Inquiry and Appreciation Regarding the 2D Material Generation Model

[zc994646659]

Dear developer,

I hope this message finds you well.

First of all, I would like to sincerely thank you for developing such an exciting and innovative generative model for 2D materials. I’ve been exploring it recently, and I find its potential incredibly inspiring for our field.

I have a few questions and would greatly appreciate your guidance:
1. Element-Specific Generation: Is it possible to control or constrain the elements used in the generated structures?
2. Training Data Design: I understand that the model’s output depends heavily on the training data. How should I design my training set if I want to bias the generation toward specific chemical compositions or structure types?
3. Use of C2DB: Would it be appropriate to include the full C2DB dataset as training input?

Thank you again for your excellent work and for your time in addressing these questions. I’m very excited to experiment more with the model and potentially contribute to its development.

Warm regards,
Chao Zhou

[chu-]

@zc994646659 sorry for the delayed reply, I missed your message.

1.Yes, it is possible to control and constrain the elements used in generated structures in the SCDM codebase, though the constraints are implemented through predefined element groups rather than arbitrary user-defined constraints.
The SCDM system implements element-specific generation through the replace_values_per_row function, which applies different element constraints based on the unique values present in each structure's atom type array.
（1）Structures with only one element type use the xenes group
（2）Binary structures with two element types use M (metals) and X (chalcogens) groups
（3）Ternary structures use A_elements, B_elements, and O_elements for complex oxide formation
2. The SCDM codebase contains several key components for creating and processing training sets that bias generation towards specific chemical compositions and structure types:The primary mechanism for biasing generation towards specific chemical compositions is the atom_mask parameter. The fundanmental solution to this problem is reverse design capability of the model. actually, it's under development.
3. Choosing the right features is essential for training effective models. Feature selection includes identifying the most relevant features in the data set, reducing dimensions, avoiding overfitting, and improving the interpretability of the model. For example, when predicting the yield strength of a material, parameters related to temperature, strain rate, and grain boundary slip mechanism should be considered. It is recommended to combine data from other databases, such as MP, Jarvis, etc.
All in all, using a complete C2DB data set as training input has potential, but it also requires careful handling of issues such as data quality, feature selection, and model verification.