Interpretable machine learning model for predicting freeze-thaw damage of dune sand and fiber reinforced concrete

The freeze-thaw (F-T) properties of ordinary concrete have been extensively studied and related models were well-established. However, these models cannot be used to accurately assess the F-T damage of dune sand and basalt fiber reinforced concrete (DSFC), which is essential for performing targeted repairers. This study established an interpretable machine learning (ML) model to predict the F-T damage indicator D of DSFC through systematic comparisons of eight ML models (four classical models and four ensemble models). As an initial step, a dataset containing 257 samples was created by combining the F-T test results of this study and previous literature. This dataset was utilized to establish the eight models, which were then benchmarked against the polynomial models to select the optimal model. Subsequently, the selected model was interpreted using importance analysis, which included permutation feature importance (PFI) and partial dependency plots (PDP), to dissect the predictive mechanism of this model. Comparisons of eight models revealed that the four ensemble learning models generally outperforms the four classical learning models. Among the four ensemble learning models, XGBoost was verified to have optimal performance, with an R2 of 0.965 and RMSE of 0.019 on the test set. The importance analysis of XGBoost further demonstrated the inhibitory effect of dune sand and basalt fiber on F-T damage revealed in the experimental results. In summary, the proposed XGBoost along with importance analysis, is a practical methodology to predict and interpret the F-T damage of DSFC.