New starlight-removal method recovers known planet 51 Eridani b in JWST data

FLIP removed starlight from JWST coronagraph images 2 to 15 times more sensitively than standard processing, but the planet it recovered, 51 Eridani b, has been known since 2015.

Peer reviewed · Method paper · Tested on known planets

A new image-processing method has pulled the known giant planet 51 Eridani b out of James Webb Space Telescope (JWST) data at 4.3 microns, a wavelength where standard processing found nothing. Rodrigo Ferrer-Chavez, Jason Wang and colleagues report in Nature on 4 November 2026 that their software, FLIP, detects the planet at a signal-to-noise ratio of 6.3. The catch: “hidden planet” headlines overstate it. Nobody found a new world. The advance is the method, which still falls short of the best a telescope can do.

Evidence at a glance

Is this a big deal?

It is a solid step for a hard technical problem, not a discovery. The paper shows that modelling the telescope’s physics beats the standard recipe on real data, and an outside specialist calls it significant. Whether the gains carry over to the much fainter planets astronomers actually want is untested.

What scientists found

51 Eridani b is not new. A team using the ground-based Gemini Planet Imager announced it in 2015: a planet of about 2 Jupiter masses (inferred from models, not measured directly), roughly 13 astronomical units (AU) from its star in projected separation, around a star about 20 million years old. Thirteen AU is a little farther out than Saturn orbits the Sun, in the gap before Uranus. The system is about 96 light-years away.

What is new is the colour. A 2025 JWST study by William Balmer and colleagues detected the planet at 4.1 microns but not at longer wavelengths. In the new paper’s analysis, standard processing also produced no detection in the F430M filter, centred near 4.3 microns, where carbon dioxide absorbs. FLIP detected the planet there at a signal-to-noise ratio of 6.3, which the authors describe as the first image of 51 Eridani b at 4.3 microns. Sensitivity at the planet’s location improved by about a factor of four over the earlier analysis.

Across the datasets, FLIP’s sensitivity to faint companions was 2 to 15 times better than the standard technique, called KLIP. For one star, TYC 5899, the gain was about 15 times at 0.5 arcseconds from the star and about 5 times at 1 arcsecond. For 51 Eridani it was about 5 times and 2.5 times at the same separations. Detection signal-to-noise on the known planets rose about 1.6 times for HIP 65426 b and about 2.1 times for 51 Eridani b in the F410M filter.

How they found it

Ferrer-Chavez told Northwestern’s newsroom that planets are about a million times fainter than their stars, so the glare must be removed before a planet shows. Webb’s coronagraph, a mask that blocks the starlight, does most of the work. What leaks past the mask is a glare pattern that shifts as the telescope changes by tiny amounts.

The standard fix, used for about 20 years according to Wang, is to photograph a reference star and subtract its pattern. Think of subtracting yesterday’s photo of a stadium floodlight from today’s, after the lamp has moved. The leftover glare can swamp a planet. The analogy breaks in one place: the reference star is real data, not a photograph of the same light.

FLIP, short for Physics-based Learning for Imaging Planets, instead builds a computer model of the star, the telescope’s optics and the detector. Northwestern’s release calls it a “digital twin”; the paper calls it a physical model, reimplemented in the PyTorch software library so every step can be differentiated. The model has four groups of free parameters: the star’s flux and position behind the mask, optical path errors before and after the mask, and each pixel’s response. In stage one it is calibrated on a reference star. In stage two a subset of parameters is tuned on the science image, with about a million free parameters optimised in roughly 35 minutes on one NVIDIA A30 graphics processor. To avoid fitting the planet itself, the team injects fake planets and stops when their signal-to-noise peaks.

The datasets were HIP 65426 (F444W filter), 51 Eridani (F410M and F430M) and TYC 5899, a red dwarf with no known companions used for injection tests. Wang stressed that the method uses the computational tools behind AI models but no actual AI, because the physics is already known.

Why it matters to you

For readers outside astronomy, this is a wonder-lens story with no near-term effect on daily life. It matters for what comes next. NASA’s proposed Habitable Worlds Observatory aims to photograph Earth-like planets around Sun-like stars, a task that demands extraordinary telescope stability. The authors suggest FLIP-style processing could relax part of that requirement, and the same idea could help the Nancy Grace Roman Space Telescope. That is a proposal, not a demonstration.

