Prediction of punching shear strength in flat slabs: ensemble learning models and practical implementation

This study proposes new models to predict the punching shear strength of flat slabs without transverse reinforcement by harnessing the power of machine learning through ensemble learning models. Leveraging two distinct databases—one with 522 samples with six input variables and another comprising 745 samples with four essential input variables. Six ensemble learning models, including Random Forest, AdaBoost, Light GBM, GBRT, CatBoost, and XGBoost, are evaluated. Through a combination of Bayesian optimisation and tenfold cross-validation technique, the CatBoost model emerges as the standout performer, achieving a coefficient of determination (R²) of 0.97 for both training and testing datasets. Notably, these models exhibit their superior prediction accuracy as compared to existing design codes and empirical equations. To further validate robustness of the models and evaluate the randomness of the databases, Monte Carlo simulations are employed. Additionally, the adaptability of XGBoost, CatBoost, and GBRT was tested using 200 random data points that extended beyond the original database's range, showcasing their capacity to provide reliable predictions in extended scenarios. Finally, a user-friendly interface application is developed for estimating the punching shear strength of RC flat slabs.