Ethylene-Oxygen Combustion: From Machine Learning Potential Function Construction to Molecular Dynamics Simulation

Understanding and accurately modeling combustion processes of complex fuels remain challenging. In this study, we constructed a reactive machine learning potential function (MLP) for the C₂H₄-O₂ system to investigate the corresponding combustion behavior. The MLP was trained using DeePMD-kit with 890,682 configurations, achieving an average fitting error of 0.014 eV for ab initio points on the C₂H₄-O₂ MLP. The training data set was mainly generated by random sampling configurations from molecular dynamics (MD) simulations performed with DPGen, complemented by representative structures extracted from key reactions and collisions, enabling an accurate description of the energetics relevant to C₂H₄-O₂ combustion. The accuracy and transferability of the developed C₂H₄-O₂ MLP were assessed in the NVT ensemble using MD simulations of a system initially composed of 100 C₂H₄ and 300 O₂ molecules at 3000 K. A total of 175 species and 633 reactions were identified, which were subsequently reduced to two simplified reaction networks leading to the formation of H₂O and CO₂. The predicted reaction rates and reaction network show good consistency with established combustion mechanisms. This work provides a basis for extending MLP to the simulation of combustion processes involving more complex fuels.