Prediction of Punching Shear Capacity for Fiber-Reinforced Concrete Slabs Using Neuro-Nomographs Constructed by Machine Learning

Punching shear capacity is an important parameter in designing structural elements. Accurate estimation of punching shear capacity typically requires rigorous calculation schemes. Especially for fiber-reinforced slabs, traditional design methods may not be sufficient to predict the interaction between different influencing parameters affecting punching shear capacity for such slabs. In this study, multiple state-of-the-art machine learning (ML) algorithms were utilized, namely, regression learner, ensemble tree (bagged and boosted), support vector machine (SVM), regression decision tree, Gaussian process regression (GPR), and artificial neural networks (ANN). A comprehensive evaluation of the six ML techniques was conducted with respect to model accuracy and computational efficiency. The results demonstrated that the ANN-based algorithms outperformed other ML approaches based on the values of root mean square error (RMSE), mean absolute error (MAE), and correlation coefficient (${\mathrm{R}}^2$). Furthermore, the analysis of the results has shown that the slab effective depth has the most significant effect on the predicted punching shear, followed by the width of applied load and concrete compressive strength. Python coding (with the assist of Pynomo software) was utilized to create nomographs integrated with weights resulting from the neural network model. Such neuro-nomographs can be used to simulate the results of the developed ANN model. Moreover, the values of tested punching shear capacities over predicted values (${\mathrm{V}}_{\mathrm{test}}/{\mathrm{V}}_{\mathrm{pred}}$) using the neuro-nomograph have shown mean and coefficient of variation (COV) values of 1.00 and 0.05, respectively, indicating remarkably minor scatter in the prediction.