Rapid Abrasion-Resistance Prediction of Recycled Aggregates Using Improved Whale Optimization-Tuned Gaussian Process Regression and SHAP Analysis

Abstract

High Friction Surface Treatment (HFST) relies heavily on wear-resistant aggregates to ensure roadway safety, yet the conventional evaluation of aggregate abrasion resistance is time-consuming and resource-intensive. In this study, a machine learning framework was developed to predict the abrasion-induced angularity evolution of recycled high-alumina aggregates from their initial morphological characteristics, thereby enabling rapid abrasion-resistance screening. Six regression models were compared under leave-one-group-out cross-validation, and an improved whale optimization algorithm (IWOA) was proposed to tune the Gaussian process regression (GPR) model, incorporating five enhancements and a regularized fitness function to restrain overfitting. The models were trained on 42 samples from six aggregates, whose angularity, Form 2D, micro-texture, sphericity, and F:E ratio were measured with the AIMS II device before and after successive abrasion cycles. The IWOA-GPR model achieved the best performance, with an R2 of 0.8909, an RMSE of 150.98, an MAE of 120.55, and a MAPE of 4.60%. The SHAP analysis identified the abrasion revolutions, the initial Form 2D, and the initial angularity as the dominant contributors to the worn angularity. Moreover, the early angularity loss after the first 500 revolutions correlated strongly with the measured Los Angeles abrasion value (r = 0.935), which allows the LAA of a candidate aggregate to be estimated after a single abrasion cycle. The proposed framework therefore provides a rapid and reliable tool for screening wear-resistant aggregates for HFST applications and supports the clean utilization of recycled solid wastes in anti-skid pavements.