Interpretable machine learning for predicting ash fusion temperatures of biomass-sludge hybrid feedstocks with stage-dependent patterns

Co-gasification of biomass and sludge offers an effective pathway for waste valorization and energy recovery. However, the fusion behavior of mixed ash is highly complex and variable, often leading to slagging and operational instability. To address the difficulty of predicting ash fusion behavior in such complex systems, a dataset comprising 92 biomass-sludge mixed ash samples was assembled, and an interpretable multi-target regression framework was developed. The model simultaneously predicts four characteristic fusion temperatures, namely deformation temperature (DT), softening temperature (ST), hemispherical temperature (HT), and flow temperature (FT), based on chemical composition and thermochemically derived features. Through correlation-based feature selection and Bayesian optimization, the optimized artificial neural network (ANN) achieved high predictive performance (R2 ≈ 0.94) with zero monotonicity violations among predicted temperatures. Shapley Additive Explanations (SHAP) analysis identifies stage-dependent patterns in feature importance and interactions across different ash fusion stages. For example, alkali components such as K2O show negative associations with predicted temperatures at lower-temperature stages, while oxides such as Al2O3 and CaO exhibit increasingly positive contributions at higher-temperature stages. These results suggest that the influence of ash components varies systematically across temperature ranges in complex multi-component systems. This study demonstrates that interpretable multi-target machine learning provides a practical framework for analyzing ash fusion behavior in biomass-sludge mixtures, offering a data-driven basis for slagging risk evaluation and feedstock blending optimization in co-gasification processes.