Advanced soft computing techniques for predicting punching shear strength

Abstract

Accurate prediction of punching shear strength is crucial for ensuring the structural integrity and safety of reinforced concrete (RC) flat slab-column connections. However, estimating punching shear strength in complex configurations, such as RC flat slabs, remains a challenging task due to intricate interactions between various parameters. To address this challenge, this study introduces an innovative approach that combines machine learning techniques with hyperparameter optimization strategies to enhance the precision and efficiency of predicting punching shear strength. Specifically, a comprehensive investigation is conducted to assess the performance of XGBoost optimized by different hyperparameter optimization algorithms such as the forensic-based investigation (FBI), Bayesian optimization (BO), grid search (GS), random search (RS), and default setting (DS). The experimental results of 40 runs on three different datasets confirmed the FBI-XGBoost model as the most appropriate for predicting punching shear strength in terms of efficiency and computation time by achieving the highest RMSE, MAPE, MAE, and R2 values. In addition, the statistical results exposed the infeasibility of using RS for the XGBoost model. While the GS method was able to configure the XGBoost model well, doing so required extensive computational time, while the BO-FBI model achieved acceptable prediction accuracy in a reasonable training time. The FBI-XGBoost model was further compared with empirical methods, including ACI-318-08, BS8110-97, and CEB-FIP-90, based on the predictive-to-actual ratio. FBI-XGBoost was identified as the best model by attaining the ratio value closest to 1 and the lowest standard deviation (1.011 ± 0.108). Overall, the statistical experimental results and hypothesis test values firmly support the FBI-XGBoost model as the most reliable free-formulae method for predicting punching shear strength in RC flat slab-column systems.