The interpretable deep learning-based coke quality prediction strategy aligned with domain knowledge for sustainable smart cokemaking applications

• An interpretable coke quality prediction model aligned with domain knowledge was proposed. • The model can reasonably identify the reasons why some coal samples produce low-quality coke. • Aligning the learned correlation with domain knowledge can effectively reduce prediction errors. • Factors influencing the accuracy of the proposed interpretable neural network were explored. Accurate prediction of coke quality is critical for optimizing coal blending schemes to improve the efficiency of the coke production process, thereby reducing energy consumption and CO 2 emissions to advance the coking industry towards carbon neutrality. However, existing prevalent prediction methods of coke quality, primarily based on black-box Artificial Neural Networks (ANNs), have poor interpretability. They often learn spurious correlations that contradict fundamental cokemaking behaviors, degrading model performance and making them untrustworthy for practical applications. In pursuit of trustworthy smart cokemaking, this study is the first to propose an interpretable neural network model aligned with domain knowledge for Coke Reactivity Index (CRI) and Coke Strength after Reaction (CSR) prediction. 422 Chinese coal samples were collected for analysis, with coking coals composed of 18 properties and coal blends composed of 7 properties. Based on established domain knowledge of the cokemaking process, important coal properties whose importance contradicted domain knowledge in the original ANN model were identified with the SHapley Additive exPlanations (SHAP) algorithm and corrected using the proposed Contextual Decomposition Regression Explanation (CDRE) algorithm. Experimental results demonstrate that the optimized ANN aligned with domain knowledge provides more reasonable feature importance and lower prediction error in coking coals and coal blends. The model can also reasonably analyze the reasons behind some coal samples producing low-quality coke. The learned high-level representations of coal samples from the optimized ANN exhibited a wider distribution, allowing for better differentiation of their disparities. Moreover, the coefficient value of the loss function and the number of network layers were identified as critical factors influencing model performance. The findings of this study provide a feasible solution for interpretable coke quality prediction, effectively supporting the development of trustworthy smart cokemaking.