Tree-Based Ensemble Methods: Predicting Asphalt Mixture Dynamic Modulus for Flexible Pavement Design

The hot mix asphalt (HMA) dynamic modulus ( divide E* divide ), or stress/strain response measurement under dynamic loading, is considered the primary mechanical property input for flexible pavement in the Mechanistic-Empirical Pavement Design Guide (MEPDG). The current MEPDG software, AASHTOWare Pavement ME Design (PMED) Version 2.5, employs the Witczak equation for divide E* divide estimation when laboratory-measured divide E* divide data is unavailable. This study investigates the feasibility of developing alternative divide E* divide prediction models with modern machine learning techniques based on an established data library of Georgia HMA mixtures involving varying binder sources, binder grades, and nominal maximum aggregate sizes. Specifically, tree-based ensemble methods, including bagging, random forest, and gradient boosting, were applied considering their superior performance, and balanced versatility and interpretability. The results revealed that the gradient boosting model produced the most accurate predictions with an R-2 coefficient of determination of 0.982 as compared to the nationally calibrated Witczak model, which has an R-2 of 0.392, using the same test data set. The stark discrepancy in the divide E* divide predictions emphasizes the need of local calibration for the MEPDG models and the high potential for developing greatly improved models using modern machine learning techniques with the local data sources. Ultimately, this study provides an excellent example for other state highway agencies to follow to facilitate their MEPDG implementation.

Automated summary

This study investigated the feasibility of developing alternative prediction models for hot mix asphalt dynamic modulus using modern machine learning techniques, with the gradient boosting model producing the most accurate predictions. Local calibration for MEPDG models is necessary, and there is high potential for greatly improved models using modern machine learning techniques and local data sources.