How DeepMind AI Created Millions of New Materials with GNoME

Materials science is a field that studies the properties and applications of different substances, such as metals, ceramics, polymers, and composites. Materials science is crucial for developing new technologies and solving various challenges, such as energy, health, and environment. However, discovering new materials is a difficult and time-consuming process, as it involves experimenting with various combinations of elements and conditions.

To speed up the process of materials discovery, researchers from Google DeepMind, a leading AI company, have developed a new tool called GNoME, which stands for Graph Neural Network for Materials Exploration. GNoME is an AI system that can generate recipes for millions of potentially stable and synthesizable new inorganic compounds, based on existing materials data.

How does GNoME work?

In the case of GNoME, the nodes are atoms and the edges are bonds, and the graphs represent the structures of materials. GNoME uses a large database of known materials, such as the Materials Project, which contains information about 48,000 inorganic compounds. GNoME then tweaks the combinations of atoms and bonds in these graphs to create new graphs that represent new materials.

GNoME also uses a set of rules and constraints to ensure that the new graphs are valid and realistic. For example, GNoME checks the valence of the atoms, the coordination of the bonds, and the symmetry of the structures. GNoME also predicts the stability and synthesizability of the new materials, based on their energy and similarity to existing materials.

What are the results of GNoME?

Using the Materials Project as the input, GNoME generated 2.2 million new graphs that represent new inorganic compounds. Out of these, 1.8 million were predicted to be stable, which means they have a low energy and are unlikely to decompose. Moreover, 381,000 of the stable compounds were predicted to be synthesizable, which means they can be made in a laboratory or a factory.

The new materials generated by GNoME include many interesting and useful substances, such as battery and solar cell components, that were missed by prior algorithms. For example, GNoME predicted a new lithium-ion battery cathode material, Li2MnO3, that has a higher capacity and voltage than the commonly used LiCoO2. GNoME also predicted a new perovskite solar cell material, Cs2AgBiBr6, that has a lower toxicity and higher stability than the widely studied CsPbI3.

Why is GNoME important?

GNoME is a breakthrough in materials science, as it demonstrates the power and potential of AI to accelerate the discovery of new materials. GNoME is able to create millions of new materials in a matter of hours, compared to the years or decades that it would take for human researchers to do the same. GNoME is also able to explore a vast and diverse space of materials, and find novel and promising candidates that could lead to new technologies and applications.

GNoME is also an example of open science, as DeepMind has made the code and the data of GNoME publicly available for anyone to use and test. This means that other researchers and developers can build on GNoME’s work, and verify and validate its predictions. GNoME could also inspire new collaborations and innovations across different disciplines and domains, such as chemistry, physics, engineering, and biology.

Conclusion

GNoME is a new tool by Google DeepMind that uses a graph neural network to generate recipes for millions of new inorganic compounds, based on existing materials data. GNoME predicts the stability and synthesizability of the new materials, and finds many interesting and useful substances, such as battery and solar cell components. GNoME is a breakthrough in materials science, as it shows how AI can speed up the process of materials discovery, and create new opportunities and challenges for the field.