Molecular Dynamics Study on the Mechanism of Coal High-Temperature Pyrolysis Based on Machine Learning Potential

Understanding the atomic-scale mechanisms of coal pyrolysis is essential for efficient coal utilization and carbon-neutral energy strategies, yet conventional computational approaches often struggle to balance between the high accuracy of quantum-chemical calculations and the efficiency of reactive force fields. To overcome this limitation, we proposed a multiscale computational framework integrating high-throughput density functional theory (DFT) calculations, ReaxFF-based configuration sampling, YARP reaction enumeration, and DPA3-based machine learning potentials (MLPs). Two coal-specific MLPs, DPA3-coal and DPA3-coal@dftb, were constructed and systematically benchmarked on both small molecular systems and larger C20–30 coal fragments extracted from MD simulations. DPA3-coal@dftb model demonstrated significantly improved accuracy over ReaxFF in predicting energies and atomic forces while maintaining good transferability. To balance computational efficiency and accuracy in large-scale simulations, the DPA3-coal model was employed to perform accelerated reactive molecular dynamics simulations of a Solomon-type bituminous coal molecule from 1600 to 2600 K. The simulations revealed temperature-dependent evolution of coke, tar, and gas products, including secondary condensation and deep-cracking processes at elevated temperatures. Higher-level DFT calculations further confirmed the thermodynamic consistency of key reaction pathways involving radical formation, H-transfer, recombination, and CO generation, indicating that coal-specific MLPs provide an effective atomistic tool for investigating mechanistic trends in coal pyrolysis.