An interpretable large language model framework with the cokemaking-oriented prompt and fuzzy-enhanced explanation for coke quality prediction

Abstract

Although many data-driven coke quality prediction models have been proposed, existing approaches primarily rely on structured data and fail to effectively utilize the large amount of accumulated domain knowledge, such as published technical reports and scientific literature in the cokemaking field. This limitation restricts further improvements in prediction performance. Moreover, there is still a lack of a unified framework that can achieve high prediction accuracy, handle incomplete data, provide interpretability, and support interactive analysis. To address these challenges, this study proposes a novel Large Language Model (LLM) framework with a cokemaking-oriented prompt and a fuzzy-enhanced explanation mechanism for predicting Coke Reactivity Index (CRI) and Coke Strength after Reaction (CSR). The input coal properties are represented as both textual tokens and graph tokens, which are further integrated within the LLM to capture complementary semantic and structural information. A total of 928 Chinese coal samples are used for model development, with the testing set entirely composed of real industrial cokemaking data. The proposed model achieves superior performance compared with mainstream machine learning approaches, with mean absolute errors of 1.72 for CRI and 2.37 for CSR, and demonstrates good generalization ability on 39 Australian coal samples. It is further found that cokemaking-oriented prompt outperforms generic prompts by activating domain-specific knowledge, leading to more compact and well-separated feature representations in the learned LLM embedding space. In addition, the fused token representations outperform single-modality inputs due to complementary information integration. The model also provides interpretable results, generating feature importance scores and explanations that are highly aligned with domain knowledge.