Foundational Model

This tutorial demonstrates how to load a pretrained foundational model and attach it as an ASE calculator.

For more advanced topics—such as fine-tuning, or alternative interfaces, please refer to the corresponding tutorials.

Model Selection

The models below are ordered from efficiency to accuracy.

Model	Indicative system size on a single 80 GB GPU
TACE-OMat24-7M	About 70,000 atoms (EquivariantX kernel)
TACE-OMat24-L	About 90,000 atoms (EquivariantX kernel)
TECE-OMat24-RRA-1.0	About 30,000 atoms (EquivariantX kernel)

Enable a supported backend as described in Acceleration when comparing performance.

Model Overview

TACE uses atomic cluster expansion; TECE additionally uses edge cluster expansion. RRA denotes radial rotary complex attention. OMat24 models are trained on OMat24, whereas OAM models are subsequently trained on sAlex and MPtrj.

Model	Size	Training data	Required TACE version
TACE-OMat24-7M	M	OMat24	>=0.2.0
TACE-OAM-7M	M	OMat24 → sAlex + MPtrj	>=0.2.0
TECE-OMat24-RRA-1.0	XL	OMat24	>=0.2.0
TECE-OAM-RRA-1.0	XL	OMat24 → sAlex + MPtrj	>=0.2.0
TACE-OMat24-RRA-1.0	XL	OMat24	==0.2.0
TACE-OAM-RRA-Preview	XL	OMat24 → sAlex + MPtrj	==0.2.0
TACE-OMat24-L	L	OMat24	>=0.2.0
TACE-OAM-L	L	OMat24 → sAlex + MPtrj	>=0.2.0
TACE-v1-OMat24-M	M	OMat24	==0.1.0
TACE-v1-OAM-M	M	OMat24 → sAlex + MPtrj	==0.1.0
TACE-v1-LES-REICO-5-PdAgCHO	M	REICO-5-PdAgCHO	==0.1.0

Dataset Overview

Dataset	Domain and coverage	Level of theory
OMat24	Inorganic bulk materials; non-equilibrium structures and relaxation trajectories	PBE+U
sAlex	Inorganic crystals; subsampled Alexandria relaxation trajectories	PBE+U
MPtrj	Inorganic crystals; Materials Project relaxation trajectories	PBE+U
MatPES PBE	Inorganic materials; equilibrium structures and MD-sampled configurations	PBE, without Hubbard U
MatPES r²SCAN	Inorganic materials; equilibrium structures and MD-sampled configurations	r²SCAN, without Hubbard U
REICO-5-PdAgCHO	Heterogeneous catalysis; Pd-Ag catalysts and C/H/O-containing species	PBE+D3

Model Download and Cache

Pass a registered model name directly to load_tace, TACEAseCalc, TACETorchSimCalc, or a command’s --model argument. The weights are downloaded automatically and cached locally. Files ending in .pt, .pth, .ckpt, or .pt2 are loaded as local files instead.

from tace.lightning import load_tace

model = load_tace("TACE-OAM-7M", device="cuda", dtype="float32")

tace-eval -m TACE-OAM-7M -i structures.xyz
tace-export-eval -m TACE-OAM-7M --backend aoti --device cuda

Names must match the registry exactly, including version suffixes such as TECE-OAM-RRA-1.0. An unknown name raises an error and lists available models.

By default, all models are stored under:

~/.cache/tace/

If your network connection is unstable or restricted, you may manually download the pretrained models from the TACE model collection on Hugging Face. The model weights are distributed under CC BY 4.0.

For automatic loading, place the checkpoint directly under ~/.cache/tace/ using the filename expected by the registry. For example, TACE-OAM-7M uses ~/.cache/tace/TACE-OAM-7M.pt. To list the registry keys supported by your installation:

from tace.foundations import tace_foundations

print(tace_foundations.list_models())

For a release not listed in the registry, download its checkpoint from Hugging Face and pass its local path as model to TACEAseCalc.

Minimal ASE Example

Below is a minimal working example showing how to use a TACE Foundational Model as an ASE calculator:

import torch
from ase.io import read
from tace.interface.ase import TACEAseCalc, add_dispersion

# Load a pretrained foundational model
# The model will be auto-downloaded to ~/.cache/tace if not present
model = "TACE-OAM-7M"

dtype = "float32"
device = "cuda" if torch.cuda.is_available() else "cpu"

# Fidelity fidelity_idx (0 corresponds to the first fidelity)
fidelity_idx = 0

atoms = read("../unrelaxed.xyz", index=0)
calc = TACEAseCalc(
    model=model,
    dtype=dtype,
    device=device,
    fidelity_idx=fidelity_idx,
)
atoms.calc = calc

Dispersion Correction (Optional)

Dispersion interactions can also be supported by calling third-party libraries. For detailed instructions, see ase guide.

Honors and Milestones

• 2026-07-08 — Matbench Discovery SOTA. TECE-OAM-RRA-1.0 ranked first on the default Matbench Discovery leaderboard by Combined Performance Score (CPS), with a score of 0.908.

Matbench Discovery default ranking as of July 8, 2026.