Freeze-thaw (F-T) damage, characterized by microcrack propagation and material degradation caused by repeated water-ice phase transitions, is a critical factor leading to durability degradation of concrete structures in cold regions. Fly ash (FA) as an industrial byproduct is widely used to replace cement, reducing carbon emissions, while enhancing mechanical properties. However, conventional prediction methods for freeze-thaw resistance of fly ash concrete (FAC), which rely on the empirical formulas or simplified mechanical models, often exhibit limitations in efficiency and accuracy. These limitations stem from their inability to capture the complex nonlinear interactions among multifactorial parameters such as material composition, environmental conditions, and loading histories. As modern engineering increasingly demands high-precision durability assessments for infrastructure lifespan prediction, there is an urgent need to develop advanced predictive tools. This study was to propose an artificial intelligence (AI)-based framework to predict F-T resistance (i.e., mass loss rate, MLR; relative dynamic modulus of elasticity, RDME) with high precision and interpretability, addressing a critical deficiency in material design and durability assessment.

A comprehensive database of 537 experimental datasets was constructed, including 14 input variables (i.e., cement content, FA dosage, water-to-binder ratio (W/B), F-T cycles (NFT)). To address the curse of mitigate multicollinearity, a hybrid feature selection strategy combining the Maximum Information Coefficient (MIC) and Lasso regression was implemented. The MIC quantified nonlinear correlations between input variables and target responses (i.e., mass loss rate, MLR, and relative dynamic elastic modulus, RDME), while the Lasso regression performed variable subset selection through L1 regularization. Three machine learning algorithms-Random Forest (RF), Support Vector Regression (SVR), and Extreme Gradient Boosting (XGBoost) were subsequently employed to develop prediction models. Hyperparameter optimization for each algorithm was conducted via the Bayesian optimization with 5-fold cross-validation to prevent overfitting. The model performance was rigorously evaluated using multiple metrics, including coefficient of determination (R²), mean absolute error (MAE), and root mean squared error (RMSE). Furthermore, the SHapley Additive exPlanations (SHAP) values and Partial Dependence Plots (PDP) were utilized to dissect the contribution hierarchy and nonlinear marginal effects of critical parameters on the FAC freeze-thaw performance, thereby enhancing the model interpretability.

The XGBoost-T model (with feature selection) achieves the a maximum accuracy for MLR (R²=0.9224) and RDME (R²=0.9127), outperforming XGBoost model by 9.5% and 4.7%, respectively. The feature selection reduces a mean absolute error (MAE) by 26.1% (MLR) and 62.6% (RDME), proving its necessity. The study also analyzes the contributions and nonlinear effects of key parameters by the Shapley Additive Explanations (SHAP) and Partial Dependence Plots (PDP). The findings indicate that cement content and total binder content increase can reduce MLR and improve RDME, while water-to-binder ratio and freeze-thaw cycles have the opposite effects. The PDP analysis provides recommendations for optimizing FAC mix designs in cold regions, such as a water content range of 130-180 kg/m³ and a sand content range of 550-700 kg/m³ for optimal RDME.

This research introduced an AI-driven framework for predicting the freeze-thaw resistance of FAC, integrating feature engineering, intelligent optimization, and explainable machine learning. The framework could overcome the limitations of conventional empirical models via leveraging data-driven insights and high-dimensional nonlinear modeling. The XGBoost-based prediction model demonstrated state-of-the-art accuracy and robustness, while the PDP-derived parameter sensitivity analysis offered actionable strategies for enhancing FAC durability in cold environments. This study could contribute to the theoretical advancement of FAC durability design and underscore the transformative potential of artificial intelligence in civil engineering materials research. Future work could focus on expanding the database to include emerging cementitious materials and integrating multi-physics simulations for holistic durability assessments.