CyrilPKU/CrystalREPA-Models

CrystalREPA Models

Pretrained checkpoints for CrystalREPA: Transferring Physical Priors from Universal MLIPs to Crystal Generative Models.

Accepted at NeurIPS 2026.

Paper · Code · Data

CrystalREPA aligns atom-wise representations of crystal generative models with those of frozen machine-learning interatomic potentials (MLIPs), using an element-aware contrastive objective (EA-NCE).

This release provides one MatterGen baseline and ten CrystalREPA models trained on MP-20. Sampling requires neither the teacher models nor their cached representations.

Generate crystals

Install the environment following the code repository. Run the following commands from its root directory.

Download the model trained with DPA-3.1-3M as the teacher:

hf download CyrilPKU/CrystalREPA-Models \
  --repo-type model \
  --include "mattergen/dpa_3_1_3m/*" \
  --local-dir checkpoints

Generate 1,024 crystals:

python scripts/mattergen.py generate \
  --model_path=checkpoints/mattergen/dpa_3_1_3m \
  --checkpoint_epoch=1899 \
  --output_path=outputs/generated/dpa_3_1_3m \
  --batch_size=64 \
  --num_batches=16 \
  --record_trajectories=False

Outputs include EXTXYZ structures and a ZIP archive of CIF files. For a quick trial, set --num_batches=1.

Sampling does not require downloading the training dataset. If GPU memory is limited, reduce the batch size and increase the number of batches to preserve the total sample count.

Available models

All checkpoints correspond to epoch 1899, after 1,900 training epochs.

Directory under mattergen/	MLIP teacher
baseline	None
dpa_3_1_3m	DPA-3.1-3M
dpa_3_1_3m_ft	DPA-3.1-3M-FT
dpa_3_1_mptrj	DPA-3.1-MPtrj
mace_mpa_0	MACE-MPA-0
mace_mp_0	MACE-MP-0
orb_v3	ORB v3
orb_v2	ORB v2
sevennet_omni_i12	SevenNet-Omni-i12
sevennet_l3i5	SevenNet-l3i5
mattersim_v1_5m	MatterSim-v1-5M

DPA-3.1-3M-FT denotes DPA-3.1-3M further fine-tuned on MPtrj.

Each model directory contains:

config.yaml
checkpoints/epoch=1899.ckpt

These are trained MatterGen checkpoints, not the original MLIP teacher weights. Each checkpoint includes model parameters, optimizer/scheduler states and training progress, and is approximately 465 MB.

To use another model, replace dpa_3_1_3m in the download and generation commands with its directory name.

Download all MatterGen models:

hf download CyrilPKU/CrystalREPA-Models \
  --repo-type model \
  --include "mattergen/*" \
  --local-dir checkpoints

The file mattergen/models.json provides model metadata and checkpoint SHA256 checksums. Paths inside the catalog are relative to the code repository's layout.

Training and reproducibility

The REPA models use EA-NCE with alignment weight 1.0, temperature 0.1 and alignment layer 2. The baseline uses no representation alignment.

Training data and precomputed teacher representations are available in CrystalREPA-Data. Training from the cached representations does not require running the teacher models.

See the training instructions and the MatterGen documentation for training and continuation from a checkpoint.

Preprocessing: the released baseline and dpa_3_1_3m checkpoints were trained with primitive-cell preprocessing. The other nine checkpoints used the preprocessing of the released shared cache, which applies Niggli reduction without primitive-cell conversion. This distinction is recorded in training_preprocessing_matches_release in the model catalog.

Evaluation and intended use

Evaluation uses the MatterGen evaluation pipeline. Relaxation and reference-based metrics require the corresponding potential and reference datasets.

These models are intended for research on inorganic crystal generation. Generated structures are candidates: generation alone does not establish thermodynamic stability or experimental synthesizability. See the paper for evaluation results and methodology.

Citation

The paper has been accepted at NeurIPS 2026. The camera-ready version is forthcoming; the URL below points to the arXiv preprint.

@inproceedings{zhang2026crystalrepa,
  title={CrystalREPA: Transferring Physical Priors from Universal MLIPs to Crystal Generative Models},
  author={Zhang, Chengqian and Jin, Yucheng and Zhang, Duo and Li, Tiejun and Wang, Han},
  booktitle={Advances in Neural Information Processing Systems},
  year={2026},
  note={Accepted at NeurIPS 2026; camera-ready version forthcoming},
  url={https://arxiv.org/abs/2605.08960}
}

Acknowledgments and license

The released MatterGen checkpoints are provided under the MIT license.

This implementation builds on Microsoft MatterGen. CrystalREPA is inspired by REPA. We thank their authors and the developers of the MLIP teachers for their work and open-source contributions.

Training data and upstream teacher models retain their respective licenses.