Inquiry About Model-Supported Material Properties and Validation

[lima-zt]

Hello! I would like to know how to view the specific physical property parameter settings (such as band gap, elastic modulus, etc.) supported by each generative model, and whether the system provides validation functions for these parameters (rather than just evaluating general properties like structural stability or novelty). Are there any relevant documents or operational guidelines available for reference? Thank you!

[danielzuegner]

Hi @lima-zt,

For space group, any one of the 230 are supported. For all other properties, you can input arbitrary values, but the quality of the model's prediction will depend on whether there are training samples with similar values. For instance, inputting energy above hull of -1000.0 is unlikely to generate a meaningful sample, but is generally possible. I suggest you inspect the training data for the respective ranges of the properties you're interested in.

Our code does not include code to evaluate these properties, you'll have to use your own DFT- or ML-based workflows to do so.

Hope this helps!

Daniel

[lima-zt]

Hi @danielzuegner ,
Thank you for your response. I have reviewed the config.yaml file of the mattergen_base model, which indicates that the attributes include dft_bulk_modulus, dft_band_gap, dft_mag_density, ml_bulk_modulus, hhi_score, space_group, and energy_above_hull. However, after defining the mattergen_base model and specifying the output path, when I executed the command mattergen-generate RESULTS_PATH --model_path=MODEL_PATH --batch_size=2 --properties_to_condition_on="{'space_group': 225, 'dft_band_gap': 4.90}" --diffusion_guidance_factor=2.0, it reported an attribute matching error during conditional generation. Does this mean the mattergen_base model does not support defining these attributes?

Furthermore, for other official models (e.g., dft_mag_density_hhi_score), are their adjustable properties strictly limited to the parameters explicitly declared in the model name (i.e., does this model only support the two conditional attributes, dft_mag_density and hhi_score)?

Thank you for your time and consideration.

LiMa

[danielzuegner]

Hi @lima-zt,

As stated in the README, the base model is not trained on any properties:

mattergen_base: unconditional base model trained on Alex-MP-20

To sample with a property constraint you will have to use the model with the respective matching name.

Furthermore, for other official models (e.g., dft_mag_density_hhi_score), are their adjustable properties strictly limited to the parameters explicitly declared in the model name (i.e., does this model only support the two conditional attributes, dft_mag_density and hhi_score)?

Yes, the models only support the properties indicated by their name.

Best,

Daniel