Abstract
Accurately predicting the unconfined compressive strength (UCS) of microsilica-lime stabilized sulfate sand (MSLSS) is critical for the safe and efficient design of infrastructure in arid regions, yet it remains challenging due to the highly nonlinear relationships among influencing factors. This study pioneers the development of hybrid machine learning (ML) models, integrating the Sparrow Search Algorithm (SSA) with XGBoost (XGB), Random Forest, and Decision Tree, for predicting UCS of MSLSS. These models were trained and tested on experimental datasets incorporating input variables: lime content, microsilica content, curing days, curing condition, optimum moisture content (OMC), and maximum dry density. Comprehensive performance evaluation using metrics such as R(2), MAE, MSE, and MRE demonstrated that SSA optimization markedly enhanced the predictive accuracy and generalization capability of all base models, with the RF model exhibiting the most substantial improvement. The hybrid XGB-SSA model achieved the highest overall predictive accuracy, yielding excellent performance on the testing set (R(2) = 0.982, MAE = 1.358). The standard XGB model also displayed competitive results, presenting a practical alternative when model complexity is a concern. SHAP-based interpretability analysis revealed OMC and microsilica content as the most influential input variables. This study provides valuable support for geotechnical design and engineering applications in relevant contexts.