Reduced-Feature Hybrid Machine Learning for Coal Ash Sintering-Temperature Prediction Using Chemistry and...

Accurate prediction of the coal ash sintering temperature is essential for mitigating slagging, agglomeration, fouling, and operational instability in combustion and gasification systems. In this study, three reduced-feature machine-learning frameworks were developed and compared for predicting the coal ash sintering temperature using experimentally measured ash composition data collected from multiple literature sources. The first framework employed a chemistry-informed extreme gradient boosting (XGBoost) model using the most influential oxide descriptors identified through feature importance analysis. The second framework utilized a physics-derived approach based on weighted physicochemical descriptors, including log-transformed weighted viscosity at the melting point, weighted melting point, and weighted surface tension. The third framework integrated chemistry-informed and physics-derived descriptors within a hybrid XGBoost architecture. Feature-ranking analysis identified CaO, TiO₂, and SO₃ as the dominant chemistry descriptors, while log-transformed weighted viscosity at melting point, weighted melting point, and weighted surface tension emerged as the most influential physics-based descriptors. The chemistry-informed XGBoost model achieved R² values of 0.740 under Leave-One-Out Cross-Validation (LOOCV) and 0.505 under independent 50/50 validation. The physics-derived framework exhibited lower predictive performance, with Random Forest and Gradient Boosting models achieving R² values of 0.457 and 0.253, respectively. Hybrid physics-chemistry models consistently outperformed both individual frameworks. The Top-2 Chemistry + Top-2 Physics configuration achieved the highest LOOCV performance (R² = 0.869), while the Top-3 Chemistry + Top-3 Physics configuration demonstrated the strongest independent validation performance (R² = 0.741, MAE = 20.55 °C, and RMSE = 24.80 °C). The improved performance of the hybrid models suggests that combining oxide chemistry with property-derived descriptors improves prediction of coal ash sintering temperature within the present data set. The results highlight the value of hybrid machine-learning framework using property-derived descriptors for improving predictive accuracy, enhancing interpretability, and reducing model complexity. The proposed framework provides a promising methodology for predicting ash behavior in coal combustion, gasification, PFBC systems, and other thermal conversion technologies.