A Multiple-Fidelity Method for Accurate Simulation of MoS₂ Properties Using JAX-ReaxFF and Neural Network Potentials

Area of Science:

• Computational materials science
• Chemical physics
• Nanotechnology

Background:

• Reactive force fields (ReaxFF) are crucial for atomic-level chemical reaction modeling.
• JAX-ReaxFF offers efficient parametrization but can yield inaccurate predictions due to analytical formulas.
• Neural network potentials (NNPs) provide accuracy but demand extensive training data.

Purpose of the Study:

• To introduce a novel multiple-fidelity method integrating JAX-ReaxFF and NNPs.
• To enhance the accuracy and efficiency of material simulations.
• To apply and validate the method on Molybdenum disulfide (MoS2) for flexible electronics.

Main Methods:

• Developed a hybrid approach combining JAX-ReaxFF and NNPs.
• Utilized ReaxFF with implicit physical information for cost-effective pretraining data generation.
• Applied the method to the Mo-S-H system for simulation and validation.

Main Results:

• The multiple-fidelity method successfully simulated MoS2, a 2D semiconductor.
• Pretraining data generation using ReaxFF significantly improved NNP accuracy.
• Achieved a 20% reduction in root-mean-square energy errors in the Mo-S-H system.

Conclusions:

• The combined JAX-ReaxFF and NNP approach offers a powerful strategy for accurate material simulations.
• This method accelerates computational research by reducing data requirements and improving prediction accuracy.
• The approach is extensible to diverse material systems, advancing materials discovery.