Performance prediction and interpretable analysis of concrete based on automatic machine learning

Traditional machine learning (ML) algorithms exhibit good generalization capabilities and prediction accuracy in concrete performance prediction; however, their complex hyperparameter tuning process significantly constrains work efficiency. To address this, this study introduces an automated machine learning (AutoML) approach incorporating ensemble automatic hyperparameter optimization techniques, establishing an efficient and automated framework for concrete performance prediction. Firstly, a large amount of experimental data encompassing mix proportions, compressive strength, slump, and porosity was collected, and the correlations between influencing factors and concrete properties were analyzed. Secondly, AutoML models based on various algorithms were developed, error analysis was performed on the model prediction results, and model performance was comprehensively evaluated using five metrics including the coefficient of determination (R²). Finally, interpretability analysis of the prediction results was conducted based on SHAP value theory. The results demonstrate that the AutoGluon model delivered the optimal prediction performance for compressive strength (65.2% of prediction errors distributed within ±2 MPa, R²=0.959) and porosity (R²=0.973), while the MLjar model achieved the best prediction for slump (R²=0.879). SHAP analysis revealed that cement content, binder content, and coarse aggregate content are the key factors influencing compressive strength, porosity, and slump, respectively.