A 0.62-AUC model made a 65,578-candidate materials search tractable, doubling hit rate
bravo_abad · x · 2026-09-16
bravoabad walks through Ran et al.'s work on discovering altermagnetic metal-organic frameworks, arguing that ML models in scientific discovery don't need to be accurate — they just need to decide what's worth calculating next.
- The target material class is rare: the initial DFT dataset had only 6 positives among 350 candidates
- A class-weighted XGBoost on molecular fingerprints served as a coarse filter; its AUC of 0.62 looks modest on a conventional benchmark, but it narrowed 65,578 candidates to 145 for expensive DFT calculations
- Six new altermagnets were found, raising the hit rate from 1.7% to 4.1%
- Takeaway: in rare-event discovery, judge models by how much they enrich the experiment pipeline, not benchmark scores
More from Research
- AI researchers predicted 2054 for a Millennium Problem; reality arrived far sooner — ben_j_todd · 2026-09-16
- Paper2Agent turns research papers into interactive AI agents (Nature) — EricTopol · 2026-09-16
- CPAL 2027 heads to Tokyo: small conference on parsimony and learning — YiMaTweets · 2026-09-16
- Discrete Beckmann Transport Models: One-Step Discrete Generators Without a Teacher — rishabh16_ · 2026-09-16
- AI models are becoming appliances: great at one thing, bad at everything else — dbreunig · 2026-09-16
- Google's Fuse paper benchmarks 12 LLMs on inferring hidden social motives — dair_ai · 2026-09-16