A HydroGEN consortium team developed a generalizable defect graph neural network modeling approach for predicting defect formation enthalpies using the ideal (defect-free) host crystal structure and properties as input.
HydroGEN Advanced Water Splitting Materials Consortium
August 11, 2023
minute read time
A HydroGEN consortium team developed a generalizable defect graph neural network modeling approach for predicting defect formation enthalpies using the ideal (defect-free) host crystal structure and properties as input that can negate the need for many orders of magnitude more expensive DFT supercell relaxations. Critically, the methodology also:
- Requires no manual feature engineering and is therefore not limited to specific crystal/symmetry classes or chemistries
- Provides models that can be systematically improved with more training data
- Readily integrates other state-of-the-art convolution functions as they are developed/published.
Publication
Witman, M.D., Goyal, A., Ogitsu, T. et al. Defect graph neural networks for materials discovery in high-temperature clean-energy applications. Nat Comput Sci (2023). https://doi.org/10.1038/s43588-023-00495-2
Background
High-accuracy prediction of vacancy defect formation enthalpies elucidate the primary and critical figure of merit needed to assess a material’s utility across a large variety of applications. Previous efforts to predict defect properties and vacancy formation enthalpies span various methods and material classes within which the models are applicable. Deep learning techniques, such as graph or convolutional neural networks, can circumvent such limitations.
Authors and Affiliations
- Sandia National Laboratories - Matthew D. Witman & Anthony H. McDaniel
- National Renewable Energy Laboratory - Anuj Goyal & Stephan Lany
- Indian Institute of Technology Hyderabad - Anuj Goyal
- Lawrence Livermore National Laboratory - Tadashi Ogitsu