AI-accelerated materials discovery

Core thesis

DeepMind's GNoME used deep learning to predict 2.2M new crystal structures, with 380,000 identified as stable. Of these, hundreds have been independently synthesised — validating AI predictions. For batteries: 528 novel lithium conductors found (25x previous studies). For superconductors: 52,000 graphene-like materials identified. Berkeley Lab's Materials Project integrates AI foundation models with LLM chatbots for interactive exploration. The bottleneck has shifted from discovery to synthesis — we can predict materials faster than we can make them.

State of the art (2026)

The frontier has moved from prediction to autonomous synthesis. DeepMind GNoME and Microsoft MatterGen settled the discovery question - MatterGen now ships on Azure Quantum Elements, with a model-designed compound synthesised at Shenzhen close to its target bulk modulus. The live race is the self-driving lab that closes the predict-make-measure loop. Berkeley A-Lab made 41 of 58 novel compounds in 17 days; the DOE Genesis Mission has put critical materials on a national agentic-AI footing. The defining 2026 event is Periodic Labs, founded by GNoME lead Ekin Dogus Cubuk on a $300M a16z-led seed, building robotic powder-synthesis labs aimed at higher-temperature superconductors. Value accrues to whoever owns the proprietary experimental data, not the model.