Physics-informed machine learning for universal ash fusion temperatures prediction: A novel categorical chain differential framework.

Accurate prediction of ash fusion temperatures (AFTs) is crucial for ensuring the operational efficiency and safety of solid-fuel boilers and gasifiers. However, conventional machine learning methods typically treat each characteristic temperature as an independent prediction target, resulting in temperature inversions that violate the required physical ordering of AFTs. This study aimed to develop a Categorical Chain Differential framework for the coupled and physically consistent prediction of the four AFTs across heterogeneous solid fuels. Based on a regressor chain architecture, the framework incorporates fuel category information to account for data heterogeneity. It predicts the temperature differences between adjacent fusion stages and applies non-negative constraints to ensure that the predicted temperatures follow DT ≤ ST ≤ HT ≤ FT. The results demonstrate that the framework achieved an enhancement in predictive accuracy over traditional models using hyperparameters, with the cumulative R2 increasing by up to 0.45. The optimized model reduced the prediction error for the deformation temperature from 84.2 °C to 62.2 °C, while restricting the hemispherical temperature error to 49.7 °C. The framework also eliminated all temperature-sequence violations. These results demonstrate that combining fuel category information with ordered temperature-difference prediction can improve predictive performance while ensuring physically consistent outputs. The framework provides a practical tool for assessing ash fusibility and supporting slagging-risk management during the thermochemical conversion of biomass and waste-derived fuels.