Mechanical–data-driven performance interpretation and design of concrete-filled steel tubes under lateral impact

Efficiently and accurately assessing the impact performance of concrete-filled steel tubular columns, as well as incorporating impact-resistance considerations into the structural design process, remains a significant challenge in engineering practice. In this study, a database comprising 1575 samples was generated via numerical simulations, and six machine learning algorithms were employed to develop predictive models for assessing the post-impact damage of concrete-filled steel tubular columns. The artificial neural network model demonstrated the strongest accuracy and generalisation capability, even in an extra validation under unfamiliar conditions. The SHapley Additive exPlanations approach was implemented to interpret the individual importance of input features, and the joint effects of structural and loading features were investigated by progressive variation analysis. The results indicate that steel tube diameter, impact velocity and impact mass are the most influential features governing the post-impact damage of columns, while the effect of the axial compression ratio exhibits a two-stage pattern changing from beneficial to detrimental as it exceeds an inflexion point. Increasing the steel tube diameter and steel ratio are the most effective in improving the impact-resistance performance of concrete-filled steel tubular columns, but both exhibit significant diminishing returns. Finally, a machine learning-based inverse prediction model was developed to achieve cost-effective structural design by optimising a balance among steel tube diameter, steel ratio, and axial compression ratio. The preliminary design programs were also developed.