Developing Hybrid Machine Learning Models to Determine the Dynamic Modulus (E*) of Asphalt Mixtures Using Parameters in Witczak 1-40D Model: A Comparative Study.

Wenjuan Xu,Mengmeng Zhou,Xin Huang,Zhengjun Yang,Jiandong Huang

doi:10.3390/ma15051791

Wenjuan Xu, Mengmeng Zhou + Show 3 more

Open Access

https://doi.org/10.3390/ma15051791

Copy DOI

Abstract

To characterize the dynamic modulus (E*) of the asphalt mixtures more accurately, a comparative study was shown in this paper, combining six ML models (BP, SVM, DT, RF, KNN, and LR) with the novelly developed MBAS (modified BAS, beetle antennae search) algorithm to check the potential to replace the empirical model. The hyperparameter tuning process of the six ML models by the proposed MBAS algorithm showed satisfactory results. The calculation and evaluation process demonstrated fast convergence and significantly lower values of RMSE for the five ML models (BP, SVM, DT, RF, and KNN) to determine the E* of the asphalt mixtures. Comparing the performances of the six ML models in the prediction of the E* by the statistical coefficients and Monte Carlo simulation, the RF model showed the highest accuracy, efficiency, and robustness.

Highlights

A crucial parameter that distinguishes the flexible pavement performance conditioning on multiple temperatures and loading circumstances in the Mechanistic-Empirical Pavement Design Guide (MEPDG) is the E* of asphalt mixtures [1–3]
To address the above-mentioned issues, six machine learning (ML) models, including the back-propagation (BP), decision tree (DT), k-nearest neighbors (KNN), logistic regression (LR), random forest (RF), and support vector machine (SVM), were investigated and compared to predict the E* of asphalt concrete in the present study. These ML algorithms were used to establish the relationships between the design parameters and E* of asphalt mixtures, and the beetle antennae search (BAS) algorithm was used to adjust the hyperparameters
During the calculation process of the traditional BAS algorithm, the step size is kept constant or decreasing in one iteration, which leads to the BAS algorithm falling into the state of local optimal

Summary

Introduction

A crucial parameter that distinguishes the flexible pavement performance conditioning on multiple temperatures and loading circumstances in the Mechanistic-Empirical Pavement Design Guide (MEPDG) is the E* (dynamic modulus) of asphalt mixtures [1–3]. It should be noted that the various ML techniques mentioned above were used in predicting E*, there were still some problems that need to be solved: (1) only a limited number of advanced ML algorithms have been used to predict E*, and the reliability and computational efficiency of other advanced ML algorithms, such as random forest (RF), have not been deeply studied in previous studies; (2) the application of ML algorithm always requires proper selection of the hyperparameters, but the optimization ability of beetle antennae search (BAS) for the hyperparameters has not been fully studied in the prediction of asphalt mixture; (3) the existing ML models may have great differences in the calculation results and efficiency when applied to predict the E* of the asphalt mixture, but there is still a lack of systematic and quantitative comparison of these algorithms

Research Objective

MBAS Algorithm

K-Fold Cross-Validation

Results of the Hyperparameter Tuning