Performance Evaluation of Three-Phase Conductive Asphalt Concrete and Machine Learning–Driven Damage Prediction

Abstract

This study aims to develop a high-performance three-phase conductive asphalt concrete (TP-CAC) composed of steel slag, graphite, and carbon fiber, and build a machine learning model for predicting damage from electrical resistance changes in beam specimens. A complete specimen preparation process was first designed to address carbon fiber agglomeration, followed by electrical resistance measurement and mechanical tests (Marshall stability, rutting, and low-temperature bending) to determine the optimal mix ratio. Using fracture energy ( G f ) as a damage indicator, gradient boosting models [CatBoost, Light gradient boosting machine (GBM), extreme gradient boosting (XGBoost)] and their ensemble model were built to analyze the relationship between electrical resistance and G f , with CatBoost selected as the optimal predictive model. The results showed that TP-CAC with 50% steel slag (by volume), 30% graphite (by mass), and 0.2% carbon fiber (by mass) exhibits excellent smart properties (resistance-sensitive to damage) and mechanical performance. All four models achieved a coefficient of determination ( R 2 ) > 0.9 , confirming the feasibility of monitoring damage via resistance changes. Although the initial root-mean square error (RMSE) values were relatively high, feature engineering (sliding window extraction) and hyperparameter optimization (Optuna) improved the CatBoost model to R 2 = 0.97 , with a 25% reduction in RMSE. This study integrates experimental optimization with data-driven modeling to advance smart materials.