A multiple-fidelity Method for Accurate Simulation of MoS 2 Properties Using JAX-ReaxFF and Neural Network Potentials

Reactive force field (ReaxFF) is one of the most commonly used force field to model the chemical reactions on atomic level. Recently, JAX-ReaxFF, combined with auto- matic differentiation, has been used to efficiently parameterize ReaxFF. However, pre- dicted properties using parameterized ReaxFF may be inaccurate due to the inductive bias of its analytical formula. While neural network-based potentials (NNPs) trained on density functional theory (DFT)-labeled data offer a more accurate method, it re- quires a large amount of training data to be trained from scratch. To overcome these issues, we present a multiple-fidelity method that combines JAX-ReaxFF and NNP, and apply the method on MoS 2 , a promising two-dimensional (2D) semiconductor for flexible electronics due to its excellent mechanical, optical, and electronic properties. By optimizing ReaxFF for MoS 2 and incorporating implicit prior physical information in the functional forms, we show that ReaxFF can serve as a cost-effective way to generate pretraining data, facilitating more accurate simulations of MoS 2 properties, such as the convex hull diagram, sulfur vacancy formation, and interaction with S 8 using SchNet. Moreover, in the Mo-S-H multi-element system, the pretraining strat- egy can reduce root-mean-square errors(RMSE) of energy by 20%. This approach can be extended to a wide variety of material systems, accelerating their computational research.