Research
AI tool GNoME finds 2.2 million new crystals, including 380,000 stable materials that could power future technologies
Modern technologies from computer chips and batteries to solar panels are based on inorganic crystals. To enable new technologies, the crystals must be stable or they can decay, and behind each new, stable crystal can be months of painstaking experimentation.
Today, in one document published in Nature, we share the discovery of 2.2 million new crystals – equivalent to nearly 800 years of knowledge. Introducing Graph Networks for Materials Exploration (GNoME), our new deep learning tool that dramatically increases the speed and efficiency of discovery by predicting the stability of new materials.
With GNoME, we have multiplied the number of technologically viable materials known to mankind. Of its 2.2 million predictions, 380,000 are the most stable, making them promising for experimental synthesis. Among these candidates are materials that have the potential to develop future transformative technologies ranging from superconductors, power supercomputers and next-generation batteries to enhance the performance of electric vehicles.
GNoME demonstrates the potential of using artificial intelligence to discover and develop new materials at scale. Outside researchers in laboratories around the world have independently created 736 of these new structures experimentally in parallel work. In collaboration with Google DeepMind, a team of researchers at Lawrence Berkeley National Laboratory has also published a second paper inside Nature This shows how our AI predictions can be leveraged for autonomous material synthesis.
We have made GNoME forecasts are available in the research community. We will contribute 380,000 materials that we predict will be stable to the Materials Project, which is now processing the compounds and adding them to its online database. We hope these resources will advance inorganic crystal research and unlock the promise of machine learning tools as guides for experimentation
Accelerate material discovery with AI
About 20,000 of the experimentally identified crystals in the ICSD database are computationally stable. Computational approximations drawn from the Materials Project, the Open Quantum Materials Database and the WBM database increased this number to 48,000 stable crystals. GNoME expands the number of stable materials known to mankind to 421,000.
In the past, scientists searched for new crystal structures by modifying known crystals or experimenting with new combinations of elements—a costly trial-and-error process that could take months to yield even limited results. In the last decade, computational approaches have led from the Materials project and other groups have helped discover 28,000 new materials. But until now, new AI-driven approaches have hit a fundamental limit in their ability to accurately predict materials that might be experimentally viable. The discovery of 2.2 million materials by GNoME would be equivalent to about 800 years of knowledge and demonstrates an unprecedented scale and level of accuracy in predictions.
For example, 52,000 new graphene-like multilayer compounds that have the potential to revolutionize electronics with the development of superconductors. Previously, approx 1,000 such materials had been identified. We also found 528 potential Li-ion conductors, 25 times more than a previous studywhich could be used to improve the performance of rechargeable batteries.
We release the predicted structures for 380,000 materials that have the highest chance of being successfully fabricated in the lab and used in sustainable applications. For a material to be considered stable, it must not decay into similar compositions with lower energy. For example, carbon in a graphene-like structure is stable compared to carbon in diamonds. Mathematically, these materials are in the convex hull. This project discovered 2.2 million new crystals that are stable by current scientific standards and lie below the convex hull of previous discoveries. Of these, 380,000 are considered the most stable and are in the ‘ultimate’ curved hull – the new standard we have set for material stability.
GNoME: Leveraging Graph Networks for Materials Exploration
GNoME uses two pipelines to discover low-energy (stable) materials. The structure pipeline generates candidates with structures similar to known crystals, while the synthesis pipeline takes a more randomized approach based on chemical formulas. The results of both pipelines are evaluated using established Density Functional Theory calculations and these results are added to the GNoME database, informing the next round of active learning.
GNoME is a state-of-the-art graphical neural network (GNN) model. The input data for GNNs is in the form of a graph that can be likened to connections between atoms, which makes GNNs particularly suitable for discovering new crystalline materials.
GNoME was initially trained with data on crystal structures and their stability, openly available through the Materials project. We used GNoME to generate new candidate crystals, as well as to predict their stability. To evaluate the predictive power of our model during progressive training cycles, we repeatedly tested its performance using well-established computational techniques known as Density Functional Theory (DFT), which are used in physics, chemistry, and materials science to understand the structures of the atoms, which is important for evaluating the stability of crystals.
We used a training process called “active learning” that dramatically increased the performance of GNoME. GNoME would generate predictions for the structures of new, stable crystals, which were then tested using DFT. The resulting high-quality training data was then fed into our training model.
Our research increased the discovery rate of material stability prediction from about 50%, to 80% – based on an external benchmark set by previous state-of-the-art models. We were also able to scale the efficiency of our model by improving the discovery rate from below 10% to over 80% – such performance increases could have a significant impact on how much computation is required per discovery.
Artificial intelligence “recipes” for new materials
The GNoME project aims to reduce the cost of discovering new materials. External researchers have independently created 736 of GNoME’s new materials in the lab, demonstrating that our model’s predictions for stable crystals accurately reflect reality. We have released our database of newly discovered crystals to the research community. By giving scientists the full list of promising “recipes” for new candidate materials, we hope it will help them test and potentially make the best ones.
After completing the last discovery efforts, we searched the scientific literature and found that 736 of our computational discoveries were made independently by outside groups around the world. Above are six examples ranging from a first-of-its-kind alkaline earth diamond-like optical material (Li4MgGe2S7) to a potential superconductor (Mo5GeB2).
The rapid development of new technologies based on these crystals will depend on the ability to manufacture them. In work led by our collaborators at Berkeley Lab, researchers have shown that a robotic laboratory could rapidly produce new materials with automated synthesis techniques. Using materials from the Materials Project and stability insights from GNoME, the stand-alone lab created new crystal structure recipes and successfully synthesized more than 41 new materials, opening up new possibilities for AI-based materials synthesis.
A-Lab, a facility at Berkeley Lab where artificial intelligence guides robots to make new materials. Photo: Marilyn Sargent/Berkeley Lab
New materials for new technologies
To build a more sustainable future, we need new materials. GNoME discovered 380,000 stable crystals that have the potential to develop greener technologies – from better batteries for electric cars to superconductors for more efficient computers.
Our research – and that of collaborators at Berkeley Lab, Google Research and groups around the world – shows the potential of using artificial intelligence to guide materials discovery, experimentation and synthesis. We hope that GNoME along with other AI tools can help revolutionize materials discovery today and shape the future of the field.