Incorporating Coulomb interactions with fixed charges in moment tensor potentials and equivariant tensor network...

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning in Physics

Background:

• Short-range machine-learning interatomic potentials (MLIPs) often struggle with accurately modeling electrostatic interactions.
• Accurate modeling of charged molecules is crucial for understanding chemical reactions and material properties.

Purpose of the Study:

• To enhance MLIPs by incorporating long-range electrostatic interactions using the Coulomb model with fixed charges.
• To improve the accuracy of MLIPs for charged organic molecules and their binding properties.

Main Methods:

• Incorporation of the Coulomb model with fixed charges into the functional form of moment tensor potentials and equivariant tensor network potentials.
• Training and validation of the enhanced MLIPs on datasets of organic dimers of charged molecules.
• Comparison of MLIP predictions with results from density functional theory (DFT).

Main Results:

• Explicit inclusion of Coulomb interactions reduced energy fitting errors by over four times for short-range MLIPs.
• The developed long-range MLIPs demonstrated significant improvements in predicting binding curves for charged organic dimers.
• MLIP results showed good agreement with DFT calculations for the studied systems.

Conclusions:

• Integrating long-range electrostatic interactions is a highly effective strategy for improving MLIP accuracy, especially for charged systems.
• The enhanced MLIPs provide a computationally efficient and accurate alternative to DFT for studying charged organic molecules.
• This work paves the way for more reliable molecular simulations involving charged species.