MLIP models

Machine-learned potential weights are downloaded at runtime. They are not stored in git and are not included in the Python wheel.

Catalog

Id	Runtime	Approximate size	Description
nequix-mp-1	Nequix	2.8 MB	MP-1 universal potential
nequix-mp-1-pft	Nequix	2.8 MB	MP-1 PFT variant
nequix-mp-1-pft-nocotrain	Nequix	2.8 MB	MP-1 PFT without co-training
nequix-oam-1	Nequix	2.9 MB	OAM-1
nequix-oam-1-pft	Nequix	2.8 MB	OAM-1 PFT
nequix-omat-1	Nequix	2.8 MB	OMat-1
nequip-s	NequIP	5.6 MB	NequIP-OAM small
nequip-l	NequIP	71.8 MB	NequIP-OAM large
equiformer	Equiformer	559.9 MB	Equiformer V3 OMat24 direct
equiformer-gradient	Equiformer	486.2 MB	Equiformer V3 OMat24 gradient

Select a catalog model

from atomli.calculators import MLIP

atoms.calc = MLIP("nequix-mp-1")

The first construction downloads the model. Later constructions reuse the cached file.

Inspect the catalog

from atomli.models import default_models_dir, ensure_model, list_models

print(list_models())
print(list_models(runtime="nequip"))
print(default_models_dir())
path = ensure_model("nequix-mp-1")

Set ATOMLI_MODELS_DIR to choose a cache location. Otherwise Atomli uses $XDG_CACHE_HOME/atomli/models when that variable is set, and the platform cache directory when it is not: ~/Library/Caches/atomli/models on macOS, ~/.cache/atomli/models on Linux and other platforms.

Verification and delivery

Atomli checks the local cache first. If the file is not in the cache, Atomli downloads it from tako.atom.li. If that download fails, Atomli uses the mlip-models GitHub Release as the backup source. The same catalog delivers the SKALA 1.1 QC checkpoint for QC(xc="SKALA-1.1").

Each catalog entry pins a SHA-256 digest and byte size. Cached files are reused only when the digest matches. Downloads use a temporary partial file and become visible at the final cache path only after completion.

Prefetch for offline work

From the Atomli repository:

./scripts/download-mlip-models.sh

Prefetch uses the same cache path as MLIP(). In managed environments, set ATOMLI_MODELS_DIR before prefetching and at runtime.

Runtime selection

Catalog ids select their recorded runtime automatically. For a direct local path, the MLIP calculator guide explains suffix inference, explicit runtime selection, stress, and device behavior.