Coal pyrolysis is a dynamic process involving competition among multiple molecules and reactions. Experimental measurements provide only statistical structural information, while traditional molecular simulations have difficulty covering large systems and long coking times. This study develops a reinforcement learning method to predict continuous and traceable coal pyrolysis pathways. Multiple coking coal samples are used for model training and testing. A kinetic model selects the reaction type and temperature range, while a reinforcement learning model based on graph neural network selects the target molecule and reaction site. The different coal tests show that the model can predict bond breaking, bridge bond changes, aromatic cluster connections, cross-linking, and volatile release for coal samples not included in training. The predicted structural trends agree with experimental observations. The predicted molecular connections and reaction pathways can support 3D structure calculations. Furthermore, this mechanism prediction algorithm helps the following downstream applications, such as coke quality prediction, carbon-structure control, and studies of coke degradation in hydrogen-rich blast furnaces.

Coal pyrolysis,Reinforcement Learning,Reaction pathway prediction