Graph construction for large-unit cell crystals (e.g., MOFs)

[meiirbek-islamov]

Hello,

I'm currently utilizing ALIGNN for property prediction of MOFs, but I'm encountering significant delays in the process due to the time-consuming steps of reading CIF files, processing them, and constructing graphs. Is there any approach to expedite this procedure? I'm aware that ALIGNN has been trained with approximately 137,000 MOFs for CO2 capture performance. I'm curious to know how you managed such a substantial quantity of MOFs. Additionally, is it possible to save the constructed graphs for future use, avoiding the need to recreate them during each training session?

Thank you!

[knc6]

Hi,

We had to use high CPU memory something like 100 GB. So, make sure
#SBATCH --mem=100G or something like that.

We had to use a smaller batch_size on the order of 32 or lower. It took about 30 hours for training on a single V100 GPU.

You can save the graphs if you save save_dataloader as true. This is good for debugging, but not recommended.

You can checkout the job.out file from any of these figshare files to learn about additional details e.g., https://figshare.com/ndownloader/files/31459198.