Explainable machine learning for mechanism analysis and slag optimization of boron removal from silicon using slag refining

The production of high-purity silicon for advanced applications is critically dependent on the removal of boron, a key impurity notoriously difficult to eliminate. While slag refining is a promising industrial route, its optimization is hindered by complex, multi-parameter interactions that challenge traditional experimental methods. This study employs a machine learning (ML) framework to overcome these limitations. A comprehensive database of 4000 entries was constructed by integrating literature data and thermodynamic calculations. Among six ML algorithms evaluated, an optimized XGBoost model achieved superior predictive accuracy for boron removal efficiency (R² > 0.97, MSE < 0.0003). SHAP analysis identified temperature and the content of the third component (e.g., CaCl₂, CaF₂) as the most influential process parameters. Guided by these insights, an orthogonal experimental design was implemented for optimization. The CaO-SiO₂-CaCl₂ system achieved a peak boron removal efficiency of 88 pct, with the CaO-SiO₂-CaF₂ and CaO-SiO₂-Al₂O₃ systems also reaching high efficiencies of ≥ 86 pct and ≥ 75 pct, respectively. This work establishes a data-driven paradigm that integrates machine learning with metallurgical experimentation for intelligent process optimization.