Ab Initio-Trained Machine Learning Molecular Dynamics Model for Radical Reactions in Hydrogen Combustion

Machine learning-based molecular dynamics (ML-MD) has emerged as a powerful tool that bridges the accuracy of quantum mechanical methods with the efficiency of classical molecular dynamics. However, existing ML-MD models often fall short in accurately predicting individual reaction rates due to incomplete sampling of reaction pathways and the absence of a multifidelity framework that combines data of varying precision. In this work, we develop and validate integrated ab initio-trained ML-MD (aML-MD) models for two key radical reactions in hydrogen combustion: H + HO₂ → H₂ + O₂ and O + HO₂ → O₂ + OH. Using the DeePMD-kit framework, we first construct an aML-MD model trained solely on high-accuracy ab initio data and demonstrate its ability to predict individual reaction rates across distinct pathways via quasi-classical trajectory (QCT) calculations with excellent agreement to PES-based results. We then apply transfer learning to build another aML-MD model that incorporates a large set of moderate-accuracy DFT/PES data and a small subset of ab initio data. This multifidelity approach reduces computational cost by over fivefold while maintaining predictive accuracy within 20% of PES-based benchmarks. Further NEB simulations indicate that transfer learning yields a more accurate reaction barrier, thereby enhancing the model’s fidelity in capturing the underlying reaction dynamics. The results demonstrate the effectiveness of combining multifidelity data and transfer learning to enable general-purpose ML-MD models capable of accurately capturing complex reaction dynamics, with promising applications to larger, multi-reaction chemical systems.