A focus on delocalization error poisoning the density-functional many-body expansion

Area of Science:

• Computational Chemistry
• Materials Science
• Quantum Mechanics

Background:

• Density-functional approximations (DFAs) are widely used for electronic structure calculations.
• Machine learning potentials (MLPs) and force fields rely on accurate energy calculations for simulations.
• Many-body expansions (MBEs) are crucial for developing potentials for large-scale simulations.

Purpose of the Study:

• To investigate the behavior of self-interaction errors (SIEs) from DFAs within MBEs.
• To assess the reliability of DFAs for developing accurate MLPs and force fields.
• To highlight the implications of SIEs for large-scale simulations.

Main Methods:

• Explicit decomposition of self-interaction errors (SIEs) from various density-functional approximations (DFAs).
• Analysis of SIEs within the three- and four-body terms of many-body expansions (MBEs).
• Evaluation of the stability and magnitude of MBE terms derived from DFAs.

Main Results:

• Many commonly used DFAs exhibit significant oscillations in their MBE terms, particularly the three-body terms.
• These oscillations can become problematically large, indicating potential unreliability.
• The presence of intrinsic SIEs raises concerns about the transferability of MLPs trained on small systems to large-scale simulations.

Conclusions:

• Widely used DFAs may not be suitable for developing robust MLPs and force fields due to unmanaged self-interaction errors.
• The findings challenge the assumption of stable energetics from DFAs in the context of MBEs.
• Careful consideration of SIEs is essential for the reliable application of MLPs in multiscale modeling.