Explainable machine learning models for punching shear strength estimation of flat slabs without transverse reinforcement

Abstract

Flat slabs, despite their aesthetic qualities and widespread use in construction, are susceptible to brittle shear failure. In addition, although design provisions are available, they are often associated with high bias and variance. This study evaluates the efficiency of machine learning-based approaches in establishing accurate prediction models for the punching shear strength of flat slabs without transverse reinforcement. To this end, 380 experimental results from various literature are assembled in this study. In addition to linear regression, seven machine learning methods as ridge regression, support vector regression, decision tree, K-nearest neighbors, random forest, adaptive boosting, and extreme gradient boosting-are considered in this study to obtain the best prediction model for the punching shear strength of flat slabs. Based on random assignment of the data into training and test sets and a performance evaluation of the test set, the extreme gradient boosting model is shown to have the highest coefficient of determination and lowest mean square error estimate. The superior performance of this machine learning model is further underscored through comparison of its shear strength predictions with those of existing code provisions and empirical models. It is noted that the extreme gradient boosting model has a coefficient of determination of 0.98, and the associated coefficient of variation is 0.09. This study also employs the SHapley Additive explanation method to explain the importance and contribution of the factors that influence the punching shear strength in the extreme gradient boosting model.