Learning QM/MM potential using equivariant multiscale model.

Area of Science:

• Computational chemistry
• Materials science
• Quantum mechanics

Background:

• Machine learning (ML) models are efficient for electronic structure theory but limited to closed systems.
• Existing ML models struggle to capture external potentials and polarization effects from the environment.
• High-order electrostatic interactions are crucial for accurate modeling of chemical systems in external fields.

Purpose of the Study:

• To develop a novel ML model that accurately accounts for external potentials, polarization, and long-range interactions.
• To improve the prediction accuracy and transferability of ML surrogate models for electronic structure calculations.
• To incorporate high-order electrostatic terms and equivariant representations into ML models.

Main Methods:

• Incorporated high-order terms of the Taylor expansion of an electrostatic operator into the ML model.
• Utilized an equivariant model generating rotationally covariant high-order tensors.
• Applied multipole-expansion equations to represent polarization and intermolecular interactions.
• Adapted strategies for deriving long-range interactions between systems and environment media.

Main Results:

• The novel ML model demonstrates higher prediction accuracy compared to existing methods.
• The model exhibits improved transferability across diverse environmental media.
• Accurate representation of external potentials and electrostatic environments was achieved.

Conclusions:

• The proposed ML model effectively addresses limitations in modeling chemical systems within external electrostatic environments.
• The inclusion of high-order electrostatic terms and equivariant representations enhances predictive power.
• This approach advances the application of ML in computational chemistry for complex systems.