README.md

metadata

base_model: microsoft/mattergen
library_name: mattergen
tags:
  - diffusion
  - mattergen
  - crystal-generation
  - materials-science
  - transition-metal-oxides
  - energy-above-hull

Simergy Diffusion TMO Alpha

Fine-tuned MatterGen diffusion checkpoint for transition-metal oxide crystal generation. The model was full-finetuned from the MatterGen chemical_system_energy_above_hull checkpoint and is conditioned on:

• chemical_system
• energy_above_hull

Recommended diffusion guidance factor: 9.0.

Files

• config.yaml: MatterGen/Hydra config for loading this checkpoint directory.
• checkpoints/last.ckpt: final training checkpoint, used by MatterGen when checkpoint_epoch="last".
• checkpoints/epoch=59-loss_val=0.36.ckpt: best monitored validation checkpoint.
• metrics.csv: training and validation metrics emitted by Lightning.
• training_overrides.yaml: Hydra overrides used for the finetuning run.
• hparams.yaml: Lightning hparams file.

Usage

Download the repository and pass the snapshot directory as MatterGen model_path:

from huggingface_hub import snapshot_download

model_dir = snapshot_download("HishaamA/simergy-diffusion-tmo-alpha")
print(model_dir)

Then generate with MatterGen:

python -m mattergen.scripts.generate ./results/tmo_samples \
  --model_path="$MODEL_DIR" \
  --checkpoint_epoch=last \
  --properties_to_condition_on="{'chemical_system':'Li-O','energy_above_hull':0.05}" \
  --diffusion_guidance_factor=9.0 \
  --record_trajectories=False

For the best monitored checkpoint, use --checkpoint_epoch=best.

Training Summary

• Base checkpoint: chemical_system_energy_above_hull
• Finetuning mode: full finetuning
• Dataset config: tmo_ehull
• Trainer config: single_gpu
• Max epochs: 100
• Accumulated gradient batches: 8
• Learning rate: 5e-6

Generated structures should be independently screened and validated before any downstream scientific or engineering use.