Materials Discovery
The techniques transforming molecular biology apply directly to physical materials. Generative models propose compositions and structures; simulation tells you which ones are worth a furnace.
Areas
Electrodes, electrolytes, and interfaces under transport and stability objectives.
Composition and defect search guided by electronic structure.
Activity and selectivity screened before synthesis.
Property targets translated into monomer and sequence design.
Storage, conversion, and thermal management.
Functional compounds with multi-property specifications.
Discover materials before making them.
Why AI
Plausible four-element compositions exceed a trillion. The standard structure database holds about 184,000 entries spanning 9,141 structure types, because human discovery proceeds outward from structures it already knows.
Machine-learned interatomic potentials reproduce quantum-chemistry energies at a small fraction of the cost, turning exhaustive stability screening from impossible into routine.
Instead of filtering an enumerated list built from known prototypes, they sample structures conditioned on a target property, so a proposal need not belong to any existing structural family.
When a robot reads its own diffraction pattern and picks the next recipe, the model's next prediction is shaped by yesterday's failed synthesis rather than by a literature review.
Evidence
The landmark results in this field arrived with published critiques attached. Both belong on the same page.
A graph network trained on 48,000 known stable crystals proposed 2.2 million structures, 381,000 of them predicted stable, deliberately exploring compositions the authors describe as escaping previous human chemical intuition. 736 turned up in other labs' experiments while the work was under way.
Merchant et al., Nature, 2023 ↗Materials chemists inspected a random sample of the stable-structure database and found an existing entry for every one, usually at higher symmetry, meaning the new compound was an ordered version of a known disordered phase. Their verdict was scant evidence of novelty, credibility and utility together.
Cheetham & Seshadri, Chemistry of Materials, 2024 ↗Language models proposed recipes from the synthesis literature, a scheduler chose the next experiment, and the system read its own diffraction. The original report claimed 41 of 58 and the word novel; after materials chemists disputed it, Nature issued a correction in 2026 lowering the count and removing that word.
Szymanski et al., Nature 2023; Author Correction 2026 ↗Orthodoxy held that fast ion conduction needs a narrow range of ionic environments. An AI-assisted, expert-supervised search produced a material with fifteen crystallographically distinct lithium sites that conducts on a par with liquid electrolytes. Other groups have since built on it.
Han et al., Science, 2024 ↗The honest other half, and the reason we build the way we do: predicted-stable is not synthesisable, and synthesisable is not novel or useful. The recurring failure across the flagship results is identical and structural, because pipelines place each element on its own site and so re-report a well-known disordered phase as a new ordered compound. Every published critique lands on the same prescription rather than on rejection, which is that crystallographic and synthesis expertise has to be inside the loop rather than bolted on afterwards.
