Proposing an inherently interpretable machine learning model for shear strength prediction of reinforced concrete beams with stirrups

Abstract

Advanced machine learning (ML) models are utilized for accurate shear strength prediction of reinforced concrete beams (RCB), but their lack of interpretability makes it unclear how models make specific predictions, reducing their reliability and applicability. Mainstream model-agnostic interpretation methods require numerous additional computing procedures, limiting the interpretation efficiency, and the model prediction process remains unknown. This study proposes an inherently interpretable shear strength prediction model for RCB with stirrups based on Explainable Boosting Machine (EBM) and Bayesian optimization. The EBM algorithm decomposes the predicted shear strength into individual shape functions, thus thoroughly revealing the prediction process. Besides, the ML tasks of selecting a model with optimal hyperparameters are automatically performed by Bayesian optimization to reduce computational cost. The developed model is validated on a database including 372 specimens. Compared with shear design codes, empirical models and prominent ML models, the EBM model obtains most accurate predictions for the test set with R2, MAE, and RMSE of 0.9293, 35.42 kN, and 50.99 kN, respectively. The accuracy of EBM is robust to varying input variables through trend analysis. Its interpretability fully discloses the contribution of individual features to the shear strength, and the model rationality is verified by comparing feature contribution with existing mechanisms. The proposed EBM model achieves inherently interpretable shear strength predictions while maintaining high accuracy, which promotes the model applicability in structural assessment.

Keywords

• Shear strength
• Reinforced concrete beam
• Machine learning
• Explainable boosting machine
• Interpretability
• Bayesian optimization