Old Webb Photos Give Up a Planet Hiding in the Glare

A team led by Rodrigo Ferrer-Chavez has built software that removes starlight from James Webb Space Telescope images far more cleanly than standard methods, and it has already pulled a planet out of data where earlier processing saw nothing ^[1][2].

The researchers report that their method, called FLIP, makes faint planets two to 15 times easier to detect, with the gain varying by observation. They describe it in a Nature paper published on November 4, 2026 ^[1].

One point first. The planet is not new to astronomy. It is 51 Eridani b, a known giant planet, and the paper reports its first detection in Webb data at a wavelength of 4.3 microns, where standard techniques had produced a non-detection ^[1][2]. The news is that a better reading of observations already collected can show things the first reading did not.

The problem is glare

To photograph a planet directly, you have to separate its faint light from the light of the star it orbits. The Nature paper says young giant planets can be about 100,000 times fainter than their host stars in near-infrared light ^[1]. Removing the star's image accurately is the central obstacle to direct imaging ^[1].

Webb carries a coronagraph in its NIRCam camera, a mask that blocks most of a star's light. Some always leaks through, and it forms a speckled pattern that can bury a planet.

The usual cure is to build a reference picture of the star from other observations and subtract it. That works until something shifts. A small change in pointing or temperature makes the reference differ from the actual image, and residual glare remains ^[1].

Modelling the telescope instead

FLIP, short for Physics-based Learning for Imaging Planets, takes another route. It starts from a physical model of the instrument's optics and detector and calibrates it against a reference observation. It then adjusts the model until it matches the starlight pattern in the science image ^[1].

The tuning is done with automatic differentiation and gradient descent, techniques from machine learning that nudge millions of model parameters step by step toward a better fit ^[1].

There is an obvious danger in a method that fits itself to an image: it might fit the planet too and subtract it along with the star. The team guarded against this by injecting artificial planets into the data and tracking their signal during optimization ^[1].

The paper compares FLIP's results with published analyses that used standard pipelines. To the authors' knowledge, it says, this is the first demonstration of substantially improved sensitivity from this kind of model-based subtraction on real coronagraphic observations ^[1].

The approach did not come easily. Ferrer-Chavez said his first attempt, training machine-learning models to predict stellar glare directly, did not work. Jason Wang, a co-author, said that early exploration helped the team understand the problem ^[2].

What the archive might hold

The practical promise is that no new observing time is needed. Ferrer-Chavez said the tool could let researchers look for planets that were missed in Webb's archive, or that they tried to image but could not detect at the time ^[2].

Telescope time is rationed and expensive. Reprocessing images that already exist costs computing power instead.

Katie Bouman, a Caltech co-author, framed the result as a shift in what software can do for hardware. "What's exciting is that we're showing how computation can help overcome limitations of the telescope itself," she said ^[2].

The title of the paper makes the same claim in technical form: "Imaging exoplanets below the telescope stability limit" ^[1]. The idea is that a telescope's physical steadiness caps how well standard subtraction can work, and a model can do better than the cap.

The larger prize

The stakes extend past Webb. NASA's planned Habitable Worlds Observatory has a goal that includes directly imaging Earth-like planets around Sun-like stars ^[1][2]. The Nature paper notes that such work faces an exceptionally demanding telescope-stability problem ^[1].

The paper argues that model-based processing could ease that requirement ^[1]. That is a prospect, not something FLIP has done. The paper demonstrates the method on selected Webb data and does not claim to have searched the full archive or to show that every dataset will yield the same gain ^[1].

The team says it plans to improve FLIP and apply it to other instruments, including the Nancy Grace Roman Space Telescope and ground-based telescopes ^[2].

What is not yet known

The evidence so far is one research group's paper plus its institution's announcement. Caltech's account is not independent validation ^[2]. No other team has yet been reported to have reproduced the 51 Eridani b detection or the sensitivity gains.

The spread from two- to 15-fold is also wide. How the gain varies with the target star, the filter and the distance between planet and star will decide how many old observations are worth reprocessing.

The next test is whether FLIP finds a planet nobody knew about. The team has not said how many more Webb observations it will process or when. A detection of an unknown planet in old data would show the method is a tool for finding planets, and not only for re-measuring ones astronomers already had on their lists.