Skip to content

Latest commit

 

History

History
78 lines (59 loc) · 3.59 KB

README.md

File metadata and controls

78 lines (59 loc) · 3.59 KB

Model Card

Alex-20

The pre-trained model is available on Google Drive and Hugging Face Model Hub.

Model Parameters

params, transformer = make_transformer(
        key=jax.random.PRNGKey(42),
        Nf=5,
        Kx=16,
        Kl=4,
        n_max=21,
        h0_size=256,
        num_layers=16,
        num_heads=16,
        key_size=64,
        model_size=64,
        embed_size=32,
        atom_types=119,
        wyck_types=28,
        dropout_rate=0.1,
        attn_rate=0.1,
        widening_factor=4,
        sigmamin=1e-3
)

Training dataset

Alex-20: contains ~1.3M general inorganic materials curated from the Alexandria database, with $E_{hull} < 0.1$ eV/atom and no more than 20 atoms in unit cell. The dataset can be found in the Google Drive or Hugging Face Datasets.

Alex-20 RL

  • $E_{hull}$ reward: The checkpoint is available on Google Drive and Hugging Face Model Hub. The reward is chosen to be the negative energy above the hull, which is calculated by the Orb model based on the Alexandria convex hull.

  • Dielectric FoM Reward: The checkpoint is available on Google Drive and Hugging Face Model Hub. The reward is chosen to be figures of dielectric figure of merit (FoM), which is the product of the total dielectric constant and the band gap. We use the pretrained MEGNet to predict the band gap. The checkpoint of the total dielectric constant prediction model can be found in the Google Drive. You can load the model using matgl package.

MP-20

Important

The load the MP-20 checkpoint, you need to switch the CrystalFormer to version 0.3 The current version of the model is not compatible with the MP-20 checkpoint.

Checkpoint

The pre-trained model is available on Google Drive and Hugging Face Model Hub.

Model Parameters

params, transformer = make_transformer(
        key=jax.random.PRNGKey(42),
        Nf=5,
        Kx=16,
        Kl=4,
        n_max=21,
        h0_size=256,
        num_layers=16,
        num_heads=16,
        key_size=64,
        model_size=64,
        embed_size=32,
        atom_types=119,
        wyck_types=28,
        dropout_rate=0.5,
        widening_factor=4,
        sigmamin=1e-3
)

Training dataset

MP-20 (Jain et al., 2013): contains 45k general inorganic materials, including most experimentally known materials with no more than 20 atoms in unit cell. More details can be found in the CDVAE repository.