Specify the number of atoms for generations

[yuhao1982]

For generation, can we specify the number of atoms we want. e.g. 200 ? It seems that you sample the number of atoms according to the distribution of alex-mp data which is the same as the training data. So you want train and test phase are in the same distribution ?

[danielzuegner]

Hi @yuhao1982,

Yes, by default MatterGen samples the number of atoms from the ALEX_MP_20 distribution (1–20 atoms), matching the training data. You can customize this in a few ways:

Option 1: Create your own num_atoms distribution

Edit mattergen/common/data/num_atoms_distribution.py and add a new entry to NUM_ATOMS_DISTRIBUTIONS. For example, to always generate structures with exactly 10 atoms:

NUM_ATOMS_DISTRIBUTIONS = {
    "ALEX_MP_20": { ... },  # existing
    "FIXED_10": {
        10: 1.0,
    },
}

Or a custom range (probabilities must sum to 1.0):

"UNIFORM_15_20": {
    15: 0.1667,
    16: 0.1667,
    17: 0.1667,
    18: 0.1667,
    19: 0.1667,
    20: 0.1665,
},

Then pass the name via --sampling_config_overrides:

mattergen-generate --pretrained_name=... --output_path=... \
  --sampling_config_overrides='["+condition_loader_partial.num_atoms_distribution=FIXED_10"]'

Option 2: Use --target_compositions for exact compositions

If you know the exact composition you want, you can bypass the distribution entirely:

mattergen-generate --pretrained_name=... --output_path=... \
  --sampling_config_name=csp \
  --target_compositions='[{"Si": 8, "O": 16}]'

This generates a 24-atom structure with a fixed Si₈O₁₆ composition (requires a CSP-trained model).

Important caveat: The model was trained on structures with up to 20 atoms, so generating structures with significantly more atoms (e.g., 200) would be out of distribution and likely produce low-quality results.

Hope this helps, feel free to follow up if you have further questions.

[danielzuegner]

Closing due to inactivity.