Predicting shear capacity of slender FRP-reinforced concrete beams

To address the complex shear transfer mechanisms and brittle failure modes of fiber-reinforced polymer (FRP) reinforced concrete beams without stirrups, this study introduces an extractable predictive framework built on an experimental dataset stratified by the longitudinal reinforcement ratio (ρ_f). Identical hyperparameters were applied across three optimization algorithms (Quasi-Newton (L-BFGS), Stochastic Gradient Descent (SGD), and Adaptive Moment Estimation (Adam)) to conduct a comparative analysis using 5-fold cross-validation. The L-BFGS algorithm yielded the lowest prediction errors for this specific structural dataset. The final non-linear artificial neural network (ANN) model achieved a test R² of 0.924 and a Mean Absolute Error (MAE) of 9.91 kN, representing approximately 12.7% of the dataset’s mean experimental capacity (78.10 kN). Statistical evaluation via a Wilcoxon signed-rank test (p<0.05) indicates that the model yielded lower prediction errors compared to 13 international design codes and empirical equations. Furthermore, the architecture was extracted into explicit algebraic matrices and embedded within a standalone software application. Advanced model interpretability via permutation feature importance indicates statistical associations in which web width (b_w), effective depth (d), and reinforcement ratio (ρ_f) operate as the dominant predictors. This explicit framework provides an accessible computational alternative to existing empirical codes, facilitating optimized material usage and refined safety margins in the design of sustainable FRP-reinforced concrete infrastructure.