Geopolymer concrete (GPC) is a low-carbon alternative to ordinary Portland cement concrete, but its compressive strength depends on complex interactions among mixture proportions and curing conditions, making laboratory-based mix design time-consuming. This study proposes an ensemble machine-learning framework optimized using the forensic-based investigation (FBI) algorithm to accurately predict GPC compressive strength and support rapid mixture assessment. A dataset of 350 experimental GPC samples was used to develop and evaluate seven FBI-optimized models (XGBoost, LightGBM, GBM, AdaBoost, random forest, bagging SVM, and decision tree) under a 10-fold cross-validation method. Overall, FBI optimization improved predictive performance, with boosting-based ensembles showing the most consistent accuracy. Among all models, FBI-XGBoost achieved the best testing performance (RMSE = 2.93 ± 0.61 MPa, MAE = 2.11 ± 0.44 MPa, MAPE = 7.08 ± 1.16%, and R 2 = 0.89 ± 0.07), substantially outperforming the corresponding non-optimized boosting baselines by reducing RMSE by approximately 20–22% and MAPE by approximately 30–40%. Feature-importance analysis using SHapley Additive exPlanations (SHAP) indicates that sodium hydroxide concentration and sodium silicate content are the most influential variables governing compressive strength. The proposed FBI–XGBoost provides an accurate, robust, and reproducible tool for strength prediction, offering practical value for reducing trial mixes and accelerating sustainable GPC design. • FBI metaheuristic optimizes ensemble models for GPC strength. • FBI-XGBoost achieves superior accuracy with a MAPE of 7.08%. • Boosting frameworks outperform bagging approaches. • SHAP analysis identifies alkali activators as key strength factors.
Nguyen et al. (Sun,) studied this question.