Shear strength prediction of RC beams using mechanics-informed machine learning

In this paper, machine learning (ML) techniques are employed to accurately predict the shear strength of reinforced concrete (RC) beams. Two types of ML models are developed and compared: data-driven ML and mechanics-informed ML. A total of 193 experimental data from RC beams with stirrups that failed in shear are used for training and testing. Nine ML algorithms and four performance evaluation metrics are applied to select the optimal model. The data-driven ML models show high predictive accuracy, but the feature importance results obtained from the SHapley Additive exPlanations (SHAP) method are inconsistent across the nine algorithms, revealing limitations in physical interpretability. To address this issue, the mechanics-informed ML model is developed by applying feature engineering based on mechanics theory. The mechanics-informed ML model exhibits consistent feature importance results maintaining high accuracy, thereby improving interpretability. To derive a closed-form equation representing the ML model’s prediction process, SHAP interpretation results and multiple regression analysis based on the mechanics-informed ML model are combined. The derived equation shows a high-performance score with R2 = 0.95 for the entire dataset. The results indicate that the developed mechanics-informed ML model provides high accuracy and interpretability, enabling the formulation of the prediction process and enhancing reliability in structural design applications.