AI can invent millions of materials — but how many actually work?

Silicon Valley’s push to turbocharge materials discovery with AI has drawn both fanfare and skepticism. DeepMind says its GNoME model predicted 2.2 million stable crystalline materials and helped guide hundreds of lab syntheses, while Microsoft’s MatterGen aims to directly generate crystals with desired properties and Meta touted AI-flagged MOFs for carbon capture. Critics counter that many AI-proposed compounds are impractical, duplicate known materials, or assume perfectly ordered crystal structures unlikely at real temperatures—undermining claimed breakthroughs. A robotic “A-Lab” that used AI-devised recipes also faced challenges over mischaracterization and novelty, prompting a broader debate about standards and validation. Proponents say AI is a powerful signpost, not an end product, and that tight coupling with experimental chemists—and models that account for disorder—will determine whether the hype translates into usable materials.

Read more