Hybrid optimised CatBoost model for accurate prediction of geopolymer concrete strength

Geopolymer concrete (GPC) represents a significant advancement in sustainable construction due to its low energy consumption and reduced carbon footprint. Therefore, developing a reliable model to predict the compressive strength of GPC is essential for saving time, energy and cost. In this study, experimental data were collected from Egyptian literature, resulting in a dataset comprising 195 samples with 14 input variables used to predict compressive strength. Six machine learning models were employed, including categorical boosting (CatBoost), extreme gradient boosting, gradient boosting, light gradient boosting machine, bagged tree and random forest. Among all models, the CatBoost algorithm demonstrated the best performance (test R2 = 0.937, MAE = 2.139, EVS = 0.937, MeAE = 1.267, and RRMSE = 0.057). SHapley Additive exPlanation was applied to the CatBoost model to interpret feature importance. The results indicate that higher proportions of slag and fly ash precursors generally lead to increased predicted compressive strength. This research uses Egyptian data to model GPC strength, thereby contributing to regional knowledge. It also offers methodological advancement in both model selection and interpretability. This research contributes to streamlining experimental processes, reducing labour and material requirements, enhancing time efficiency and offering valuable insights for optimising GPC mix design.