Machine learning-based modeling to predict and parametrically optimize the compressive strength of nanomaterial concrete composites exposed to elevated temperatures

The global shift toward sustainable infrastructure has increased the use of environmentally friendly binders as substitutes for Portland cement. These binders provide technical, economic, and environmental benefits. This study applies machine learning (ML) models to predict the compressive strength (CS) of nano-silica (NS) concrete exposed to elevated temperatures. A dataset of 380 experimental mix compositions with eight input parameters was collected and preprocessed. Eight ML models were developed and tested, including AdaBoost (ADB), Decision Tree (DT), Gradient Boosting (GB), K-Nearest Neighbors (KNN), Light Gradient Boosting (LGB), CatBoost (CatB), Random Forest (RF), and Extreme Gradient Boosting (XGB). Model performance was evaluated using a 5-fold cross-validation approach. Based on the outcomes, the GBR, CatB, and RF models achieved the highest accuracy with mean test R 2 values of 0.980, 0.978, and 0.975, respectively. The XGB model also showed strong predictive capability with R 2 values above 0.937. GBR performed the best, with mean RMSE 1.60 MPa in training and 3.28 MPa in testing, followed by CatB, which had mean RMSE 2.41 MPa in training and 3.50 MPa in testing. SHapley Additive exPlanation (SHAP) and Partial Dependence Plot (PDP) analyses were used to examine feature importance and optimize mix compositions. Temperature (mean SHAP: 10.83) and water content (9.63) were identified as the most influential factors on CS. PDP shows that CS increases with cement above 450 kg/m 3 , NS up to 15 kg/m 3 , and coarse aggregate above 800 kg/m 3 , while decreasing with extra fine aggregate, water above 140 kg/m 3 , and high temperature. The findings demonstrate that ML models can serve as efficient tools for forecasting concrete performance, optimizing mix design, and reducing reliance on costly experimental trials.