Optimizing design code predictions of concrete interface shear via physics-informed Gaussian processes

Reliable prediction of shear strength at concrete-to-concrete interfaces is essential for the safe and economical design of composite and rehabilitated concrete structures. However, existing international design standards exhibit systematic, parameter-dependent biases arising from their empirical formulations. This study develops a physics-informed Gaussian Process Regression framework to quantify and correct these structured biases while preserving the interpretability of current design models. A database of 460 push-off tests spanning diverse concrete types, surface conditions, and reinforcement levels is used to evaluate six major design provisions from North American and European standards. Evaluation across these standards reveals significant systematic biases and substantial scatter, identifying strong functional relationships between model errors and key design parameters rather than random scatter as traditionally assumed. Comprehensive analysis shows that a rational quadratic kernel provides the most robust performance, reducing the coefficient of variation of residuals by up to 84% and achieving mean prediction ratios near unity with coefficients of determination exceeding 0.94. Kernel selection based on the Bayesian Information Criterion and confirmed by ten-fold cross-validation demonstrates consistent generalization and effective removal of bias across the input space. Compared with conventional global professional-factor adjustments, the proposed post-processor yields markedly narrower residual distributions and lower dispersion while leaving existing design equations unchanged. Beyond correction capabilities, the framework serves as a comprehensive diagnostic instrument, providing feature-importance analysis that enables evidence-based refinement of design code formulations. The resulting hybrid methodology provides a transparent, data-driven pathway to modernize empirical shear design provisions and is transferable to other bias-correction problems in structural engineering.