Performance and Cost Assessment of Machine Learning Interatomic Potentials.

Area of Science:

• Materials Science
• Computational Chemistry
• Condensed Matter Physics

Background:

• Machine learning interatomic potentials (ML-IAPs) are advancing materials simulations.
• Developing accurate ML-IAPs requires understanding local atomic environments.

Purpose of the Study:

• To comprehensively evaluate four local environment descriptors for ML-IAPs.
• To assess the performance of ML-IAPs across diverse materials and properties.

Main Methods:

• Utilized atom-centered symmetry functions (ACSF), smooth overlap of atomic positions (SOAP), spectral neighbor analysis potential (SNAP) bispectrum components, and moment tensors.
• Generated a diverse dataset using high-throughput density functional theory (DFT) calculations for metals (Li, Mo, Cu, Ni) and semiconductors (Si, Ge).

Main Results:

• All evaluated descriptors significantly outperformed classical interatomic potentials (IAPs) in predicting energies and forces.
• ML-IAPs accurately predicted material properties like elastic constants and phonon dispersion curves.
• A clear trade-off exists between model accuracy and computational cost, influenced by descriptor complexity.

Conclusions:

• Machine learning interatomic potentials offer a powerful approach for accurate materials modeling.
• Model selection for applications like molecular dynamics should consider the balance between accuracy and computational expense.