Modeling and predicting the fusion behavior of municipal solid waste incineration fly ash: A comparative study between machine learning and thermodynamic simulation

Predicting the fusion behavior of municipal solid waste incineration fly ash (MSWI-FA) is crucial for energy-efficient melting treatment. Traditional experimental methods and thermodynamic simulations are limited by high costs, inefficiency, and low accuracy, with FactSage particularly suffering from significant overestimation errors due to its sensitivity to basicity and idealized thermodynamic assumptions. This study presents the first comprehensive comparison between machine learning (ML) and FactSage simulation for predicting the flow temperature (FT) of MSWI-FA. ML approaches overcome FactSage's limitations by directly learning from experimental data to capture complex, non-linear relationships between ash components without relying on equilibrium assumptions, demonstrating the superior accuracy and generalization performance. Among 6 evaluated ML models, the Random Forest model performed the best, achieving a mean absolute error of 40.7 °C across the whole dataset, which is due to its ensemble learning architecture that effectively handles feature interactions and prevents overfitting. Additionally, ablation experiments presented the necessity of principal component analysis dimensionality reduction for model performance, improving the R2 by 11.54 %. Model interpretability analysis revealed the high importance of basicity and the silicon-to-aluminum ratio, as well as the potential profound impact of feature interactions on the FT. SHAP dependence plots confirmed these key feature relationships, demonstrating the data-driven model's physical consistency. A model-based FT prediction program achieved an average relative residual error of 1.97 % on independent validation samples, offering high accuracy and robustness. This study alleviates the pressure for the efficient melting treatment of MSWI-FA, providing engineering support for optimizing melting temperature and for intelligent co-melting formulations of solid waste.