Millions of modern materials have been discovered thanks to deep learning

GNOME AI tool finds 2.2 million modern crystals, including 380,000 stable materials that could power future technologies

State-of-the-art technologies, from computer chips and batteries to solar panels, rely on inorganic crystals. To enable modern technologies, crystals must be stable or they may degrade, and behind each modern, stable crystal may lie months of painstaking experimentation.

Today in A article published in Nature, we share the discovery of 2.2 million modern crystals – equivalent to almost 800 years of knowledge. Introducing Graph Networks for Materials Exploration (GNoME), our modern deep learning tool that dramatically increases the speed and efficiency of discovery by predicting the stability of modern materials.

Thanks to GNOME, we have multiplied the number of technologically feasible materials known to mankind. Of the 2.2 million predictions, 380,000 are the most stable, making them promising candidates for experimental synthesis. These candidates include materials that have the potential to develop future transformative technologies, ranging from superconductors, powering supercomputers and next-generation batteries to augment the performance of electric vehicles.

GNOME shows the potential of using artificial intelligence to discover and develop modern materials at scale. Outside researchers in laboratories around the world independently created 736 of these modern structures, experimentally, in parallel work. In collaboration with Google, DeepMind was also published by a team of researchers from Lawrence Berkeley National Laboratory second article in Nature this shows how our AI predictions can be used to autonomously synthesize materials.

We did GNOME forecasts available to the research community. We will donate 380,000 materials that we believe will be stable to the Materials Project, which is currently processing these compounds and adding them to your online database. We hope these resources will advance inorganic crystal research and unlock the potential of machine learning tools to guide experiments

In the past, scientists searched for modern crystal structures by modifying known crystals or experimenting with modern combinations of elements – a costly process of trial and error that could take months to produce even circumscribed results. Over the last decade, computational approaches led by Project Materials and other groups helped discover 28,000 modern materials. However, so far, modern AI-based approaches have reached a fundamental limit in their ability to accurately predict materials that could be experimentally feasible. GNOME’s discovery of 2.2 million materials would represent approximately 800 years of knowledge and demonstrates an unprecedented scale and level of prediction accuracy.

For example, 52,000 modern graphene-like layered compounds that could revolutionize electronics with the development of superconductors. Previously o 1,000 such materials have been identified. We also found 528 potential lithium-ion conductors, 25 times more than a previous studywhich can be used to improve battery performance.

We publish predicted structures for 380,000 materials that have the greatest chance of being successfully made in the laboratory and used in real applications. For a material to be considered stable, it must not decompose into similar lower energy compositions. For example, carbon with a graphene-like structure is stable compared to the carbon in diamonds. Mathematically, these materials lie on a convex hull. This project discovered 2.2 million modern crystals that are stable by current scientific standards and lie below the convex envelope of previous discoveries. Of these, 380,000 are considered the most stable and lie on the “final” convex hull – a modern standard we have set for material stability.

GNOME: Using graph networks for material mining

GNOME is a state-of-the-art graph neural network (GNN) model. The input to GNNs is in the form of a graph that can be compared to the connections between atoms, which makes GNNs particularly suitable for discovering modern crystalline materials.

Originally, GNOME was trained on publicly available data on crystal structures and stability Project Materials. We used GNOME to generate novel candidate crystals and also predict their stability. To assess the predictive power of our model over progressive training cycles, we repeatedly tested its performance using established computational techniques known as density functional theory (DFT), used in physics, chemistry, and materials science to understand atomic structure, which is significant for assessing crystal stability.

We used a training process called “active learning” that dramatically improved GNOME performance. GNoME would generate predictions about the structures of modern, stable crystals, which would then be tested using DFT. The resulting high-quality training data was then fed back into our training model.

Our research increased the discovery rate of material stability predictions from approximately 50% to 80%. The discovery of MatBench, the external benchmark set by previous cutting-edge models. We were also able to augment the performance of our model, improving the discovery rate from less than 10% to over 80% – such a performance augment can have a significant impact on the amount of computation required per discovery.

AI “recipes” for modern materials

The GNOME project aims to reduce the cost of discovering modern materials. Third-party researchers have independently created 736 modern GNOME materials in the lab, demonstrating that our model’s predictions for stable crystals accurately reflect reality. We have made our database of newly discovered crystals available to the research community. We hope that by providing scientists with a full catalog of promising “recipes” for modern candidate materials, it will lend a hand them test and potentially create the best ones.

The rapid development of modern technologies based on these crystals will depend on the possibility of their production. In a paper led by our colleagues at Berkeley Lab, scientists demonstrated that a robotic laboratory can rapidly produce modern materials using automated synthesis techniques. Using materials from the Materials Project and stability knowledge from GNoME, the autonomous lab created modern crystal structure recipes and successfully synthesized over 41 modern materials, opening modern possibilities for AI-powered materials synthesis.

Fresh materials for modern technologies

To build a more sustainable future, we need modern materials. GNoME has discovered 380,000 stable crystals that have the potential to develop greener technologies – from better batteries for electric cars to superconductors for more proficient computing.

Our research – and that of collaborators at Berkeley Lab, Google Research and teams around the world – shows that artificial intelligence has the potential to lend a hand with materials discovery, experimentation and synthesis. We hope that GNOME, along with other AI tools, will lend a hand revolutionize materials discovery today and shape the future of the field.