Finding new inorganic crystals can take a long time. Google Deepmind’s GNoME tool aims to speed up that search by proposing structures and estimating their stability. The team says its discoveries could amount to more than would otherwise have been found over 800 years.
Millions of candidates from a materials model
Google Deepmind reported that GNoME identified more than 2.2 million new crystals. Among them are about 380,000 compounds considered particularly stable, which could be relevant to future technologies such as superconductors, better batteries and new types of semiconductors.
These results are potential starting points for research, rather than ready-made technologies. A predicted structure still needs to be evaluated and, where possible, produced in a laboratory before its practical value can be established.
The 380,000 most stable candidates have been made freely available to researchers around the world. Sharing the data gives scientists a broad set of materials to investigate as they work toward new technologies.
How active learning improved predictions
GNoME began with publicly available information from the Materials Project, including data on known materials’ crystal structures and stability. The system then proposed new materials, which were assessed using quantum physical density functional theory.
That process is described as active learning: results from the stability checks were fed back into the model’s training. In this way, each round of testing could inform later predictions, rather than leaving the model with only its initial training data.
According to Deepmind, this approach raised the system’s discovery rate for predicting material stability from around 50 percent to over 80 percent. The team also reported an efficiency improvement, taking the discovery rate from under 10% to over 80%. It said that could affect how much computing is needed for each discovery.
Predictions meet laboratory work
Some of GNoME’s proposed crystals have already moved from computation to experiment. Other laboratories have synthesized 736 of the predicted new crystals, providing experimental confirmation for those candidates.
Automated laboratories may help researchers work through more possibilities. At Berkeley Lab, the experimental robotics laboratory called A-Lab synthesized 41 new materials in an automated process within 17 days. The system selected ingredients, carried out synthesis and analyzed the results.
This example shows how automated production can connect a search for materials with attempts to make them. But a large number of predicted structures could still exceed what automated laboratories can produce, leaving researchers with a selection problem.
Choosing what is worth making
The gap between prediction and production is central to the next stage of materials research. A model can generate candidates faster than laboratories can test them, so estimating which structures deserve experimental attention could help direct limited synthesis work.
GNoME’s predictions, the experimentally synthesized crystals and the A-Lab process each address a different part of that challenge: suggesting structures, checking whether they can be made and automating parts of the work. Together, they point to a research pipeline in which AI helps prioritize possibilities while laboratory results remain essential.
The potential applications are still prospects. Stable candidates could matter for batteries, superconductors or semiconductors, but researchers first need to determine which materials can be produced and how useful they are. Open access to the most stable candidate data gives research groups a starting point for that work.