Machine Learning Prediction of the Compressive Strength of Nano-Silica-Modified Hybrid Geopolymer Mortar

Geopolymer materials are increasingly recognized as sustainable alternatives to conventional cementitious materials due to their lower environmental impact and promising engineering performance. Recent studies have demonstrated that incorporating nanomaterials can further enhance the properties of geopolymer systems. In particular, nano-silica has been reported to significantly improve the mechanical performance of geopolymer materials. However, accurate prediction of compressive strength remains challenging because of the complex nonlinear interactions among mix design parameters, activator chemistry, and curing conditions. This study develops a machine learning framework to predict the 28-day compressive strength of nanosilica-modified hybrid geopolymer mortar using a dataset of 73 mixes compiled from the literature. Five regression models, Linear Regression, Random Forest, CatBoost, Extreme Gradient Boosting, and Extra Trees Regressor, were benchmarked using 10-fold cross-validation and 100 repeated Monte Carlo simulations. The three best baseline-performing models, namely Extra Trees Regressor, CatBoost, and Random Forest, were then optimized using Bayesian hyperparameter tuning with the Optuna framework and Tree-structured Parzen Estimator sampler across 200 optimization trials. The optimized Extra Trees Regressor achieved the best performance, with R² values of 0.906 and 0.844, RMSE values of 5.43 and 6.24 MPa under 10-fold cross-validation and Monte Carlo simulation, respectively. Feature-importance analysis showed that binder composition, including slag, fly ash, and metakaolin contents, was the dominant factor, contributing more than 50% of the model importance, followed by nano-silica dosage at approximately 15%. While NaOH molarity, alkaline activator content, and curing conditions exhibited comparatively lower influence, their contributions remained consistent with established geopolymerization mechanisms. In addition, the optimized model was deployed as an open-access web-based prediction tool to support practical strength estimation and mix-design decision-making.