An Interpretable and Automated Optimized Stacking Learning Model for Predicting the Shear Capacity of Reinforced Concrete Beams

To address the limited predictive accuracy of existing models and the challenges in optimizing stacking models, this study proposes an interpretable stacking learning model that is automatically optimized using a genetic algorithm (GA) to predict the shear capacity of reinforced concrete (RC) beams with and without stirrups. A diverse candidate model pool was constructed using 11 classical regression algorithms spanning five major domains: ensemble learning, neural networks, Bayesian theory, kernel methods, and nonlinear fitting. A Tree-structured Parzen Estimator (TPE)-based optimization algorithm was then employed to automatically and finely tune the hyperparameters of each individual model, and GA was introduced to perform global automated optimization and dynamic updating of the base model and meta-model combinations within the stacking architecture. The results demonstrate that the proposed GA-Stacking exhibits excellent predictive accuracy and generalization compared to R-Stacking and code-based models, while achieving search efficiency far superior to random search. In addition, both global and local Shapley additive explanations (SHAP) analyses indicate a positive correlation of section geometry (b, h0) and material strength (fc, fsv) to the shear capacity. For λ < 2.80, the SHAP predictions exhibit a pronounced enhancement in shear capacity, followed by a subsequent degradation trend as the shear span-to-depth ratio increases. These findings provide valuable methods and insights for studying and applying RC beam shear capacity prediction.