Could CrystaLLM deal with the disordered crystal structures?

[yuedawang]

Dear developers,

Could CrystaLLM deal with (e.g. training and sampling) the disordered crystal structures with the atoms partially occupied in the cif files?

[lantunes]

Hi,

Thank you for your question and your interest in this project. The current version of CrystaLLM was not trained on disordered crystal structures. Therefore, the model will not be able to deal with any composition that implies a disordered structure, and it will not be able to deal with non-stoichiometric compositions. It will also not generate a CIF file with partially occupied sites.

However, it should be possible, in theory, to train (or fine-tune) the model on CIF files containing partially occupied sites. This would be a very interesting project.