Predicting the fracture load of asphalt concrete under TPB test using POA-optimized machine learning methods

Abstract

Fracture load P_f is a critical parameter for evaluating the fracture behavior of asphalt concrete (AC). However, traditional experimental methods for determining P_f are time-consuming and costly. Thus, this study employs machine learning (ML) to predict the P_f of AC under three-point bending (TPB) test based on aggregate gradation, specimen dimensions, porosity, and temperature. The Pelican Optimization Algorithm (POA) is utilized to optimize the hyperparameters of the Random Forest Regressor (RFR), Multi-Layer Perceptron (MLP), Generalized Additive Model (GAM), and LSBoost. Model performance is compared through error analysis, and the effective fracture resistance (K_eff) computed with the optimized models is evaluated against predictions by the traditional Maximum Tangential Stress (MTS) and Maximum Tangential Strain (MTSN) fracture criteria. The results suggest that conventional ML algorithms tend to exhibit lower predictive accuracy, weaker correlation, and signs of overfitting. By applying POA optimization, all four models show improvements in predictive accuracy, R² values, and a notable reduction in overfitting. Among the optimized models, LSBoost-POA demonstrates the highest predictive accuracy and robustness. Under Mixed-Mode I/II loading conditions, the K_eff predictive performance of RFR-POA and LSBoost-POA significantly outperforms that of MTS and MTSN, with LSBoost-POA achieving the best results. SHapley Additive exPlanations (SHAP) analysis reveals that for LSBoost-POA, the specimen size factor λ is the most influential parameter in P_f prediction, while the specimen height H has the least impact. These insights hold promising implications for applying machine learning to the study of fracture behavior in AC.