Google DeepMind GNoME: 2.2M Crystal Structures, 380K Stable, 736 Independently Synthesized

Title

Google DeepMind GNoME: 2.2M Crystal Structures, 380K Stable, 736 Independently Synthesized

Content

Google DeepMind's Graph Networks for Materials Exploration (GNoME) is a deep-learning system that has discovered 2.2 million new inorganic crystal structures, equivalent to nearly 800 years' worth of knowledge in the field of computational materials science. The project was published in Nature in November 2023. GNoME uses two pipelines to discover low-energy stable materials. The structural pipeline creates candidates with structures similar to known crystals, while the compositional pipeline follows a more randomized approach based on chemical formulas. The outputs of both pipelines are evaluated using Density Functional Theory (DFT) calculations, and the results are added to the GNoME database, informing the next round of active learning. GNoME is a graph neural network (GNN) model whose inputs take the form of graphs representing connections between atoms, which makes GNNs particularly suited to discovering new crystalline materials. DeepMind reported a dramatic improvement in discovery rate. "Our research boosted the discovery rate of materials stability prediction from around 50%, to 80% - based on MatBench Discovery, an external benchmark set by previous state-of-the-art models. We also managed to scale up the efficiency of our model by improving the discovery rate from under 10% to over 80%." The combination of graph neural networks and diffusion models enabled GNoME to screen billions of candidate crystals in days on GPU clusters. External researchers have independently synthesized 736 of GNoME's predictions in laboratories across the globe, demonstrating that the model's stability predictions accurately reflect physical reality. Examples include a first-of-its-kind Alkaline-Earth Diamond-Like optical material (Li4MgGe2S7) and a potential superconductor (Mo5GeB2). In a Berkeley Lab-led follow-up, a robotic lab used Materials Project and GNoME data to synthesize more than 41 new materials autonomously. DeepMind has released the full database of newly discovered crystals to the research community as AI 'recipes' for new materials. Looking ahead, DeepMind plans to expand GNoME into organic and hybrid organic-inorganic crystals, integrate it with autonomous synthesis platforms like the Berkeley A-Lab for closed-loop discovery, and continue releasing candidate databases for downstream researchers to test.