Comparing mechanistic and data-driven models for predicting punching shear capacity of FRP-RC slabs

The deterioration of reinforced concrete infrastructure due to steel corrosion has accelerated the adoption of Fiber-Reinforced Polymer (FRP) bars as a durable alternative. However, the linear-elastic nature of FRP introduces brittle failure modes in flat-plate systems, necessitating accurate predictive tools for punching shear capacity. This study addresses the challenge of accurately predicting punching shear capacity in FRP-reinforced concrete slab-column connections. Traditional design codes, such as ACI, CSA, and JSCE, were evaluated against an extensive experimental database of 148 tests. The results showed that these codes have low predictive accuracy, with coefficients of determination (R²) as low as 0.190. In parallel, three data-driven machine learning models, Gradient Boosted Regression Trees (GBRT), Random Forest (RF), and an Artificial Neural Network (ANN), were developed and optimized. These models were validated using 10-fold cross-validation, and their performance on unseen data was found to be significantly superior. For instance, the GBRT model achieved an R² of 0.923 and a mean absolute error (MAE) of 49.90 kN, which is a 56% improvement over the best-performing code-based model. The consistency of the predictions was also markedly better, with the coefficient of variation (CoV) of the experimental-to-predicted ratio decreasing from 0.25 for the best-performing code-based models to 0.08 for the machine learning models. To ensure transparency, Explainable Artificial Intelligence (XAI) techniques were used to diagnose model behavior. Unlike standard black-box applications, this approach revealed that effective depth is the most influential parameter, quantitatively confirming the localized nature of the failure. The study concludes that the developed machine learning models provide significantly higher predictive accuracy and reduced scatter compared to traditional design codes, offering a viable alternative for structural design.