Advanced Modelling for Coke Quality Prediction Based on Neural Networks

This thesis develops an advanced artificial intelligence–based system to predict coke quality, including the Coke Reactivity Index (CRI) and Coke Strength after Reaction (CSR) with carbon dioxide, more accurately and reliably. It integrates various deep learning techniques, such as image-based, domain adaptation, expert knowledge, and graph-based modelling, to address the limitations of traditional empirical models, including low accuracy, poor generalization, limited interpretability, and the inability to handle coal samples with missing properties. The developed framework can help reduce production costs, shorten experimental testing time, and improve the efficiency of coal blending decisions, contributing to the advancement of smart cokemaking.