FlowMM trained on MP-20
A FlowMM model for unconditional crystal generation, trained on the MP-20 dataset. It is used as the prior for the material generation experiments in the NeurIPS 2026 paper ESS-Flow: Training-free guidance as Bayesian inference in source space. The code is at https://github.com/Adhithyan8/essflow.
No pretrained FlowMM weights were released by the original authors, so we trained this model ourselves with FlowMM's default configuration for de novo generation:
python scripts_model/run.py data=mp_20 model=abits_params
The checkpoint is from epoch 1894.
Files
checkpoints/flowmm_mp20.ckpt: PyTorch Lightning checkpoint..hydra/config.yaml: the training configuration, which FlowMM needs to load the checkpoint..hydra/overrides.yaml: the command-line overrides used for training.
Usage
Download the files into the ESS-Flow repository:
hf download Adhi94/essflow-flowmm-mp20 --local-dir pretrained/flowmm_mp20
To load the model directly with FlowMM:
from flowmm.model.eval_utils import load_model
cfg, model = load_model("pretrained/flowmm_mp20/checkpoints/flowmm_mp20.ckpt")
FlowMM finds the configuration by searching the parent directories of the checkpoint for .hydra/config.yaml, so keep the folder structure as it is.
License
FlowMM is released under CC BY-NC 4.0, and so is this checkpoint.