Bayesian optimization reconstructs 3D atomic structures from one microscopy image in ~75s
bravo_abad · x · 2026-11-08
Xiankang Tang and coauthors show how Bayesian optimization can replace days-to-weeks of iterative simulation and expert tuning when reconstructing 3D atomic arrangements from a single electron microscopy image.
- Working on BaTiO₃, they pair an electron microscopy simulator with a Gaussian process that learns which imaging parameters and atomic configurations best reproduce the experimental image
- Physical knowledge enters the search as geometric constraints: candidates must satisfy them before reaching the expensive simulator
- Imaging parameters converge in roughly 75 seconds, reaching 98.1% correlation between experiment and simulation
More from Research
- giffmana: paper claims zero real savings—no scaling trend or FLOPs-matched results — giffmana · 2026-11-08
- Embodiment vs physical instantiation: researchers debate whether LLMs already have a toehold — AnnaCiaunica · 2026-11-08
- JEPA debate: Berton says LeJEPA plateaus sooner than JEM, demands scaling laws — gabriberton · 2026-11-08
- Retuned LeJEPA Still Loses to JEM; Scaling Comparison Draws LeCun Into SSL Debate — gabriberton · 2026-11-08
- Study tests continued pretraining on 1B-8B open models to refresh factual knowledge — burkov · 2026-11-08
- Tencent unveils WeLM-Image: 31.7B MoE image model tops open-source benchmarks, 5s per image on one H20 — Crazy-Repeat-2006 · 2026-11-08