Skip to content

Latest commit

 

History

History
 
 

pre-trained

README.md

Pre-trained CGCNN models

This directory includes several pre-trained CGCNN models that one can use to predict the material properties of new crystals. We encourage users to report their own CGCNN models for other properties which might benefit the community.

Pre-trained models

Regression

File Property Units Data Ref. Model Ref.
formation-energy-per-atom Formation Energy eV/atom Jain et al. Xie et al.
final-energy-per-atom Absolute Energy eV/atom Jain et al. Xie et al.
band-gap Band Gap eV Jain et al. Xie et al.
efermi Fermi Energy eV/atom Jain et al. Xie et al.
bulk-moduli Bulk Moduli log(GPa) Jain et al. Xie et al.
shear-moduli Shear Moduli log(GPa) Jain et al. Xie et al.
poisson-ratio Poisson Ratio — Jain et al. Xie et al.

Classification

File Positive Negative Data Ref. Model Ref.
semi-metal-classification Metal Semiconductor Jain et al. Xie et al.

Before using pre-trained models

  • CGCNN models (and machine learning models in general) can only generalize to crystals from the same ditribution as training data. It is adviced to check Data Ref. before using pre-trained models. For instance, Materials Project uses ICSD structures as input, which includes experimentally synthesized crystal structures. Significant errors can be expected if a CGCNN model trained on Materials Project is used to predict the properties of imaginary, thermadynamically unstable crystals.
  • CGCNN models have prediction errors. It is adviced to check Model Ref. to understand their accuracy before using pre-trained models.

How to cite

If you used any pre-trained models, please cite both Data Ref. and Model Ref. because data and model are equally important for a successful machine learning model!

How to share your pre-trained models

Please send an email to [email protected] if you want to share your own pre-trained models. Since we don't have time to check the validity of your model, we only accept peer reviewed works.

To submit, be sure to include:

  1. A pth.tar file storing the CGCNN model.
  2. The type of the model and the target property.
  3. The links to the data reference and model reference.