A hybrid strategy of AutoML and SHAP for automated and explainable concrete strength prediction

Abstract

The precise prediction of concrete compressive strength is essential for ensuring safe and reliable infrastructure design and construction. However, traditional empirical models often struggle to accurately predict compressive strength due to the complex nonlinear relationship between concrete properties and target strength. This study introduces an AutoML-SHAP (Automatic Machine Learning - SHapley Additive exPlanations) strategy, designed to automatically predict the compressive strength of concrete and provide insightful interpretations of the predictive outcomes. The AutoML model uses K-fold bagging and multilayer stacking to automate model selection and hyperparameter tuning. The integration of AutoML and SHAP offers synergistic benefits, facilitating the development of a precise, efficient, and comprehensively interpretable model. Results demonstrate that AutoML-SHAP model outperforms other machine learning models for predicting compressive strength without human intervention. The AutoML model is automatically established within 174 s and exhibits comparable predictive performance with R2 = 0.96, RMSE = 3.63, and MAE = 2.41. SHAP provides a global explanation of the impact of mixing parameters on compressive strength, and a local explanation of feature contribution to each prediction, making the process transparent and reliable. Feature dependence analysis reveals the influence tendency of mixing parameters on strength.

Keywords

• Concrete Compressive Strength
• Auto Machine Learning
• SHapley Additive exPlanations
• Predictive model