Reducing the Cost of Neural Network Potential Generation for Reactive Molecular Systems.

Area of Science:

• Computational Chemistry
• Materials Science
• Molecular Dynamics

Background:

• Machine learning potentials (MLPs) have advanced molecular simulations.
• Constructing robust training sets for MLPs is computationally expensive, often requiring extensive ab initio calculations.
• The need for comprehensive geometry coverage limits MLP applicability for complex systems.

Purpose of the Study:

• To develop a computationally efficient method for generating training sets for MLPs.
• To mitigate the high cost associated with reference ab initio simulations for training set generation.
• To enable accurate molecular simulations for systems with complex potential energy landscapes.

Main Methods:

• Introduced transition tube sampling to generate relevant geometries around a transition path.
• Utilized sparse local normal mode expansions for thermal geometry generation.
• Employed an active learning protocol for efficient geometry selection, creating a focused training set.

Main Results:

• The developed method generates training sets that capture essential transition path geometries without costly reference trajectories.
• MLPs trained with this method perform comparably to those trained with full ab initio data.
• The approach provides significant computational speedup, characteristic of ML potentials.

Conclusions:

• Transition tube sampling effectively reduces the computational burden of training set generation for MLPs.
• The method yields accurate ML potentials suitable for both classical and path integral simulations.
• This technique enhances the practicality and efficiency of applying ML potentials to diverse molecular systems.