Generator of Neural Network Potential for Molecular Dynamics: Constructing Robust and Accurate Potentials with Active...

Area of Science:

• Computational materials science
• Chemical physics
• Machine learning in chemistry

Background:

• Neural network potentials (NNPs) are crucial for large-scale molecular dynamics (MD) simulations, offering accuracy comparable to ab initio methods.
• However, maintaining simulation stability for long durations remains a challenge due to uncharted potential energy surfaces (PES).
• Existing methodologies lack effective solutions for ensuring NNP stability in extended MD simulations.

Purpose of the Study:

• To develop an automated framework for generating robust and accurate NNPs.
• To enhance the stability and reliability of long-duration MD simulations for organic materials.
• To enable simulations of complex systems exceeding 10,000 atoms for extended timescales.

Main Methods:

• An active learning (AL) framework was developed for the automatic generation of NNPs.
• The framework integrates data set creation, NNP training, evaluation, and structure sampling/screening.
• A novel sampling strategy focuses on unstable structures with short interatomic distances, combined with efficient screening based on interatomic distances and structural features.

Main Results:

• The developed NNP generator enables stable MD simulations of systems with over 10,000 atoms for up to 20 nanoseconds (ns).
• Applied to liquid propylene glycol (PG) and polyethylene glycol (PEG), the generated NNPs demonstrated high simulation stability.
• Predicted physical properties, including density and self-diffusion coefficients, showed excellent agreement with experimental data.

Conclusions:

• This work presents a significant advancement in creating robust and accurate NNPs for organic materials.
• The active learning-based approach overcomes limitations in NNP stability for long-duration MD simulations.
• The methodology paves the way for reliable, large-scale MD simulations of complex molecular systems.