Data-based modeling of the punching shear capacity of concrete structures

Abstract

Flat slabs have been widely used in construction in recent decades because they allow greater clear heights and more versatile architectural layouts compared to beam-column solutions. In such structures, punching shear failure is often critical. This type of failure is brittle and occurs without pronounced advance notice, and the alternative load paths cannot be ensured. Therefore, this failure can lead to progressive collapse. For this reason, proper reliability assessment of slabs against punching is of great importance. Improved accuracy and minimization of uncertainty in the predictive models are critical. Currently, there is a wide debate in the international community about the best predictive model for punching shear capacity and there is no clear consensus among researchers. This study focuses on the development of predictive models for punching shear capacity of concrete structures using machine learning (ML) algorithms such as the Gaussian Process Regression (GPR) and Support Vector Regression (SVR). The models were developed using a set of 501 experimental punching shear tests from existing theses, dissertations, technical papers, and research reports. The performance of the developed prediction model is evaluated statistically and by comparison with EC 2. Based on the soft computing approach of the ML model, it is possible to achieve precise and accurate prediction of the ultimate punching shear capacity with significant speed. Thus, this method is suitable for implementation in probabilistic models in a reliability framework. And it can be used to explore parameter sensitivities of the punching shear failure mechanism. The results of this study indicate that the Gaussian Process Regression and specifically the Matern 5/2 algorithm can describe very accurately the punching shear capacity and it is suitable for probabilistic evaluations.