Untrained Network as Prior: Recovering 3D Chemistry from Just 16 Electron-Microscope Views
bravo_abad · x · 2026-09-19
In electron tomography, reconstructing a 3D chemical map normally requires many viewing angles, but samples can't tilt far enough and more measurements mean more radiation damage — leaving an inverse problem with big data gaps. Del Pozo Bueno et al. skip pretraining entirely: an untrained 3D convolutional network is optimized directly against the measured projections, with the architecture itself acting as a prior favoring structured 3D solutions over noise. A clever twist jointly reconstructs all chemical elements to exploit cross-element constraints — a 'deep image prior' approach applied to electron-microscopy chemistry.
More from Research
- A 7B model can score pretraining data quality — and measurably improve downstream benchmarks — cephaloform · 2026-09-19
- 10 Machine Learning Books That Build Real Understanding, From Goodfellow to Chip Huyen — thisguyknowsai · 2026-09-19
- 25 Fields Medallists incl. Terence Tao push back on AI: solving problems is 'only a tool and proxy' — beglen · 2026-09-19
- Google DeepMind turns AI on genome switches that run our cells every second — borowcy · 2026-09-19
- LDDM: 3D generative model designs novel drug hits validated by X-ray across 5 targets — mmbronstein · 2026-09-19
- CARL: RL agent learns to create and steer self-organizing Lenia patterns — danbri · 2026-09-19