Electron Alchemy with Machine-Learned Interatomic Potentials: Case Studies of Local Charge in Bond Dissociation...

Area of Science:

• Computational Chemistry
• Materials Science
• Quantum Mechanics

Background:

• Machine-learned interatomic potentials (MLIPs) aim to achieve density functional theory (DFT) accuracy at lower computational cost.
• Current MLIPs show high average fidelity but suffer from failure modes like spurious bond formation and inconsistent long-range interactions.

Purpose of the Study:

• Investigate the origins of MLIP failure modes, focusing on spurious bond formation.
• Benchmark different MLIP models (UMA, ORB, MACE, AIMNet2) against DFT reference data.
• Analyze the impact of atomic charge representation on MLIP accuracy.

Main Methods:

• Benchmarking MLIPs against DFT bond dissociation curves for various species.
• Comparing models with and without explicit atomic charge resolution.
• Evaluating agreement with restricted (rDFT) and unrestricted (uDFT) DFT energies.

Main Results:

• MLIPs lacking explicit charge resolution incorrectly predict stable bonds between like-charged halide anions.
• Models with atomic partial charge equilibration accurately predict repulsion in these systems.
• Inconsistencies were observed in agreement with uDFT/rDFT energies and core-region treatment.

Conclusions:

• Spurious bond formation in MLIPs is intrinsically linked to the inability of global charge specification to differentiate local geometries across different charge and spin states.
• Artifacts from core repulsion and electrostatics are addressable through improved physical priors and data curation.
• Charge representation is critical for accurate MLIP simulations, especially for ionic systems.