Machine learning-based optimization of concrete strength using interpretable models

Predicting the compressive strength of concrete is vital for optimizing mix designs and ensuring structural reliability. This study integrates the XGBoost machine learning algorithm with SHAP (SHapley Additive exPlanations) to achieve accurate predictions and interpretable insights into the factors influencing concrete compressive strength. Using a dataset of 1,078 instances, five key features - water-to-binder (w/b) ratio, curing age, superplasticizer, water, and sand-to-aggregate ratio (SAR) - were analyzed. The XGBoost model achieved a test R-2 of 0.909 and RMSE of 4.649 MPa, demonstrating robust predictive performance. SHAP analysis identified w/b ratio and curing age as the dominant contributors, with mean absolute SHAP values of 9.27 and 7.85, respectively, aligning with concrete science principles. Segmenting the dataset by curing age (<= 28 days and >28 days) significantly enhanced predictions for mature concrete (R-2 =0.917, p=0.022), while a reduced three-feature model (w/b ratio, curing age, SAR) retained 95% of the full model's accuracy (R-2 =0.867). These findings guide engineers in optimizing w/b ratio and curing age for enhanced strength, offering a practical framework for data-driven, sustainable concrete design.