The application of explainable artificial intelligence (XAI) in civil engineering has garnered increasing attention due to its ability to enhance transparency in machine learning (ML) models for material behavior prediction. However, the comparative performance and interpretability of various XAI models in predicting concrete properties under freeze–thaw degradation remain underexplored. This study presents a comprehensive evaluation of five XAI approaches: GAMI-Net, Explainable Boosting Machine (EBM), SHapley Additive exPlanations (SHAP), Local Interpretable Model-agnostic Explanations (LIME), and Shapash, for predicting the elastic dynamic modulus of concrete subjected to freeze–thaw cycles. Using a dataset comprising 347 concrete test records, three predictive models (GAMI-Net, EBM, and XGBoost) were developed, with post-hoc explanations applied to XGBoost outputs. Model performance was assessed through accuracy metrics and interpretability analyses, including feature importance ranking, global and local explanation visualizations, and dependence plots. The results reveal that GAMI-Net achieved the best overall performance (RMSE = 0.04, R² = 0.95) and provided the most stable and physically consistent interpretations. Freeze–thaw cycles and water content were identified as the most critical factors, with their interaction accounting for the greatest influence on concrete degradation. Two-dimensional heatmaps from GAMI-Net and EBM offered more intuitive interaction insights compared to SHAP-based visualizations. This study contributes theoretically by validating the efficacy of inherently interpretable models for materials science applications and practically by proposing GAMI-Net as a robust decision-support tool for frost-resistant concrete design. Future research should address data limitations, extend interpretability frameworks to time-dependent degradation, and explore uncertainty quantification for broader real-world deployment.