C.P. Wang https://orcid.org/0000-0002-4832-2848 (PTC), Washington state department of transportation (WSDOT)
Pavement Guide, University of Washington, Seattle, Washington.
http://training.ce.washington.edu/WSDOT/Modules/02_pavement_
References types/02-5_body.htm#joints
Santana EJ, Santos FR, Mastelini SM, Melquiades FL, Jr SB (2020)
Aguiar GJ, Santana EJ, Mastelini SM, Mantovani RG, Jr SB (2019) Improved prediction of soil properties with multi-target stacked
Towards meta-learning for multi-target regression problems. arXiv: generalisation on EDXRF Spectra. arXiv:2002.04312v1, DOI:
1907.11277v1, DOI: 10.48550/arXiv.1907.11277 10.48550/arXiv.2002.04312
American Association of State Highway and Transportation Officials Schwartz CW, Carvalho RL (2007) Implementation of the NCHRP 1-
(AASHTO) (1993) Guide for design of pavement structures – 1993. 37A design guide, Volume 2: Evaluation of mechanistic-empirical
National Cooperative Highway Research Program (NCHRP), design procedure. MDSHA Project No. SP0077B41, The University of
Washington, DC, USA Maryland, College Park, MD
American Association of State Highway and Transportation Officials Schwartz CW, Li R, Kim SH, Ceylan H, Gopalakrishnan K (2011)
(AASHTO) (2015) Mechanistic-empirical pavement design guide Sensitivity evaluation of MEPDG performance prediction. NCHRP
(A manual of practice). Washington, DC, USA 1-47. Washington, DC: The National Academies Press
Applied Research Associates Inc. (ARA) (2004) Guide for mechanistic- Wu Z, Groeger JL, Simpson AL, Hicks GR (2010) Performance evaluation
empirical design of new and rehabilitated pavement structures. of various rehabilitation and preservation treatments. FHWA-HIF-
NCHRP 1-37A, National Cooperative Highway Research Program 10-020. Federal Highway Management, Washington, DC.
(NCHRP), Washington, DC, USA Xi X, Sheng VS, Sun B, Wang L, Hu F (2018) An empirical comparison
Asphalt Institute (AI) (2020) US State binder specifications, Retrieved on multi-target regression learning. CMC-Computers, Materials &
February 15, 2020, http://www.asphaltinstitute.org/engineering/ Continua 56(2):185-198, DOI: 10.3970/cmc.2018.03694
specification-databases/us-state-binder-specifications/ Xu D, Shi Y, Tsang IW, Ong YS, Gong C, Shen X (2019) A survey on
Borchani H, Varando G, Bielza C, Larrañaga P (2015) A survey on multi-output learning. arXiv:1901.00248, DOI: 10.48550/arXiv.1901.
 KSCE Journal of Civil Engineering 217

00248 Zhang Z, Leidy JP, Kawa I, Hudson WR (2000) Impact of changing traffic
Yang G, Yu W, Li Q, Wang K, Peng Y, Zhang A (2019) Random forest- characteristics and environmental conditions on flexible pavement.
based pavement surface friction prediction using high-resolution 3d Transportation Research Record: Journal of the Transportation Research
image data. Journal of Testing and Evaluation, DOI: 10.1520/ Board, No. 1730, Washington, DC, 125-131
JTE20180937