A Universal Augmentation Framework for Long-Range Electrostatics in Machine Learning Interatomic Potentials

Area of Science:

• Computational materials science
• Machine learning
• Quantum chemistry

Background:

• Current machine learning interatomic potentials (MLIPs) often neglect long-range electrostatic interactions.
• Accurate modeling of electrostatics is crucial for many chemical and physical phenomena.

Purpose of the Study:

• To present the Latent Ewald Summation (LES) method as a standalone library for integrating long-range electrostatics into existing MLIPs.
• To demonstrate the effectiveness of LES in improving the accuracy and capabilities of various MLIPs.

Main Methods:

• Developed the Latent Ewald Summation (LES) method to infer electrostatic interactions, polarization, and Born effective charges (BECs) from energy and force data.
• Integrated LES with several short-range MLIPs (MACE, NequIP, Allegro, CACE, CHGNet, UMA).
• Trained a universal MLIP (MACELES-OFF) on a large, diverse dataset (SPICE) including organic molecules and clusters.

Main Results:

• LES successfully captures correct electrostatics and enhances the accuracy of MLIPs across diverse systems like water, dipeptides, and surface adsorption.
• The MACELES-OFF model, enhanced by LES, shows superior accuracy compared to its short-range counterpart (MACE-OFF).
• MACELES-OFF reliably predicts dipoles and BECs, offering improved descriptions of bulk liquids and enabling electrostatic MLIPs for organic and biomolecular systems.

Conclusions:

• The LES method provides an efficient way to incorporate long-range electrostatics into MLIPs without direct training on electrical properties.
• LES enables the development of more accurate and versatile MLIPs, paving the way for electrostatic foundation models in materials science and chemistry.