Examining GNNs for Crystal Structures: Limitations and Opportunities for Capturing Periodicity

Explore the limitations and opportunities of Graph Neural Networks (GNNs) for capturing periodicity in crystal structures in this one-hour conference talk. Delve into the systematic analysis of GNNs' capabilities in representing crystal structures, comparing them with human-designed descriptors. Discover the challenges GNNs face in capturing periodicity and examine proposed solutions, including a hybrid approach combining descriptors with GNNs. Learn about the improved prediction of materials properties, particularly phonon internal energy and heat capacity, achieved through this method. Gain insights into the mechanisms behind enhanced predictions and understand how this analysis can be extended to other deep representation learning models, descriptors, and systems such as molecules and amorphous materials.

- Intro
- Molecules vs. Crystals
- Representation of Crystal Structures
- Graph Neural Networks
- Global and Local Structural Descriptors
- Periodicity
- Pooling Function
- Summary and Conclusion
- Q&A and Discussion