What the evidence does not show

No new planet appears in the paper. The release headline’s “hidden planet” refers to a planet that older processing had not recovered at this wavelength.

FLIP does not yet reach the photon-noise limit, the floor set by the light itself, on any observation, according to the authors. For HIP 65426 the central 0.6 arcseconds were not sampled because of bright planet lobes. On TYC 5899, FLIP lost more of the signal than KLIP at larger separations in some cases. The authors also note the optical models are known to be imperfect, learning rates were chosen by hand, bad pixels were masked by eye, and only one telescope roll per dataset was processed. Extensions to ground-based telescopes and to JWST’s MIRI camera are proposed, not shown.

A signal-to-noise ratio of 6.3 is a detection, but a modest one. The paper’s text does not say the new 4.3-micron measurement changes the planet’s known properties.

How this fits what we already knew

JWST had already imaged 51 Eridani b at 4.1 microns. The new paper notes earlier evidence for a metallicity (heavy-element content) of about five times the star’s, and says this detection lets future work test that. For how Webb’s instruments work, see our explainer on how the James Webb Telescope works.

Most known planets were found indirectly, from their effect on a star’s light or motion, and only a small share have been imaged directly. For the bigger census, see our guide to how many planets there are in the Milky Way.

What an outside specialist says

Dimitri Mawet, a senior research scientist at NASA’s Jet Propulsion Laboratory who was not involved, told Scientific American “This is a significant result.” He added that physics-based post-processing like this could plausibly ease part of the Habitable Worlds Observatory’s stability error budget. His hedge, “could plausibly,” matters. SciQuest has not interviewed an outside researcher.

How other headlines framed it

• Northwestern University, via EurekAlert: led with a “hidden planet” emerging. That reads like a discovery, which the paper does not report.
• Phys.org: stressed a giant planet emerging from older JWST data. Every voice quoted belongs to the study team.
• Scientific American: went with a software fix for a wobbly telescope, and brought in an outside expert.
• The Brighter Side of News: pitched it as erasing glare. Its subtitle calls the planet known.

What happens next

The paper points to testing the method with the Roman Space Telescope and other instruments, better calibration of the telescope models, and using the 4.3-micron detection to test the planet’s metallicity. Roman’s launch is covered in our Roman report. The test that would change this story is a FLIP image of a planet nobody had seen before.

In one sentence

A physics-based starlight-removal method recovered the known planet 51 Eridani b at 4.3 microns and beat standard processing by 2 to 15 times in tests, but it found no new planet and has yet to reach its theoretical limit.

Read next

• JWST maps WASP-17b’s atmosphere: another exoplanet result from Webb.
• How the James Webb Telescope works: the instrument behind this story.
• How many planets are in the Milky Way? The bigger census.

Sources and disclosure

1. Ferrer-Chavez R., Wang J.J., Bouman K.L., Freeman W.T., Pueyo L., Feng B.Y., Perrin M.D. “Imaging exoplanets below the telescope stability limit.” Nature 659, 58-62 (2026). doi:10.1038/s41586-026-11213-z. Dates from Crossref.
2. Northwestern University, “Hidden planet emerges from the glare,” 4 November 2026: EurekAlert.
3. Macintosh B. et al. “Discovery and spectroscopy of the young Jovian planet 51 Eri b with the Gemini Planet Imager.” Science 350, 64 (2015): arXiv:1508.03084.
4. Balmer W.O. et al., The Astronomical Journal (2025), DOI 10.3847/1538-3881/adb1c6: arXiv:2503.13608.
5. Krywko J., Scientific American, 4 November 2026 (Mawet quotes): article. Also read for framing: Phys.org, The Brighter Side of News. Distance to 51 Eridani (about 96 light-years) is from secondary catalogue summaries, not the paper.

Disclosure: This article was drafted with AI assistance and checked against the sources above, including the paper’s open-access full text. The authors declare no competing interests. Calculations by SciQuest: 13 AU is about 1.4 times Saturn’s average distance of 9.6 AU. The graphic is a SciQuest original built from values reported in the paper. Spotted an error? See our corrections policy.