Novel interpretable machine learning models for predicting compressive strength of nano-silica concrete

Abstract

This study presents a comprehensive comparative analysis of several machine learning (ML) models for predicting the compressive strength (CS) of nano-silica (NS)-enhanced concrete. A large dataset comprising 724 experimental mix designs was compiled from the literature, including various variables such as cement content, water-to-binder ratio, fine and coarse aggregates, nano-silica content, superplasticizer content, and curing time. Six ML algorithms were developed and evaluated: interaction model, full quadratic, artificial neural network (ANN), M5P-Tree, gradient boosting (GB), and random forest (RF). Model performance was assessed using R2, root mean square error (RMSE), mean absolute error (MAE), scatter index, and objective value (OBJ). Among all models, the RF model achieved the highest predictive accuracy with R2 values of 0.99 and 0.95 for the training and testing datasets, respectively, and the lowest prediction errors (Training: RMSE = 2.13 MPa and MAE = 1.36 MPa), followed by ANN and GB models. Sensitivity analysis revealed curing time as the most influential factor, while partial dependence plots exhibited the nonlinear effect of nano-silica quantity, with an optimal strength response at approximately 15 kg m−3. In addition, SHapley Additive exPlanations analysis was employed to enhance model interpretability, confirming the dominant influence of curing age and water-to-cement ratio on compressive strength prediction. Compared with many previous studies relying on limited datasets or single-model approaches, this study provides a robust, interpretable, and generalizable ML framework for optimizing nano-silica concrete mix design. The findings highlight the strong potential of ML, particularly ensemble models combined with explainable AI techniques, to improve prediction reliability, reduce trial-and-error experimentation, and support more cost-efficient and sustainable concrete design.