Machine Learning Prediction and Interpretability Analysis of Peak Displacement of Concrete-Filled Steel Tube Members Under Lateral Impact

To address the difficulty in rapidly and accurately evaluating the peak displacement of concrete-filled steel tube （CFST） members under lateral impact， as well as the insufficient interpretation of the internal decision-making mechanism of existing data-driven methods， a peak-displacement database of CFST members was established based on published domestic and international lateral impact test data， and a machine learning model that balances predictive accuracy and interpretability was developed. The database consists of 193 valid test samples， covering geometric parameters， material properties， and impact conditions. Eight parameters， namely hammer mass， impact velocity， section diameter， steel tube thickness， axial compression ratio， steel yield strength， concrete compressive strength， and member length—were selected as input features， and the mid-span peak displacement was taken as the output feature. Based on the experimental database， six ensemble learning algorithms were employed to establish peak-displacement prediction models， and the Optuna framework combined with five-fold cross-validation was used for automatic hyperparameter optimization. Meanwhile， To enhance the transparency of the model results， Shapley Additive Explanations （SHAP） were employed to conduct an interpretability analysis of the CatBoost model， quantifying the contribution degree， influence direction， and potential interactions among variables of each input parameter on the predicted peak displacement. The results indicate that among the six ensemble learning models， the CatBoost model exhibits the best overall performance， effectively characterizing the nonlinear variation pattern of peak displacement in CFST members under multi-parameter coupling. Impact velocity contributes the most to the prediction of peak displacement and serves as the dominant parameter. Hammer mass and section diameter rank next in importance， whereas the influence of the axial compression ratio is relatively weak. Overall， impact velocity， hammer mass， and specimen length tend to amplify the peak displacement， while section diameter， steel tube thickness， and steel yield strength generally mitigate its development. The effects of concrete compressive strength and axial compression ratio remain relatively limited within the current data range. Furthermore， a significant interaction exists between steel yield strength and concrete compressive strength.