Predicting Unconfined Compressive Strength of Geopolymer‐Stabilized Clayey Soil Using a Weighted Ensemble of ExtraTrees and CatBoost

A reproducible machine‐learning framework is developed for predicting the unconfined compressive strength (UCS) of geopolymer‐stabilized clayey soils from laboratory test data. The model is trained on 283 experimental observations covering three soil types and uses plasticity index (PI), slag and fly ash (FA) content, alkali molarity, and molar ratios as predictors. Tuned ExtraTrees and CatBoost (categorical boosting) models are embedded in a cross‐validated pipeline to prevent data leakage, and their predictions are fused through an out‐of‐fold (OOF) weighted averaging scheme. The proposed ensemble achieves the best hold‐out test performance ( R ²  = 0.9815, root mean squared error (RMSE) = 0.852 MPa, mean absolute error (MAE) = 0.554 MPa), outperforming the individual base learners. Repeated cross‐validation and paired statistical testing further confirm that the improvement over CatBoost is statistically significant ( p = 0.0077). These results show that a leakage‐resistant combination of complementary learners can deliver accurate UCS prediction within the investigated data domain and provide a robust framework for modeling geopolymer‐stabilized soils.