Comparative Prediction of Unconfined Compressive Strength of Geopolymer-Stabilized Soil Using Interpretable and Black-Box Machine Learning Models

Abstract

Accurate prediction of unconfined compressive strength (UCS) can reduce the experimental effort required for the design of geopolymer-stabilized soils. However, highly accurate machine-learning models often provide limited engineering transparency. This study addresses this issue by comparing gene expression programming (GEP), support vector regression (SVR), random forest (RF), and gradient boosting regression trees using LSBoost (GBRT). A database of 270 experimental records was used, with seven input variables describing soil plasticity, precursor contents, activator conditions, and chemical ratios. All models were evaluated using the same training–test split and held-out test set. GBRT achieved the highest predictive accuracy, with an independent-test R² of 0.9814, RMSE of 0.9771 MPa, and MAE of 0.6080 MPa. RF, SVR, and GEP followed with R² values of 0.9661, 0.9416, and 0.9023, respectively. Despite its lower accuracy, GEP was retained as the interpretable model because it provided an explicit mathematical expression involving all seven inputs. Its behavior was further examined using multi-point one-at-a-time (OAT) sensitivity analysis and exact background-averaged SHAP analysis. Both methods identified GGBS as the dominant variable, with normalized important values of 45.43% and 42.40%, respectively. SHAP further ranked PI and FA as the next most influential inputs. Higher GGBS generally increased predicted UCS, whereas higher PI and FA produced negative contributions within the investigated data domain. The proposed framework therefore distinguishes predictive superiority from model interpretability and provides a transparent basis for understanding UCS predictions in geopolymer-stabilized clayey soils.