2D Aggregate Gradation Conversion Framework Integrated with 3D Random Aggregate Method and Machine-Learning for Asphalt Concrete

Xin Wei

doi:10.1061/JMCEE7.MTENG-17430

2024-05-01

PRE

DOI：10.1061/JMCEE7.MTENG-17430

原文链接

原文求助

摘要

En 中文

Two-dimensional (2D) random mesostructure models have been broadly utilized to study mechanical behavior and damage mechanisms of asphalt concrete due to their high availability and efficiency. The 2D aggregate gradation that characterizes the area fractions of the simulated aggregates of different sizes significantly affects the simulated mesomechanical properties of asphalt concrete. However, there are very few methods for determining the 2D aggregate gradation from the actual three-dimensional (3D) one. This study proposed a 2D aggregate gradation conversion framework integrating a 3D random aggregate method and machine learning models. First, 3D asphalt concrete mesostructures were generated using polyhedral coarse aggregates with random gradations and shapes. Virtual cutting and sieving procedures were then developed to obtain the 2D aggregates at sections with 1 mm intervals in each mesostructure and acquire the area fractions of aggregates of different sizes, respectively. By averaging the results for the sections in each cutting direction, so-called 2D representative area fractions of aggregates within different grading segments in different directions were obtained. The gradation conversion data sets were established by collecting the volume fractions of coarse aggregates in the 3D mesostructures as input variables and the corresponding 2D representative aggregate area fractions as output variables. Based on the data sets, machine learning models, including linear regression, support vector regression, bagging of regression tree, and neural network (NN), were trained using a 5-fold cross-validation approach and then tested. The results showed that the NN model generally provided the most accurate predictions of the 2D representative aggregate area fractions. The effectiveness of the proposed framework was validated by using both the test sets regarding different sizes of random mesostructures and a mesostructure rebuilt from computed tomography (CT) images of actual asphalt concrete.

Keyword:

Asphalt concrete

Mesostructure

Machine learning

Aggregate gradation

Random aggregate method