Coke quality prediction is crucial for optimizing coal blending and reducing the time and cost associated with conventional cokemaking tests. However, existing prediction models primarily rely on structured coal properties and have limited ability to incorporate cokemaking domain knowledge, handle incomplete data, and provide interpretable results. To address these issues, this study proposes an interpretable coke quality prediction framework based on a large language model with cokemaking-oriented prompts to predict the Coke Reactivity Index (CRI) and Coke Strength after Reaction (CSR). Within this framework, the available numerical coal properties are converted into textual tokens through domain-specific prompts and graph tokens through graph-based feature encoding, enabling the integration of complementary semantic and structural information without requiring missing-value imputation. A total of 928 Chinese coal samples were used for model development and evaluation, with the testing set consisting entirely of industrial cokemaking data. An additional 39 Australian coal samples were employed to evaluate model generalization across different coal distributions. Experimental results demonstrate that the proposed framework outperforms the evaluated machine learning models, achieving mean absolute errors of 1.72 for CRI and 2.37 for CSR on the Chinese testing set. Further analyses show that cokemaking-oriented prompts and fused token representations contribute to improved prediction performance. The framework also provides feature importance scores and human-readable explanations which are aligned with cokemaking knowledge, supporting interpretable coke quality prediction and coal blending decisions.
 

Coke quality prediction,Deep learning,Large language model,Cokemaking-oriented prompts,Feature fusion