name	machine_learning_potentials
description	Router for MLIP skill bundles. When a task involves machine learning interatomic potentials, consult the appropriate subdirectory skill based on the backend in use or the system's chemistry. Start with general/ for strategy and backend selection, then load the matching backend skill.

Machine Learning Potentials — Skill Directory

This directory contains skill bundles for machine learning interatomic potential (MLIP) backends. Each subdirectory is a self-contained skill.

Skill index

Subdirectory	Skill file	When to use
`general/`	`general.md`	Strategy, backend selection, active learning, cross-backend validation thresholds. Load this alongside any backend skill.
`mace/`	`mace.md`	MACE equivariant MLIP — default for inorganic systems (`mace-omat-0`) and organic molecules (`mace-off23`). Supports ASE and LAMMPS `pair_style`.
`chgnet/`	`chgnet.md`	CHGNet — best choice for magnetic systems (Fe, Co, Ni, Mn, Cr); predicts magnetic moments. ASE only (no LAMMPS `pair_style`).
`deepmd/`	`deepmd.md`	DeePMD — mature framework with first-class LAMMPS support (`pair_style deepmd`) and universal foundation models (DPA-2, DPA-3).
`uma/`	`uma.md`	UMA (Meta/FAIR) — state-of-the-art universal MLIP with multi-fidelity training. Supports ASE and LAMMPS.
`orb/`	`orb.md`	Orb (Orbital Materials) — fast universal MLIP optimised for high-throughput screening. ASE only (no LAMMPS `pair_style`).

Selection guidance

Always load general/general.md alongside whichever backend skill is active — it supplies the pretrained-first strategy, active learning loop, and cross-backend acceptance thresholds that apply universally.
Choose a backend skill using this priority order:
- Magnetic systems (Fe, Co, Ni, Mn, Cr) → chgnet/
- Organic / drug-like molecules → mace/ (mace-off23) or uma/ (omol)
- LAMMPS production runs at scale → mace/ or deepmd/
- Speed-critical high-throughput screening → chgnet/ or orb/
- State-of-the-art accuracy → uma/
- Default inorganic / mixed systems → mace/ (mace-omat-0) or uma/ (omat)
If the backend is already determined by the environment (e.g., a dp binary is present → DeePMD; mace Python module is importable → MACE), go directly to the matching skill.

Machine Learning Potentials — Skill Directory

This directory contains skill bundles for machine learning interatomic potential (MLIP) backends. Each subdirectory is a self-contained skill.

Skill index

Subdirectory

Skill file

When to use

general/

general.md

Strategy, backend selection, active learning, cross-backend validation thresholds. Load this alongside any backend skill.

mace/

mace.md

MACE equivariant MLIP — default for inorganic systems (mace-omat-0) and organic molecules (mace-off23). Supports ASE and LAMMPS pair_style.

chgnet/

chgnet.md

CHGNet — best choice for magnetic systems (Fe, Co, Ni, Mn, Cr); predicts magnetic moments. ASE only (no LAMMPS pair_style).

deepmd/

deepmd.md

DeePMD — mature framework with first-class LAMMPS support (pair_style deepmd) and universal foundation models (DPA-2, DPA-3).

uma/

uma.md

UMA (Meta/FAIR) — state-of-the-art universal MLIP with multi-fidelity training. Supports ASE and LAMMPS.

orb/

orb.md

Orb (Orbital Materials) — fast universal MLIP optimised for high-throughput screening. ASE only (no LAMMPS pair_style).

Selection guidance

Always load general/general.md alongside whichever backend skill is active — it supplies the pretrained-first strategy, active learning loop, and cross-backend acceptance thresholds that apply universally.

Choose a backend skill using this priority order:

Magnetic systems (Fe, Co, Ni, Mn, Cr) → chgnet/
Organic / drug-like molecules → mace/ (mace-off23) or uma/ (omol)
LAMMPS production runs at scale → mace/ or deepmd/
Speed-critical high-throughput screening → chgnet/ or orb/
State-of-the-art accuracy → uma/
Default inorganic / mixed systems → mace/ (mace-omat-0) or uma/ (omat)

If the backend is already determined by the environment (e.g., a dp binary is present → DeePMD; mace Python module is importable → MACE), go directly to the matching skill.

machine-learning-potentials