Accurate Fourth-Generation Machine Learning Potentials by Electrostatic Embedding.

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Machine learning potentials (MLPs) are crucial for atomistic simulations in various scientific fields.
• Current MLPs often rely on local atomic energies, limiting their accuracy.
• Fourth-generation MLPs incorporate long-range electrostatics for improved performance.

Purpose of the Study:

• To investigate the impact of electrostatic potential as a descriptor in MLPs.
• To enhance the quality and transferability of MLPs by incorporating electrostatic information.
• To overcome limitations of traditional descriptors in representing atomic environments.

Main Methods:

• Development of an electrostatically embedded fourth-generation high-dimensional neural network potential (ee4G-HDNNP).
• Augmentation of descriptors with electrostatic potential alongside structural information.
• Utilizing pairwise interactions within the MLP framework.

Main Results:

• Including electrostatic potential significantly improves MLP quality and transferability.
• The extended descriptor resolves limitations of two- and three-body feature vectors for degenerate atomic environments.
• The ee4G-HDNNP accurately predicts energy differences for NaCl clusters and shows transferability to melts.

Conclusions:

• Electrostatic potential is a vital descriptor for advancing MLPs.
• The developed ee4G-HDNNP offers superior accuracy and broader applicability in atomistic simulations.
• This approach paves the way for more reliable computational modeling in chemistry and materials science.