Efficient Training of Neural Network Potentials for Chemical and Enzymatic Reactions by Continual Learning

Area of Science:

• Computational Chemistry
• Machine Learning in Chemistry

Background:

• Machine learning (ML) methods offer efficient alternatives to traditional electronic structure theory for molecular modeling.
• Current ML force fields struggle with extrapolation and transferability due to limited training data and focus on closed systems.
• Existing models perform poorly in complex condensed-phase environments like enzymatic reactions, incurring high computational costs.

Purpose of the Study:

• To enhance the transferability and applicability of ML force fields to complex, heterogeneous environments, including enzymatic reactions.
• To develop an efficient and autonomous training strategy for ML force fields using continual learning.
• To create a versatile force field applicable across various chemical reaction media.

Main Methods:

• Developed a Machine Learning/Molecular Mechanics (ML/MM) model utilizing Taylor expansion of the electrostatic operator.
• Extended the ML/MM model's transferability to complex environments, specifically enzymatic reactions.
• Implemented continual learning strategies with memory datasets for on-the-fly training with new data.

Main Results:

• The ML/MM model demonstrated high transferability across simple solvent systems.
• The extended model showed promising transferability in more complex heterogeneous environments, including enzymatic reactions.
• The combined approach of transferability and continual learning facilitates efficient force field construction for diverse applications.

Conclusions:

• The integration of ML/MM models with Taylor expansion and continual learning significantly improves force field transferability and applicability.
• This strategy enables the development of robust force fields capable of handling chemical reactions in various environmental media.
• The findings pave the way for more general and efficient computational chemistry simulations.