Electrostatic Embedding of Machine Learning Potentials.

Area of Science:

• Computational Chemistry
• Machine Learning in Chemistry
• Molecular Modeling

Background:

• Accurate molecular simulations are crucial for drug discovery and materials science.
• Existing methods often require extensive training data or complex setups.
• Developing efficient and generalizable embedding schemes is an ongoing challenge.

Purpose of the Study:

• To present a novel electrostatic embedding scheme for machine learning potentials.
• To enable the use of arbitrary machine learning potentials trained on isolated molecular systems.
• To create a generic and data-efficient model for molecular simulations.

Main Methods:

• Developed a physically motivated electrostatic embedding scheme using electronic density and polarizability models.
• The scheme requires only in vacuo single-point Quantum Mechanics (QM) calculations for training.
• Applied Gaussian Process Regression to create an embedding model for the QM7 dataset using limited atomic environments.

Main Results:

• Successfully created an embedding model for the QM7 dataset with only 445 reference atomic environments.
• The model demonstrated high accuracy when applied to the SARS-CoV-2 protease complex with PF-00835231.
• Achieved a root-mean-square error (RMSE) of 2 kcal/mol for predicted embedding energy, comparable to explicit Density Functional Theory/Molecular Mechanics (DFT/MM) calculations.

Conclusions:

• The proposed electrostatic embedding scheme is generic, data-efficient, and accurate.
• This approach facilitates the use of machine learning potentials for complex molecular systems.
• The method shows promise for accelerating drug discovery and materials science research.