CSP mode: NaCl related issueH

[k13524]

Hello,
We trained the crystal‑structure–prediction model on the MP‑20 dataset and tried to reproduce the NaCl example from the README. When we input Na:1, Cl:1, the generated trajectories and final structures differed markedly from each other. From other issue threads, it sounds like the model expects the exact number of atoms of each species. Could this be causing the discrepancy? Also, is there a recommended guideline or best‑practice note for running the CSP model? Thanks!

[danielzuegner]

Hi @k13524,

Diffusion sampling is a stochastic process, so we should not expect the model to produce the exact same sample and/or trajectory when sampling multiple times. If you are looking for a model which consistently produces the same output structure, a diffusion model is not the right approach.

Note that we do not have any CSP tasks or models in the paper, so we provide the CSP mode only for convenience for the community. Hence, there are no best practices or metrics from the paper to compare with.

Does this address your issue, or is there anything beyond the fact that the model produces different samples?

[k13524]

Thanks so much for the quick reply, Daniel!

Just to be sure I’m understanding: I know diffusion is stochastic, but for a given atomic composition I’d expect the generated cells to be in roughly the same ballpark (e.g., not jumping from ~10 × 10 × 10 Å to ~12 × 100 × 50 Å). In our case we only condition on targets like band width, and the outputs still come with specific compositions and structures. I’ve been assuming they stay within some reasonable range so DFT relaxations are still feasible.

Am I looking at this the wrong way? Is there really no implicit or explicit range on the cell metrics?

Appreciate any clarification!

[danielzuegner]

Hi @k13524,

What you're describing sounds like the cell sometimes "exploding" during the sampling process. I've experienced this before when the model was not really well-trained. What data was the model trained on? If it was trained on Alex-MP-20 I would not expect this to be a problem.

Best,

Daniel

[k13524]

Ah-ha, that could be the reason. We used MP-20. Let me try re-training with Alex-MP-20. Thank you again for your help!