Development of Robust Neural Network Potentials for Large-Scale Atomistic Simulations of Organic Systems

Neural network potentials (NNPs) facilitate large-scale molecular dynamics (MD) simulations with over 10000 atoms, achieving ab initio accuracy and playing a crucial role in material studies. However, maintaining stability in long-term simulations is challenging due to the exploration of unknown potential energy surface (PES) regions. To address this, we developed an automatic NNP generator using an active learning (AL) framework. This generator integrates initial dataset creation, NNP training, evaluation, conformation generation, screening, and labeling. Our approach includes a sampling strategy focused on reducing specific interatomic distances and a screening strategy for efficient configuration sampling. This enhances MD simulation stability, enabling nanosecond-scale simulations. We validated our NNP generator on liquid propylene glycol (PG) as a simple system, and polyethylene glycol (PEG) and nafion for complex polymer systems, achieving stable MD simulations of systems with over 10000 atoms for 20 ns. The predicted dynamic properties, such as density and self-diffusion coefficient, closely match experimental values. This work advances robust and accurate NNP generation for organic materials, enabling long-duration MD simulations of complex systems.