Neural Network Potentials Enable Accurate Molecular Dynamics of Resin Pyrolysis at Experimentally Relevant Temperatures

This is the authors' abstract. We don't add key points for this paper.

Many significant and intricate chemical reactions, such as pyrolysis, oxidation, and catalysis, occur under high-temperature conditions, making their study of great scientific and practical importance. In this work, we developed a neural network (NN) potential specifically for resin pyrolysis at experimentally relevant temperatures (750, 1000, and 1500 K) and applied it to molecular dynamics (MD) simulations to investigate product distributions and reaction pathways. The NN potential was trained on a large data set comprising energies and forces of more than 30,000 DFT-calculated structures, with iterative data augmentation and retraining to ensure transferability across diverse configurations. Compared with the conventional ReaxFF force field, which requires unrealistically high simulation temperatures (2000–3500 K) and consequently predicts mainly small molecules, our NN-MD simulations reproduce experimentally relevant conditions, accurately predict reaction products, and capture the evolution of heavier macromolecular species such as phenolic compounds. The simulations, performed for 20 ps with a 0.2 fs time step, also provided kinetic insights and mechanistic pathways inaccessible to traditional methods. This NN-based approach significantly reduces computational cost and time while retaining near-DFT accuracy, offering a more realistic and efficient framework for understanding polymer pyrolysis and extending to other high-temperature chemical processes.