Explainable ensemble machine learning framework for pavement concrete compressive strength and chloride ion penetration prediction

Pavement concrete is a core element of transport infrastructure, but many property prediction models rely on general concrete datasets that omit key pavement mixture parameters. Although explainable machine learning has improved model interpretability, most studies lack external verification against an independent experimental program. This study develops an externally verified, pavement-specific prediction framework using targeted features, including air content and slump, and engineered variables. A database of 1430 compressive strength and 246 chloride ion penetration (CIP) mixtures was compiled from technical reports and peer-reviewed studies. Gradient boosting models (CatBoost, HistGradientBoosting, and XGBoost) and their voting and stacking ensembles were compared with artificial neural network (ANN) benchmarks. The selected VotingRegressor and CatBoostRegressor achieved R² values of 0.962 and 0.820 for compressive strength and CIP, respectively, compared with 0.905 and 0.793 for the corresponding ANNs. On unseen experimental data, both models retained approximately 97% of their test-set R². SHapley Additive exPlanations analysis showed physically consistent trends for aggregate characteristics, air content, and supplementary cementitious materials. The relatively low contribution of the water-to-binder ratio reflected its restricted pavement range rather than a model limitation. Overall, the framework provides interpretable predictions within the pavement-concrete domain represented by the database.