Originally published on The Daily Flare.
AI materials discovery has expanded the search for inorganic crystals, with Google DeepMind's GNoME system predicting 2.2 million candidate structures and identifying hundreds of thousands as potentially stable.
How machine learning widened the search
Materials used in electronics, batteries, solar cells and other technologies have traditionally been found through a mix of laboratory experiments and computational screening. Searching through possible combinations of elements and crystal structures is costly and slow. Researchers also need to establish whether atoms can form a stable structure and whether the material has useful properties.
In a 2023 study published in Nature, Google DeepMind researchers described GNoME, short for Graph Networks for Materials Exploration. The team trained graph neural networks on information about known and predicted crystals, drawing on the Materials Project and related databases. GNoME then assessed new candidates, with results from successive rounds feeding back into the system's training process.
This repeated process is known as active learning. Researchers use calculations to evaluate selected candidates, add the resulting data and train the model again. The approach helped the team explore combinations that earlier searches had not covered as broadly, including structures containing five or six different elements.
The study reported 2.2 million candidate structures below the existing convex hull, a computational measure used to compare a material's energy with competing phases. A companion account in Science described 381,000 of the predicted compounds as having calculated formation energies consistent with stability.
GNoME's predictions also broadened the range of compositions considered. The researchers highlighted layered compounds, a class relevant to battery electrodes, and predicted 52,000 layered materials compared with about 1,000 identified in the Materials Project. The set included 528 predicted lithium-ion conductors, materials of interest for battery technologies.
Prediction is only the first step
A predicted crystal is not automatically a practical material. Scientists still need to synthesize it, determine what was actually produced and measure its properties. Lab conditions, competing phases and synthesis difficulties can all affect whether a theoretical candidate can be made in a useful form.
The 2.2 million figure describes predicted structures, while the smaller set identified as potentially stable reflects a computational filter. Neither number is a count of new products already made or technologies ready for use. Nature's study reported that 736 of the stable structures had already been independently realized in experiments.
A separate effort described in Nature reporting brought automation into the experimental stage. The A-Lab, an autonomous system combining artificial intelligence and robotics, generated recipes, mixed and heated ingredients, then analyzed products without human intervention during the process. Its work was presented alongside the GNoME predictions as a way to connect computational candidate lists with laboratory synthesis.
Automation could help researchers test more candidates in a consistent way, but it does not remove the need for scientific judgment. Results still require interpretation, and a material's potential application depends on measurements beyond the initial act of synthesis.
Assessing progress beyond the prediction count
The larger scientific opportunity is a more systematic route to creating new materials. Instead of relying only on familiar chemical substitutions or trial and error, researchers can use computational screening to examine a wider range of possible structures.
Potential uses named in the research and reporting include battery materials, superconductors, catalysts, electronics and solar cells. Those are areas where improved materials could matter, but a predicted candidate is only a starting point.
A 2025 Nature feature noted criticism of efforts by major technology companies to accelerate materials research, while also describing progress behind the claims. Machine learning can increase the number and diversity of proposed materials, but the practical value of those proposals depends on validation.
For now, GNoME's contribution is a larger and more varied map of possible inorganic crystals, supported by computational screening and some independent experimental realizations. The next measure of its impact is how researchers use that map: which candidates survive laboratory tests, what properties they demonstrate and whether those properties meet real technical needs.
Top comments (0)