Interpretable machine learning for predicting shear capacity of ultra-high-performance concrete beams

Accurate prediction of shear capacity in reinforced concrete beams is crucial for structural safety assessment. Conventional theoretical methods exhibit significant variability due to the complexity of shear failure mechanisms. This study presents an interpretable machine learning (ML) framework to enhance shear capacity prediction. A comprehensive database of 1175 beam specimens was developed, including normal concrete (NC) and ultra-high-performance concrete (UHPC) beams across three distinct cross-sectional geometries. The ML algorithms–support vector regression, artificial neural network, K-Nearest neighbors, decision tree, random forest, gradient boosting machine, light gradient boosting machine, adaptive boosting, categorical boosting, and extreme gradient boosting (XGBoost)–were optimized using 10-fold cross-validation and random search. The XGBoost algorithm demonstrated superior performance, achieving an R² of 0.986 on the aggregated data set. Interpretability analysis with Shapley additive explanations identified beam depth (h), shear-span ratio (m), cross-sectional area (A_c) and fibre factor (λ_f) as critical features, highlighting their individual and interactive contributions. Moreover, a unified ML-based shear strength prediction model was developed that simultaneously captures the shear behaviour of both NC and UHPC beams, incorporating physically meaningful input features derived from the data set, thereby overcoming the limitations of separate empirical formulations. The proposed ML-based model significantly improved the accuracy of shear strength predictions compared to traditional empirical methods, enhancing reliability in structural design.