On machine learnability of local contributions to interatomic potentials from density functional theory calculations

Area of Science:

• Computational materials science
• Quantum chemistry
• Machine learning

Background:

• Machine learning interatomic potentials (MLIPs) are advanced classical force fields predicting atomic energies and forces from local environments.
• A key assumption is that atomic contributions are learnable solely from short-range local environments, ensuring scalability and transferability.

Purpose of the Study:

• To challenge the assumption of local atomic contribution learnability in MLIPs.
• To investigate the spatial extent of electron density and electrostatic potential perturbations in various materials.

Main Methods:

• Density functional theory (DFT) calculations were employed.
• Quantified the decay of induced electron density and electrostatic potential in response to local perturbations.
• Analyzed insulating, semiconducting, and metallic samples across different dimensionalities, including molecules, thin layers, and bulk crystals.

Main Results:

• Disturbances in molecules and thin layers are not localized, questioning the learnability of local atomic contributions for MLIPs in these systems.
• Induced electrostatic effects from impurities or vacancies in bulk materials decay slowly, remaining significant beyond nearest neighbors.
• Geometric deformations in bulk materials exhibit localized effects within the first neighbors, inducing a vanishing Yukawa-type potential.

Conclusions:

• The learnability and transferability of MLIPs are questionable for molecules and low-dimensional systems due to non-local effects.
• MLIPs are primarily applicable to bulk materials undergoing purely geometric deformations (e.g., conformational search, thermal properties).
• MLIPs trained on local environments require corrections for long-range electrostatic effects when dealing with chemically significant defects like impurities or vacancies.