rishikeshmagar / crystal-twins Goto Github PK

Hello,

In your paper, I found you use CGCNN as the backbone, and cannot beat the ALIGNN. Just a little curious, why don't you use a newer network as the backbone? Since CGCNN itself does not work so well. Have you done this kind of experiment?

Hi,

In your paper, you used $CT_{SimSiam}$ loss, but I could not find the implementation of it in your code, only $CT_{Barlow}$ loss is implemented. I could not reproduce the result for $CT_{SimSiam}$.

Hello, and thanks for making this work accessible! I am curious if you can give a bit more detail on the matminer and hMOF datasets you used for your pretraining.

For the data from Matminer, which datasets did you use? If I download everything from Matminer's datasets (https://hackingmaterials.lbl.gov/matminer/dataset_summary.html) and filter for those containing material structure information, I get 614,427 rows, which doesn't match the 152,368 reported in the paper. Does removing duplicates reduce this down to 152,368, or did you only work with a subset of those datasets?

For the data from hMOF, downloading the full hMOF zip file from here (https://mof.tech.northwestern.edu/databases) yields 137,953 CIFs, which doesn't match the 275,907 reported in the paper. Did you include other databases from the MOFXDB?