Exploring the Connection Between Ultimate Punching Resistance and Failure Mode of Slab-Column Joints with Machine Learning Algorithms

While reinforced concrete (RC) flat slab systems offer an advantage in terms of architectural flexibility and construction efficiency, they remain susceptible to abrupt punching shear failures at the slab-column joints. The current methods for assessing failure modes and punching resistance have considerable limitations, primarily due to an incomplete understanding of how failure mechanisms affect ultimate resistance, resulting in substantial uncertainty in prediction. By utilizing machine learning (ML) algorithms, this study contributes to the persistent problem of determining a quantitative correlation between failure mode progression and punching shear capacity. The method for identifying failure modes is presented using three ML algorithms (DT, RF, and XGBoost) and approach to accurately predict punching resistance, using failure modes as fundamental input variables integrated with two metaheuristic optimization (BFO and GWO) techniques. There has been extensive validation of the BFO-XGBoost model, which has shown superior performance in the classification of failure modes (F1=93.8%, Acc=98.3%) as well as prediction of punching resistance (R2 = 0.967, MAE = 0.032MN, and RMSE = 0.049MN). By analyzing the interpretability of models using Shapley additive explanations (SHAP), crucial parameters and their interactions can be identified. It offers unique insights into the mechanism by which failure characteristics are related to ultimate resistance, as well as practical recommendations for mitigating brittle failure risks and enhancing ultimate resistance.