AI denoising recovers superconductor spectra in 12s vs 5h, with error bars on the result
bravo_abad · x · 2026-10-21
Younsik Kim and Changyoung Kim train a probabilistic denoising autoencoder that recovers the main features of a superconductor's spectrum from just 12 seconds of measurements (vs a 5-hour reference) while outputting per-point uncertainty estimates.
Key points:
- Uncertainties are propagated into the calculations extracting physical properties;
- Correlations between neighboring points are accounted for — treating them as independent would shrink error bars too much;
- The denoised spectrum suggests a small energy gap absent in the reference: 3.3 ± 1.9 meV at one standard deviation, showing why quantified uncertainty matters.
The paper's core argument: AI-denoised scientific results must carry error bars, or you risk drawing spurious physics conclusions.
More from Research
- Cohere Labs Launches BIRDS Sprint "ML Code Mode" Starting October 31 — Cohere_Labs · 2026-10-21
- CMU Researcher on Long-Horizon Agents: Information Flow, User Modeling and Social World Models — niloofar_mire · 2026-10-21
- Paper accepted at COLM oral: richer graph representations can generalise worse — PMinervini · 2026-10-21
- Toby Ord: benchmark scaling results can't easily settle the r>1 debate — tobyordoxford · 2026-10-21
- EvoMusArt 2027 opens submissions, due Nov 1, 2026 in Mainz, Germany — evomusart_conference · 2026-10-21
- LLM2R recasts 3D part assembly as sequence modeling with a decoder-only LLM — NationalUniversityofSingapore · 2026-10-21