Question about chemical formula

[DUT-CSJ]

Great work!
There is a question: Can DiffCSP predict the complete unit cell based solely on the simplest chemical formula, such as predicting the structure of C6 given C?
To my understanding, DiffCSP cannot achieve this.
Do you have any new insights regarding this situation?
Looking forward to your reply!

[DUT-CSJ]

I have read your recent work, 'Powder Diffraction Crystal Structure Determination Using Generative Models'.
It's a good work for CSP with PXRD.
However, in this paper, the detailed chemical formula is still needed to predict the structure.
In specific fields such as zeolites, their simplest chemical formula is fixed, but they exhibit numerous allotropes.
How can auxiliary information (such as PXRD or single-crystal electron diffraction) be used to predict crystal structures?
This is what we aim to achieve.
However, we lack knowledge about AI4CSP and hope you can share your helpful insights.
Looking forward to your reply! Thank you very much!!!

[jiaor17]

Hi,
Thanks for the great question! Yes, DiffCSP cannot directly predict all possible structures given the simplest formula solely. However, here are some potential solutions for this problem setting:

• Enumerate possible multiplications of the formula unit and run structure prediction under each, assigning equal or prior-informed probabilities.
• Train a separate predictor to estimate the most likely multiplication factor, possibly conditioned on additional data (e.g., PXRD), and then sample from the resulting probability distribution.
• Use known structural templates corresponding to the given formula, and then apply DiffCSP or DiffCSP++ to refine the structure.

Hope this helps!

[DUT-CSJ]

Hi, Thanks for the great question! Yes, DiffCSP cannot directly predict all possible structures given the simplest formula solely. However, here are some potential solutions for this problem setting:

• Enumerate possible multiplications of the formula unit and run structure prediction under each, assigning equal or prior-informed probabilities.
• Train a separate predictor to estimate the most likely multiplication factor, possibly conditioned on additional data (e.g., PXRD), and then sample from the resulting probability distribution.
• Use known structural templates corresponding to the given formula, and then apply DiffCSP or DiffCSP++ to refine the structure.

Hope this helps!

Thank you for your insightful reply!
I will attempt to extend existing methods to single-crystal electron diffraction. If any issues arise, I hope to discuss them further with you.