A Metaheuristic Approach of predicting the Dynamic Modulus in Asphalt Concrete

Abstract

The prediction of the asphalt dynamic modulus (E*), which measures the material's ability to withstand changes in shape or structure, is important. Previous studies indicated that the well-known Witczak 1-40D model for E* is outperformed by machine learning models. Additionally, the application of machine learning algorithms requires manual fine-tuning of their hyperparameters. In this study, the artificial Hummingbird and Harris Hawks optimization algorithms were employed in the automatic calibration of the Random Forest and Gradient Boost algorithms' hyperparameters for modeling E* using the Witczak 1-40D model and additional parameters. In addition, the model was interpreted using the Shapley value and permutation feature importance. The results indicate that the optimized artificial hummingbird algorithm model performed better, with R² reaching 0.97. The interpretability of the model suggests that the binder parameters exhibited the highest effect on the variance of E*.