GPU-accelerated data-driven framework of hybrid ReaxFF

In computational materials science, simulation techniques must balance bridging time and length scales with maintaining high accuracy at a reasonable computational cost. In this study, a data-driven parametrization method through machine learning algorithms is proposed. The proposed method aims to decrease the time required for parameter optimization, enhancing efficiency of ReaxFF potential. This innovative approach employs a combination of reactive and non-reactive molecular dynamics simulations to simulate phenomena that demand extended time scales or involve larger systems beyond the conventional capabilities of ReaxFF. ML algorithms are utilized between the reactive and non-reactive stages to forecast non-reactive forcefield parameters by considering the updated bond topology of the system. The proposed algorithm can be accelerated on the GPU, achieving optimized management of the computing power and memory requirements imposed by ReaxFF MD on computer hardware. This study is envisioned to promote the application of ReaxFF in large and complex material systems which aims to provide more efficient and accurate predictions compared to traditional ReaxFF.