Accurate learning of long-range interatomic potentials by coupling Cartesian atomic cluster expansion and...

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Accurate molecular dynamics (MD) simulations require precise modeling of long-range interatomic interactions.
• Machine learning interatomic potentials (MLIPs) offer a promising avenue for efficient and accurate simulations.

Purpose of the Study:

• To integrate the sum-of-Gaussians neural network (SOG-Net) with a short-range descriptor for enhanced MD simulations.
• To evaluate the generalizability and performance of the SOG-Net module when coupled with different short-range descriptors.

Main Methods:

• Development of the CACE-SOG model by combining SOG-Net with the Cartesian atomic cluster expansion (CACE) descriptor.
• Investigation of SOG-Net's technical improvements, including extrapolation accuracy, charge state handling, and convergence speed.
• Evaluation of the CACE-SOG model on diverse systems: molecular dimers, salt solutions, ionic clusters, and interfaces.

Main Results:

• The CACE-SOG model demonstrated accurate fitting of long-range interaction tails with varying decay rates.
• SOG-Net proved to be a versatile module, successfully coupled with the CACE short-range descriptor.
• The model showed improved performance across various challenging systems compared to existing methods.

Conclusions:

• The SOG-Net is a highly effective module for learning long-range interatomic interactions in MD simulations.
• The CACE-SOG model represents a significant advancement in developing accurate and generalizable MLIPs.
• Further technical developments enhance SOG-Net's applicability and efficiency for complex chemical systems.