New inorganic crystals can be used in better batteries, computer chips and other technologies. Normally, the search is difficult because the number of possible compositions and structures is enormous.

GNoME scaled graph-network methods to predict material stability and guide computational searches. Its original publication reported 381,000 novel stable materials; an August 2024 dataset expansion released more than 520,000 materials.

It's not a complete solution. It identifies candidates, not ready-made products. Each material has to be synthesized, characterized and evaluated. AI has dramatically enlarged and prioritized the territory scientists can explore.

Why it matters

A large, organized search space gives laboratories promising starting points for technologies central to energy, computing and manufacturing.

Original material

Scaling deep learning for materials discovery

Google DeepMind / Nature · 2023 · dataset expanded 2024

Read the primary source