Skip to content

A comprehensive assessment of empirical potentials for carbon materials

Scope

Benchmarks carbon bond-order potentials (CBOPs) and the GAP-20 machine-learning potential against DFT across lattice data, cohesion, defects, vdW interactions, thermal stability, and mechanical response for several carbon allotropes.

Summary

The article systematically compares widely used empirical carbon bond-order potentials and the Gaussian approximation potential GAP-20 (machine-learned on carbon) to density functional theory references. Properties examined include lattice constants, cohesive energies, defect formation energies, van der Waals interactions, thermal stability, and mechanical behavior across different carbon phases. The authors highlight that potential choice strongly affects predicted allotrope behavior, with GAP-20 generally tracking DFT more closely than the assessed CBOPs for structure, defects, and several thermal/mechanical tests, while still sharing CBOP-like limitations on van der Waals treatment.

The benchmark is motivated by widespread use of CBOPs in carbon materials modeling where transferability across sp, sp², and sp³ bonding environments is not guaranteed, and by interest in whether ML potentials can reduce systematic errors without hand-tuned bond-order forms.

Methods

MD application (classical CBOPs and GAP-20). The study runs large-scale molecular dynamics with several carbon bond-order potentials (CBOPs: Tersoff/REBO/AIREBO family members and related forms as listed in the paper) and with the machine-learned GAP-20 potential on the same benchmark geometries to compare thermal stability (e.g. C\(_{60}\) heating tests) and mechanical response (e.g. nanotube fracture, graphene elastic tests in the figures). Ensemble uses NVT/NPT as stated for each thermo/stress case; timestep (fs), K-ramped T, ps–ns trajectory lengths, and P (bar) in NPT segments are documented in the computational section of APL Mater. 9, 061102—N/A to transcribe here. LAMMPS-style inputs are typical for the field; N/A if the article names a different engine—verify pdf_path. Electric field / rare-event enhanced sampling: N/A in the abstract-level summary.

Static QM / DFT (reference data). DFT supplies lattice parameters, cohesive energies, defect formation energies, and other static properties used as the reference for comparing CBOPs and GAP-20, including assessment of interlayer van der Waals-sensitive quantities (see abstract). Functional, dispersion correction choice, basis/PAW/plane-wave settings, and k-point meshes are in the main text—N/A in this wiki note (open pdf_path Section 2). Geometry relaxations and, where used, reaction-path or barrier protocols are spelled out there; N/A to re-list every NEB image count here.

ReaxFF/FF training. N/A — benchmark only.

Findings

Comparisons to DFT. GAP-20 more closely matches the DFT reference for structural and defect energetics across the crystalline allotropes and cases tabulated than the CBOPs surveyed in the work. For finite-T and mechanical tests highlighted in the abstract, GAP-20 reproduces C\(_{60}\) thermal stability and nanotube/graphene fracture or elastic trends where CBOPs are reported to struggle. N/A — the article is not a kinetics database for reactive combustion chemistry; scope is all-carbon benchmarks in the manuscript.

Limitations (authored). Like CBOPs, GAP-20 does not fully capture van der Waals interactions in the authors’ assessment; see the PDF for where that matters (layered stacking, fullerene crystals, etc.).

Sensitivity. Property-by-property trends differ across allotropes; potential choice is a strong lever (pressure in NPT stretches, T in NVT heating protocols).

Limitations

Benchmark scope is defined by the potentials and carbon systems explicitly included; reactive chemistry and hetero-elements are not the focus.

Relevance to group

External benchmark paper in corpus context: situates classical and ML carbon potentials relative to DFT for method selection in carbon MD.

Citations and evidence anchors

  • reaxff-family (context only—ReaxFF is not the subject of this benchmark)