Benchmarking 34 OpenKIM Nickel Potentials with an Emphasis on Surfaces and Extended Defects

TL;DR

The work addresses the challenge of selecting reliable interatomic potentials for nickel by systematically benchmarking a diverse set of OpenKIM Ni–Ni potentials across an extensive metric suite that includes surfaces, defects, grain boundaries, and vacancy clusters. An automated benchmarking framework computes 47 metrics and compares predictions to DFT benchmarks, using PCA to reveal cross-metric correlations and trade-offs. Key findings show that many potentials reproduce lattice parameters, elastic constants, and surface energies, but migration barriers and short-range interactions remain difficult; SNAP potentials consistently lie on the lowest-error frontier. The publicly available, reproducible framework provides a baseline for potential selection and enables benchmarking-in-the-loop development of next-generation machine-learning Ni potentials for Ni and Ni-based alloys.

Abstract

We present an automated benchmarking suite for face-centered-cubic (FCC) nickel that evaluates 47 quantitative metrics spanning both standard tests (equation of state, elastic constants, surface energies and phonons) and application-specific scenarios such as defect formation and migration, grain boundaries, step edges, close-range interactions, and vacancy cluster energetics. Using this framework, we assess 34 interatomic potentials from the OpenKIM repository, including pairwise, embedded-atom, modified-embedded-atom, angular-dependent, and spectral neighbor analysis potentials (SNAP). Results are compared against ab initio benchmarks compiled from the literature. Most potentials accurately reproduce lattice parameters, elastic constants, and surface energies, whereas predictive accuracy degrades for migration barriers and short-range compression. Principal-component analysis identifies correlated property groups and a partially orthogonal component associated with migration and short-range physics, revealing Pareto trade-offs between accuracy domains. SNAP models occupy the lowest-error frontier, although several embedded-atom potentials remain competitive across many metrics. The framework provides a reproducible baseline for potential selection, highlights systematic limitations across formalisms, and supports benchmarking-in-the-loop strategies for developing next-generation machine-learning potentials for Ni and Ni-based alloys.

Figures (10)

• Figure 1: Representative atomic configurations used in the benchmarking suite: (a) adatom on a (111) surface, (b) stacking-fault tetrahedron formed by ten vacancies, (c) (100)--(111) step edge, (d) $\Sigma9$ tilt grain boundary, (e) (332) free surface, and (f) six-vacancy void.
• Figure 2: Overall directory structure of the benchmarking suite. Folder names prefixed with the symbol '$' represent groups of folders that serve the same function for different tests.
• Figure 3: Workflow of the benchmarking suite on a high-performance computing (HPC) system using a job scheduler. The initial setup job determines the equilibrium lattice parameter and per-atom reference energy. Subsequent tests are then submitted and executed in parallel to minimize total runtime.
• Figure 4: Comparison of interatomic potentials with DFT reference values for free-surface, grain-boundary, and step-edge formation energies. Colors indicate the relative deviation from benchmark values (%). Horizontal striping in the free-surface data highlights strong intra-potential correlations: a potential that reproduces one surface energy accurately tends to reproduce others with similar accuracy.
• Figure 5: Comparison of interatomic potentials with DFT reference values for the formation and migration energies of bulk and surface point defects. Colors indicate the relative deviation from the benchmark values (%). Unlike the surface-energy benchmarks, intra-potential correlations are weak, resulting in noisier patterns across metrics. Migration energies, particularly for surface defects, are poorly reproduced by most potentials.