Graph Neural Networks for Materials Discovery
How the GNoME project used GNNs to discover new crystal structures
Welcome to Infinite Curiosity, a newsletter that explores the intersection of Artificial Intelligence and Startups. AI enthusiasts across 200 countries have been reading what I write. Subscribe to this newsletter for free to receive it in your inbox every week:
The GNoME project by Google DeepMind used Graph Neural Networks (GNNs) to discover new crystal structures. GNNs represent a powerful approach to materials science, especially in the discovery process.
Here are 8 key points to note:
1. GNNs are a type of neural network designed to process data represented in graph form. In materials science, the graph structure is particularly relevant because it can represent the atomic structure of materials. Nodes are atoms and edges are chemical bonds.
2. In the case of crystal discovery, GNNs analyze the graph representations of known crystal structures. By learning patterns and relationships within these structures, GNNs can predict the properties and stability of potential new crystals.
3. GNoME utilizes a dataset of known crystal structures and their stability from the Materials Project. The data is in the form of graphs. In these graphs, nodes and edges carry information about the atomic arrangement and chemical bonding.
4. The GNN in GNoME is trained to predict the stability of new crystal structures. Stability is a key property because it determines whether a material can be realistically synthesized and utilized in practical applications.
5. GNoME's GNN operates through two pipelines - structural and compositional. The structural pipeline generates candidates with structures similar to known crystals. The compositional pipeline takes a more randomized approach based on chemical formulas. This dual approach allows for a comprehensive exploration of potential new materials.
6. The predictions made by the GNN are evaluated using Density Functional Theory (DFT) calculations. It's a computational physics method used to investigate the electronic structure of atoms and molecules. This helps in validating and refining the predictions of the GNN.
7. GNoME employs an active learning process. What exactly is it? It's where the GNN's predictions are continuously improved by feeding back the results from the DFT calculations. This process enhances the accuracy and reliability of the predictions over time.
8. By leveraging GNNs, GNoME has been able to predict a vast number of new materials with a high degree of accuracy. This represents a significant scale-up in material discovery as compared to the older methods of experimentation that tend to be slower.
If you're a founder or an investor who has been thinking about this, I'd love to hear from you. I’m at prateek@moxxie.vc and our fund Moxxie Ventures leads seed rounds.
If you are getting value from this newsletter, consider subscribing for free and sharing it with 1 friend who’s curious about AI: