Predicting direct punching shear strength in RC flat slabs using a robust multi-layer neural network model

This research introduces a robust multilayer perceptron neural network model with a 6-7-4-1 configuration, designed to estimate the punching shear strength in flat slabs. The model's accuracy was evaluated using a validation technique called repeated random sub-sampling with a confidence level of 94.5% applied to filter the data. Numerical expressions considering factors such as slab depth, column dimensions, steel reinforcement ratio, concrete strength, steel yield, and the position of the contra flexural point were proposed to predict punching shear strength for design control purposes. Furthermore, a comparative analysis was performed between the robust multilayer perceptron neural network model and both the ACI and EC-2 codes for predicting the punching shear strength of flat slabs. The study revealed that while the different codes produced similar results for punching shear capacity, they diverged significantly from the actual experimental outcomes. In contrast, the introduced Artificial Neural Network (ANN) model provided predictions closely aligned with experimental results, underscoring its practical viability. The standard deviations (SD) of the differences between the experimental results and the predicted values were used as performance indicators. The ANN model exhibited a low SD of 53, while the ACI and EC-2 codes, serving as a reference for the comparative codes, showed a significantly larger SD, reaching 360.4 and 382.2 respectively. The ANN model also demonstrated superior performance in other error metrics, including the Mean Absolute Percentage Error (MAPE). The ANN model's MAPE was 13.3%, significantly lower compared to ACI 318-19's 78.3% and the EC-code's 80.1%. This highlights the ANN model's superior accuracy and practicality in estimating punching shear strength in flat slabs.