The development of a multimodal learning framework based on coal properties for Coke Reactivity Index and Coke Strength after Reaction prediction

Reliable coke quality prediction is essential for modern cokemaking operations, as it enables the optimization of coal blending schemes to reduce production costs and support decarbonization. Although vitrinite reflectance distributions of coal blends are widely recognized in industrial practice as a key factor influencing coke quality, they are rarely considered in prediction models. Few previous studies have explored alternative solutions, such as using aggregated numerical properties derived from vitrinite reflectance distributions. However, they can not fully capture the modality of the distribution, which limits prediction accuracy. To address this issue, this study develops a novel multimodal learning framework that integrates numerical coal properties and vitrinite reflectance distribution images for coke quality prediction, including the Coke Reactivity Index (CRI) and Coke Strength after Reaction (CSR). A total of 891 Chinese coal samples were collected for model establishment, with the testing set entirely composed of coal blends to better reflect practical industrial conditions. The proposed multimodal model achieves low prediction errors, with mean absolute errors of 1.32 for CRI and 1.58 for CSR, outperforming regression approaches using aggregated reflectance property and purely numerical models. The results further demonstrate that relying solely on numerical coal properties or vitrinite reflectance distributions is insufficient for accurate prediction, whereas their combination provides complementary information essential for robust performance. Moreover, the model demonstrates good interpretability by highlighting physically meaningful regions of the vitrinite reflectance distributions, and it can effectively explain coal blends with similar numerical properties but markedly different coke quality, which remains challenging for existing prediction models.