AI Discovers New Materials
Google's GNoME (2023) proposes 2.2 million candidate crystals (about 380,000 predicted stable) and Microsoft's MatterGen (2025) designs to target properties; lab confirmation and novelty lag.
Open in the interactive tree →About 48,000 stable inorganic crystals were known before; GNoME predicted roughly 380,000 more as stable, of which external labs had made 736 by late 2023. The A-Lab autonomous lab (Berkeley, 2023) reported 41 of 58 target compounds synthesized, but critics (Palgrave and others) argued in 2024 that many products were misidentified or already known. MatterGen (Nature, January 2025) generates crystals for target values; one synthesized example (TaCr2O6) reached a bulk modulus of 169 GPa against a target of 200 GPa.
As of October 2026
Nature's January 2026 correction of the A-Lab paper states that 'novel' meant new to the prediction platform, not necessarily new to science, and marks four reported syntheses as inconclusive. A June 2026 study finds that structures proposed by GNoME and MatterGen depart from the structural patterns of experimentally discovered materials, with MatterGen closer to them. Both remain research tools for proposing candidates; the bottleneck is making and measuring the materials, not predicting them.
Open steps
- Automatic novelty and realism check High AI leverageCheck whether proposed crystals are new to science and structurally like real materials; A-Lab's 'novel' meant new to a platform.
- Predicting how to make it High AI leveragePredict precursors, temperatures and atmospheres for a target crystal, not only whether it is stable.
- Reliable phase ID in autonomous labs Medium AI leverageIdentify mixed or disordered products from diffraction data without false success claims, and log failed attempts in a shared database.
- Beyond stability: functional properties High AI leveragePredict and verify conductivity, magnetism and ionic transport, not only stability, for generated materials.
Where AI could help
Medium AI leverage. Prediction is cheap; the remaining work is validation, where automated labs and better structure identification decide what is real and new.
- Self-driving labs that synthesize and characterize candidates every day
- Automated novelty checks against the literature and databases
- Models trained on failed and ambiguous syntheses to filter proposals before the lab
Shown so far
- In January 2026 Nature corrected the A-Lab paper: the lab synthesized 36 of 57 targets, which were new to the prediction platform and not necessarily new to science. source
- In January 2025 Microsoft published MatterGen in Nature, a generative model for inorganic crystals; one synthesized structure measured within 20% of its target property value. source
- The Matbench Discovery benchmark (August 2023) found universal interatomic potentials to be the best machine-learning method for pre-screening stable crystals, with up to 6 times faster discovery than random picking. source
Prerequisites
- Periodic Table of Elements1869
- Quantum Chemistry & Bonding1927
- Density Functional Theory1965AI materials models are trained on DFT calculations
- Deep Learning2012