Predicting the strength of microsilica lime stabilized sulfate sand using hybrid machine learning models optimized...

Area of Science:

• Geotechnical Engineering
• Materials Science
• Machine Learning Applications

Background:

• Accurate prediction of unconfined compressive strength (UCS) for microsilica-lime stabilized sulfate sand (MSLSS) is vital for infrastructure development in arid regions.
• The complex, nonlinear relationships between influencing factors pose significant challenges to traditional prediction methods.

Purpose of the Study:

• To develop and evaluate hybrid machine learning (ML) models for predicting the UCS of MSLSS.
• To investigate the efficacy of the Sparrow Search Algorithm (SSA) in optimizing ML models for enhanced predictive accuracy.

Main Methods:

• Development of hybrid ML models by integrating the Sparrow Search Algorithm (SSA) with XGBoost (XGB), Random Forest (RF), and Decision Tree algorithms.
• Training and testing models on experimental datasets including lime content, microsilica content, curing days, curing condition, optimum moisture content (OMC), and maximum dry density.
• Performance evaluation using R², MAE, MSE, and MRE metrics, complemented by SHAP-based interpretability analysis.

Main Results:

• SSA optimization significantly improved the predictive accuracy and generalization of all base ML models.
• The hybrid XGB-SSA model achieved the highest predictive accuracy (R² = 0.982, MAE = 1.358) on the testing set.
• Optimum moisture content (OMC) and microsilica content were identified as the most influential input variables.

Conclusions:

• Hybrid ML models, particularly the XGB-SSA model, offer a robust and accurate approach for predicting MSLSS UCS.
• The study provides valuable insights for geotechnical engineers, supporting safer and more efficient infrastructure design in arid environments.
• The findings highlight the potential of SSA-optimized ML models in geotechnical material characterization.