A Caltech- and Northwestern-led team has shown that a physics-based machine-learning method can strip starlight from James Webb Space Telescope images well enough to recover a known gas giant that standard analysis missed in the same data, Caltech said in a 4 November release.
The technique, Physics-based Learning for Imaging Planets (FLIP), builds a computational “digital twin” of JWST’s NIRCam, solves for millions of possible physical imperfections, models the star’s glare and subtracts it. The peer-reviewed paper by Ferrer-Chavez, Wang, Bouman and colleagues appears in Nature.
In brief
- FLIP is a digital twin of JWST NIRCam that subtracts starlight after the fact.
- Caltech says sensitivity is two to 15 times better than standard techniques on known planets.
- 51 Eridani b is a known planet newly detected in this JWST archive set — not a new discovery.
What FLIP actually did on 51 Eri b
Direct imaging is hard because a planet’s light is swamped by its star. FLIP attacks that after the exposure, by fitting a physics-based model of the telescope rather than relying only on classical point-spread-function subtraction. Caltech said the method improves detections of known planets with sensitivity two to 15 times superior to standard techniques.
In archived JWST observations, FLIP recovered 51 Eridani b, a previously known planet that the original standard analysis of that dataset did not detect. Lead author Ricardo Ferrer-Chavez said the team now has a tool to discover and study planets that could not be “seen” before, “even if that planet was already in an image in the JWST archive.” Co-author Jason Wang said: “I feel we just unlocked a new frontier,” arguing the approach can push past the telescope’s stability limit and may apply to other instruments.
Katie Bouman, a Caltech professor and co-author, said the excitement is showing “how computation can help overcome limitations of the telescope itself.” The authors note computational cost still needs to fall, and planned use on NASA’s Roman Space Telescope and ground observatories is future work, not a completed campaign. Funding credits listed by Caltech include STScI, the Sloan Foundation, an Amazon AI4Science Discovery Award and the NSF — the same broader AI-for-science funding pattern TSN covered when Anthropic expanded AI for science credits.
Why archive re-analysis is a Tech for Good story
Re-processing public JWST data with a better forward model is a cheap way to multiply the science return of an already-flown observatory. It sits with other open scientific model releases TSN has flagged, such as Deep Principle’s Hydrogen-1.0, where the point is enabling more labs to run harder problems without waiting for a new mirror to launch. It also fits the weekly Tech for Good beat TSN tracks in round-ups of applied science wins.
Known planet, new detection — a careful claim
Caltech’s release is careful on priority: 51 Eri b was already known from earlier work; FLIP’s contribution is detecting it in this JWST dataset where standard reduction did not. That distinction matters for headlines and for readers who might otherwise think JWST just found a new world. The science win is methodological sensitivity, not a discovery press conference.
Computation is becoming part of the observatory
The deeper shift is institutional. When a digital twin can beat the stability floor of a flagship space telescope, proposal time and archive mining start to matter as much as raw collecting area. Hardware coronagraphs remain essential; FLIP changes the queue of planets worth revisiting in data that already exists on disk. UK and European JWST users on the archive side can, in principle, try similar re-reductions once code and costs allow, even though this paper’s lead institutions are US-based.
There is also a funding signal: Amazon’s AI4Science award sitting beside NSF and STScI support shows how cloud and foundation money now underwrite classical astronomy software stacks. That is the same blurry line TSN keeps seeing between “AI research credit” programmes and instrument science.
Watch next whether FLIP’s code and twin are released in a form others can run, whether independent teams reproduce the 51 Eri b recovery, and whether Roman planning documents start citing post-processing twins as baseline. The confirmed claim today is methodological: better exoplanet imaging from computation on JWST data, including a known planet pulled out of an archive that standard tools left blank.
Sources
- Erasing the stars to find exoplanets (Caltech, 4 November 2026)
- Ferrer-Chavez et al., Imaging exoplanets below the telescope stability limit (Nature, 4 November 2026)
Related stories
- Anthropic Expands AI for Science Credits to 20 Research Partners
- Deep Principle Opens Hydrogen-1.0, a 753B-Parameter Science Model
- Tech for Good This Week: Cancer AI, Hospital Drones and New MXenes
Follow @tsnmediahub on X for breaking AI news as it happens.
Follow on X
