Physics-Informed Kolmogorov–Arnold Network for Accurate and Mechanically Consistent Shear Capacity Prediction of FRP-Reinforced Concrete Beams

Shear-capacity predictions for beams with fiber-reinforced polymer (FRP) longitudinal bars and stirrups can be biased, while data-driven corrections can introduce inconsistent parameter responses. Using 106 published tests, six design methods were evaluated and CNR-DT 203/2006 was selected as the baseline for a physics-informed Kolmogorov–Arnold network (PI-KAN). The network learns a multiplicative correction, with finite-difference monotonic constraints imposed on the corrected capacity at FRP-type-specific Sobol points. Model selection used five-fold cross-validation on 84 development samples; 22 samples were reserved for testing. Monotonic constraints reduced the mean validation root-mean-square error (RMSE) from 42.60 to 39.28 kN and mean violation rate from 10.19% to 4.87%. The full model achieved a test RMSE of 27.84 kN and R2 of 0.929, reducing RMSE by 62.4% relative to CNR. Loss ablation supported retaining only the monotonicity penalty. Independent auditing gave 96.85% overall satisfaction of the prescribed monotonic directions. Validation error was also lower than for multilayer perceptrons (MLPs) under matched layer widths and training settings. Structured pruning reduced the parameter count by 58.4%; the pruned model achieved a test RMSE of 24.53 kN, R2 of 0.945, and 90.9% of predictions within ±20% of measurements. Cross-fitted errors and bootstrap analysis yielded an adopted model partial factor of 2.25. The pruned model was implemented in a web calculator for individual and batch assessments. The findings support residual correction with directional constraints within the studied beam configurations and parameter ranges.