Term of Award

Spring 2023

Degree Name

Master of Science, Civil Engineering

Document Type and Release Option

Thesis (open access)

Copyright Statement / License for Reuse

Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.

Department

Department of Civil Engineering and Construction

Committee Chair

M. Myung Jeong

Committee Member 1

Soonkie Nam

Committee Member 2

Xiaoming Yang

Abstract

The complex dynamic modulus (|E*|) is a characterization property that defines the stiffness of an asphalt mixture. The dynamic modulus can be found through lab testing or predictions. Since lab testing can be time-consuming and expensive, the prediction method can be used as an alternative method. While a statistical method has been traditionally used for the |E*| prediction such as the Witczak’s predictive equations, machine learning (ML) is recently emerging as an alternative way that |E*| predictions can be made. This research attempted to predict the |E*| using several ML techniques including linear regression, support vector machines (SVM), decision trees, random forest, and deep learning.

This research includes 3906 laboratory-measured |E*| data points that come from a variety of asphalt mixtures. In the database, there is a group of conventional materials, but most of the data comes from non-conventional materials. These non-conventional materials include reclaimed asphalt pavement (RAP), recycled asphalt shingles (RAS), warm mix asphalt (WMA), asphalt rubber, air-blown asphalt, and polymer-modified. The following evaluation metrics are used to evaluate the results from ML: mean absolute error, mean squared error, root mean squared error, and explained variance score.

In this research, two comparisons were made to answer the following questions: 1) Which ML technique would provide a better prediction for |E*|? and 2) Between ML and Witczak’s predictive method, which would provide a better prediction for |E*|? It was concluded that decision trees and random forests had the best results, and linear regression results needed the most improvement. When comparing the results of the ML methods (based on the R² value), it was found that the results of decision trees, random forest, and deep learning outperformed the 1999 and 2006 Witczak's predictive equations. However, the 1999 and 2006 Witczak's predictive equations outperformed the linear regression model (based on the R² value). The 1999 Witczak's predictive equation outperformed SVM. The results for SVM and the 2006 Witczak predictive equation were close, and it appeared that the SVM may be the better method. The 1999 Witczak's predictive equation outperformed SVM.

Research Data and Supplementary Material

No

COinS