Millions of new materials discovered with deep learning

AI Tool Gnome has 2.2 million new crystals, including 380,000 stable materials that could supply future technologies

Modern technologies, from computer systems and batteries to solar panels rely on inorganic crystals. To enable new technologies, crystals must be stable, otherwise they can break down, and for every new, stable crystal may be months of tedious experiments.

Today, in Article published in NatureWe share the discovery of 2.2 million new crystals – equivalent knowledge of almost 800 years. We introduce charts for exploration of materials (GNOME), our new tool for deep learning, which radically increases the speed and performance of discovery, predicting the stability of new materials.

Thanks to Gnome, we have multiplied the number of technologically profitable materials known to humanity. Of the 2.2 million forecasts, 380,000 are the most stable, which makes them promising candidates for experimental synthesis. These candidates include materials that can potentially develop future transformation technologies, from superconductors, supplying supercomptors and new generation batteries to increase the efficiency of electric vehicles.

Gnome shows the potential for the use of artificial intelligence to discover and develop new materials on a large scale. External researchers in laboratories around the world have independently created 736 of these new structures experimentally in simultaneous work. In cooperation with Google Deepmind, a team of scientists from Lawrence Berkeley National Laboratory also published Second article in Nature This shows how our AI forecasts can be used for autonomous material synthesis.

We did Available Gnome forecasts to the research community. We will bring 380,000 materials, which we expect that we will be stable in the design design, which currently processes relationships and adds them to His online database. We hope that these resources will lead to further research on inorganic crystals and unlock the promise of machine learning tools as guides for experiments

In the past, scientists were looking for new crystalline structures, adapting known crystals or experimenting with new combinations of the roots-dogs trial process and an error whose providing even limited results may take months. Over the past decade, the calculation approach led by Material design And other groups helped discover 28,000 new materials. But to date, new approaches directed by AI have achieved a fundamental limit in their ability to accurately predict materials that can be experimentally profitable. The discovery of GNOME 2.2 million materials would be equivalent to knowledge of about 800 years and shows an unprecedented scale and level of accuracy of forecasts.

For example, 52,000 new layered compounds similar to graphene, which can revolutionize electronics along with the development of superconductors. Earlier, around 1000 such materials were identified. We also found 528 potential lithium wires, 25 times more than Previous studywhich can be used to improve battery performance.

We drop the expected constructions for 380,000 materials that have the highest chance of successful execution in the laboratory and use in profitable applications. In order for the material to be considered stable, it cannot break down into similar compositions with lower energy. For example, coal in the graphene structure is stable compared to carbon in diamonds. Mathematically, these materials lie on the convex hull. This project has discovered 2.2 million new crystals, which are stable according to current scientific standards and lie below the convex hull of previous discoveries. Of these, 380,000 are considered the most stable and lies on the “final” convex hull – a new standard that we have established for the stability of materials.

GNOME: Use of charts for exploration of materials

Gnome is the most modern model of the neural network (GNN). The input data for GNN is in the form of a chart that can be compared to connections between atoms, which means that GNN is particularly suitable for discovering new crystalline materials.

Gnome was originally trained with data on crystalline structures and their stability, openly available through Material design. We used gnome to generate new candidate crystals, as well as to predict their stability. To assess the predictive power of our model during progressive training cycles, we have repeatedly checked its performance using established computational techniques known as functional theory of density (DFT), used in physics, chemistry and material materials to understand the structure of atoms, which is important for the assessment of crystal stability.

We used the training process called “Active Learning”, which radically increased Gnome performance. Gnome would generate forecasts for the structures of new, stable crystals, which were then tested using DFT. The resulting high -quality training data was then re -introduced to our model training.

Our research has increased the material stability index of about 50% to 80% – based on Matbench discoveryThe external reference point set by the previous latest models. We also managed to increase the efficiency of our model, improving the discovery rate from less than 10% to over 80% – such efficiency increases can have a significant impact on how many calculations are required to discover.

AI “Recipes” regarding new materials

The GNOME project aims to reduce the costs of discovering new materials. External researchers have independently created 736 new GNOME materials in the laboratory, which shows that the forecasts of our model of stable crystals accurately reflect reality. We have issued our database of newly discovered crystals of the research community. By giving scientists a full catalog of promising “recipes” of new candidates, we hope that it will help them test it and potentially create the best.

The rapidly developing new technologies based on these crystals will depend on the possibility of their production. In an article conducted by our colleagues from Berkeley Lab, scientists have shown that a robotic laboratory can quickly produce new materials with automated synthesis techniques. Using materials from the Materials design and insights on the stability of Gnome, the autonomous laboratory has created new recipes for crystalline structures and effectively synthesized over 41 new materials, opening new possibilities of synthesis of materials powered by AI.

New materials for new technologies

To build a more balanced future, we need new materials. Gnome has discovered 380,000 stable crystals that have the potential to develop more green technologies – from better electric car batteries, to superconductors for more efficient calculation.

Our research – and research of collaborators from Berkeley Lab, Google Research and teams around the world – show the potential for the use of artificial intelligence to discover materials, experiments and synthesis. We hope that Gnome and other AI tools can help revolutionize the discovery of materials today and shape the future of this field.