Inquiry About Pretrained Model for Classification Task in ALIGNN

[goodluckhu]

Dear Professor DeCost,

I hope this message finds you well. My name is Jinbin, and I am currently exploring your work on "Atomistic Line Graph Neural Network for improved materials property predictions". I find your research both innovative and impactful, especially in the realm of materials property predictions.

I have a specific question regarding the classification task ROC AUC performance on the JARVIS-DFT dataset for ALIGNN models. I was unable to locate a pretrained model for Stable/unstable (ehull) classification in the repository, specifically in the file alignn/pretrained.py. The provided scripts seem to focus solely on regression models. Considering the significance of classification performance, I am keen to explore this aspect and would like to try using the classification model.

Could you kindly let me know if a pretrained model for Stable/unstable (ehull) classification is available for sharing? Additionally, I am interested in understanding whether this classification model could be applied to predict the stability of zeolite systems.

I appreciate your time and look forward to your response. @bdecost
Best regards,
Jinbin Hu

[bdecost]

Hi - I'm not sure if there is a model checkpoint for stability classification relative to the convex hull, but there are a few classification checkpoints on figshare here: https://figshare.com/articles/dataset/ALIGNN_models_on_JARVIS-DFT_dataset/17005681. I don't think they are supported by the pretrained model script, but you could grab the checkpoints and instantiate the models directly

• jv_magmom_oszicar_alignn_class
• jv_mbj_bandgap_alignn_class
• jv_n-powerfact_alignn_class
• jv_p-Seebeck_alignn_class
• jv_slme_alignn_class
• jv_spillage_alignn_class
• jv_optb88vdw_bandgap_alignn_class

@knc6 do you have a stability classification model checkpoint somewhere?

We do have a pretrained regression model for that target, but in practice it might be better to compare formation energies directly, especially if you are extrapolating to crystal structures not represented in the training data (like zeolites)