LLM-extracted chemistry rules find new crystals from just 22 examples
bravo_abad · x · 2026-11-14
Guo et al. demonstrate a crystal-discovery pipeline with radically low labeling cost: an LLM extracts rules from 902 perovskite papers, and chemists combine those rules with their own expertise to define eight numerical inputs—describing properties like how well a crystal's organic and inorganic components fit.
An XGBoost classifier learns from these features, with active learning choosing the next experiments to minimize labeling cost. The final model was trained on just 22 labelled compositions; in a separate test of 14 new candidates, 11 produced the intended crystals.
The lesson: before collecting a bigger training set, ask which parts domain knowledge can replace—chemical priors cut labeling needs by orders of magnitude.
More from Research
- New B-cell directed evolution paper rediscovers a 2004 Roger Tsien idea — NikoMcCarty · 2026-11-14
- Researchers claim 1,000× faster reasoning training with 80s-era energy models — alec_helbling · 2026-11-14
- Ai2 Asks Whether Prompting or Training Can Make LLMs Give More Useful Advice — allen_ai · 2026-11-14
- Ai2's XYBench Finds Five Top LLMs Fail to Correct Users' Misguided Questions — allen_ai · 2026-11-14
- G-STAR paper: 32–64x token compression gives 20x faster autoregressive image generation — _akhaliq · 2026-11-14
- OpenCLIP 4.2 ships with CLAP2 variable-length audio and NaFlexViT support — wightmanr · 2026-11-14