A novel small data prediction model of activity for multiple slag systems based on TabPFN and SHAP

Abstract

Accurate prediction of activity in multi-component slag systems is of great significance for optimizing metallurgical processes. Existing models, which include both thermodynamic calculations and neural networks, often rely on simplifying assumptions or require extensive hyperparameter tuning. In addition, the experimental acquisition of slag activity data is difficult due to high-temperature conditions and complex procedures, resulting in limited and scarce data. To address these limitations, a novel model was proposed integrating TabPFN and SHAP analysis for activity prediction on small data. Model performance was benchmarked against existing models, and the results show that TabPFN not only surpasses other models in terms of predictive accuracy but also enables enhanced explainability by SHAP analysis. The TabPFN model achieved an R² of 0.9816, an RMSE of 0.0451, an MAE of 0.0243, and a model response time of 0.21 s. Moreover, a SiO₂ activity prediction tool was developed, featuring user-friendly operation, fast computation, and automatic model updating without the need for extensive hyperparameter tuning when new data are introduced, offering practical value for industrial deployment. This work provides a data-driven and interpretable approach for the real-time and accurate prediction of slag activity under complex process conditions.