What a potential predicts
A machine-learning interatomic potential approximates an energy surface from atomic species and geometry; forces and stress depend on the selected model and derivative implementation. MACE supplies higher-order equivariant potentials, SevenNet supplies a NequIP-based potential package, and CHGNet supplies charge-informed atomistic modeling. MatGL is a library hosting multiple architectures, not one interchangeable checkpoint. A materials potential is an approximation to its training/reference domain, not a replacement for every DFT method or experimental observable.
mace-README.md · seven-README.md · chg-README.md · matgl-README.md
A conditional selection matrix
| Candidate | Documented distinction | Check before selection |
|---|---|---|
| MACE | MACE-MP materials and MACE-OFF organic families; ASE examples | Match model family and reference energies to the system |
| SevenNet | ASE calculator, LAMMPS parallel MD, multiple pretrained families | Exact checkpoint, elements, fidelity and interfaces |
| CHGNet | Materials Project GGA/GGA+U pretraining; magnetic-moment-informed features | Reference conventions and applicability to the structure |
| MatGL | Multiple materials architectures; current PyG implementation | Architecture, checkpoint revision, backend and asset rights |
The matrix is a starting shortlist. Supported elements do not establish coverage of every charge state, bond environment, phase, or temperature. Do not assign every candidate identical energy/force/stress tasks.
mace-README.md · seven-README.md · chg-README.md · matgl-README.md
Align units and reference calculations
CHGNet direct prediction reports energy in eV/atom, force in eV/Å, and stress in GPa. ASE conventions use eV and Å; stress is expressed in energy per volume. Therefore, raw CHGNet output and an ASE calculator output are not interchangeable without checking the adapter. Record total versus per-atom energy, atom order, force sign, stress sign/components, and unit conversions. MACE warns against comparing raw MACE-MP DFT energies directly with compatibility-corrected Materials Project energies. Compare like reference settings before interpreting an error.
chg-README.md · ase-units · mace-README.md
Version and asset gates
At the inspected MatGL revision, DGL support has been removed and PyG is the graph backend; older DGL instructions must not be reused as current compatibility evidence. Its code license does not settle the rights of an individual checkpoint. This guide does not recommend a specific MatGL checkpoint or unblock the held R04 workflow. MACE also documents a transition toward v1.0 while v0.3.x remains supported, with checkpoint conversion required for the planned migration. Pin release and model identities rather than saying “latest.”
matgl-README.md · matgl-LICENSE · mace-README.md
Validation determines the usable domain
Propose a reference set covering equilibrium structures, distortions and relevant phases. Inspect energy differences, atom-resolved force errors, and stress only where supported and required. Keep failures and exclusions visible, and separate selection data from final evaluation. Relaxation convergence and one aggregate error do not establish MD stability or transfer to unfamiliar chemistry. Use the existing atomistic-potential collection for a broader evaluation worksheet; use the SevenNet Recipe for execution details.
Related resources and reading
MACE · SevenNet · CHGNet · MatGL · ASE
Fixed-Cell Crystal Relaxation with SevenNet and ASE
Preparing an atomistic potential evaluation · AI for Materials Science: Tools, Agents and Skills
Sources and evidence boundary
Sources were inspected on 2026-10-08. Revision-pinned project documentation supports capability statements. Evaluation choices are editorial proposals. This article reports no executed workflow, measured performance, or experimental validation.