Machine Learning Driven ReaxFF Optimization for Combustion

Accurate simulation of combustion reactions is crucial for understanding combustion mechanisms. Reactive force fields (ReaxFF) offer a computationally efficient approach for simulating complex combustion processes, but their accuracy depends critically on parameterization. This work presents a comprehensive optimization of ReaxFF parameters for gas-phase combustion reactions using a machine learning driven approach. We constructed a dataset of 33 reactions, encompassing key reaction types in combustion. High-level double hybrid DFT calculations served as a benchmark to evaluate the performance of various density functionals, the semi-empirical PM7 method, and existing ReaxFF parameter sets. We then employed the JAX-ReaxFF framework to optimize the CHO2008 parameters, leveraging its efficient local gradient-based optimization algorithms. The optimized ReaxFF significantly improved the accuracy of potential energy and atomic force predictions, with the mean absolute error (MAE) for energy approaching that of PM7. Analysis of reaction pathways and potential energy surfaces further demonstrated the enhanced performance of the optimized force field, particularly near transition states. This optimized ReaxFF provides a robust and accurate tool for simulating a wide range of combustion systems, and the presented methodology offers a general strategy for developing system-specific ReaxFF parameters.