A digital twin of James Webb makes planets up to 15 times easier to see

Webb’s cameras had already caught the planet. It sat inside images of a young nearby star, buried under a haze of leftover starlight, and the standard way of cleaning such images could not pull it out. A new technique that rebuilds the James Webb Space Telescope in software, then subtracts the glare that virtual copy predicts, has now made it visible.

The gain is large. Applied to existing Webb observations, the method made faint planets between two and fifteen times easier to detect than the reference-star technique astronomers have relied on for two decades, according to a paper in Nature led by Rodrigo Ferrer-Chavez of Northwestern University. Webb’s archive may already hold worlds nobody has seen, and future telescopes built to photograph an Earth twin might not need to be quite so extraordinarily stable.

The recovered planet is 51 Eridani b, a young gas giant about 96 light-years away that ground-based telescopes first photographed roughly a decade ago. What is new is the wavelength. The method, called FLIP, short for Physics-based Learning for Imaging Planets, captured the planet at 4.3 microns, where carbon dioxide in its atmosphere absorbs strongly and where an earlier analysis of the very same Webb data had come up empty.

Recommended ReadingJWST found ancient galaxies hiding four times their measured mass in small stars

Why a planet can hide in plain sight

Even the brightest young giant planets are roughly 100,000 times fainter than their stars in near-infrared light, the paper notes. To see one, Webb’s NIRCam camera uses a coronagraph, a mask that blocks the star’s disc. Some light still leaks around it and spreads into a speckled haze, and that haze is often brighter than any planet sitting inside it.

The usual fix is subtraction. Astronomers photograph a second star, treat that image as a template of the glare and subtract it from the target. “For the past 20 years, astronomers have tackled this problem by observing a different star and using its image as a reference,” said Jason Wang, an assistant professor of physics and astronomy at Northwestern and a co-author of the study.

The weakness is that the telescope never sits perfectly still between the two shots. “Small thermal changes could cause a mirror to expand or contract by a picometer,” Wang said, adding that the motors in Webb’s cryocoolers can make the mirrors vibrate. Each change alters the glare pattern, so the template never quite matches, and the residue can bury a faint planet.

How they did it

Instead of borrowing glare from another star, FLIP calculates it. The team started from STPSF, the standard optical model of Webb maintained by the Space Telescope Science Institute, and rebuilt it in PyTorch, a programming library designed for training neural networks. That made the whole model differentiable: the software can work out how each parameter affects the final image and adjust all of them at once.

The model has four groups of free parameters. They cover the brightness of the star and its exact position behind the mask, the distortions of the light before the mask, the distortions after it, and the response of each detector pixel. In total the optimisation tunes about a million parameters, and one run takes around 35 minutes on a single NVIDIA A30 graphics card.

The model is first calibrated on a reference star, so it starts close to the right answer, and then tuned on the science image itself to absorb whatever the telescope changed in between. “We used the same computational tools that researchers use to train AI models, but we did not actually use AI,” Wang said. “Because we already know the physics, we don’t need AI to understand what’s happening.”

The obvious risk is that a flexible model learns to erase the planet along with the glare. To guard against that, the team planted fake planets in each image and tracked how clearly they stood out as the optimisation ran, stopping it once their signal peaked or levelled off. The same planted planets also measure how much real signal the method removes.

Fifteen times better, but not everywhere

The team tested FLIP on three published Webb datasets. Two target stars with known planets, HIP 65426 and 51 Eridani. The third, TYC 5899, has no known companions and was observed four months after its reference stars, a long gap in which the telescope’s optics drifted noticeably.

On HIP 65426 b, the first exoplanet Webb imaged directly, the signal came out about 1.6 times stronger than with the standard method, known as KLIP. On 51 Eridani b at a slightly shorter wavelength, it was 2.1 times stronger. At 4.3 microns the planet appeared with a signal-to-noise ratio of 6.3, clearing the five-sigma bar for a detection where the original analysis saw nothing.

The largest improvement came on TYC 5899, the drifting dataset. There, FLIP could detect planets about 15 times fainter than KLIP at half an arcsecond from the star, and about five times fainter at one arcsecond. An arcsecond is roughly the width of a coin seen from about five kilometres away.

That case matters most for discovery. Many Webb surveys use other targets observed weeks or months apart as their reference stars, and the paper counts more than 100 observations taken this way, including programmes GO 5835 and SURVEY 6005. All of them could be reprocessed.

The wavelength matters too. Carbon dioxide absorbs strongly at 4.3 microns, which dims a planet but reveals how rich in heavy elements its atmosphere is, and Earth’s own atmosphere blocks that light, so no ground-based telescope can measure it. Earlier work suggested 51 Eridani b holds about five times the heavy-element share of its star; the new image offers a way to test that.

What the method doesn’t settle

FLIP did not discover a new planet. All three planets it detected were already known, and the fake planets planted in the TYC 5899 data show what the method could find, not what it has found. The first genuine discovery, if one comes, will need to survive independent checks.

The method is also short of its own ceiling. A perfect physical model would reach the photon-noise limit, the floor set by the random arrival of light itself. FLIP gets much closer to it than KLIP but reaches it in none of the observations. The authors add that their model cannot yet report, with error bars, how much the telescope actually shifted, and that a bright nearby star in the TYC 5899 field survived the subtraction in distorted form.

The tests also cover one instrument, Webb’s NIRCam, with two kinds of coronagraph and a handful of filters. Ground-based telescopes, with far larger data volumes and a turbulent atmosphere to look through, are a harder problem that FLIP has not yet tackled.

Why a future telescope cares

NASA’s proposed Habitable Worlds Observatory is meant to photograph Earth-like planets around Sun-like stars, planets roughly ten billion times fainter than their star. With today’s approach, the paper says, such a telescope must hold its optics steady at the picometre level, 100 to 1,000 times more stable than Webb, and that requirement drives much of the mission’s difficulty.

If software can correct for drift instead of engineering it away, some of that burden lifts. “We know of no planets around other stars that are like those in our own solar system,” Ferrer-Chavez said. “It’s not that they don’t exist. It’s that we cannot see them.”

Common questions about Webb and imaging exoplanets

How does the James Webb telescope photograph exoplanets?

Webb blocks the light of a star with a coronagraph, a mask inside its NIRCam and MIRI instruments, and then removes the leftover glare in software. What remains can show a planet as a faint point of light. It works best for young, hot giant planets far from their stars, which still glow in infrared light.

What is 51 Eridani b?

It is a young gas giant orbiting the star 51 Eridani, about 96 light-years from Earth in the constellation Eridanus. The system is about 20 million years old, so the planet still glows with heat left over from its formation. It was discovered with the Gemini Planet Imager, a ground-based instrument, and announced in 2015.

Did FLIP use artificial intelligence?

Not in the sense of a trained neural network. It borrows the optimisation tools used to train AI models, including the PyTorch library and the Adam optimiser, but the model itself is built from the known physics of the telescope.

Could this help find an Earth-like planet?

Not with Webb directly, but it could make future telescopes easier to build. The Habitable Worlds Observatory, planned for launch in the 2040s, must be extremely stable to image Earth twins, and software that corrects for instabilities could relax that requirement, the authors argue.

What comes next

The study was published in Nature on 4 November 2026. The team will use the new 4.3-micron detection to measure the planet’s brightness at that wavelength and redo the analysis of its atmosphere, and wants to extend FLIP to MIRI, Webb’s mid-infrared instrument, and to NASA’s Nancy Grace Roman Space Telescope. With the Gaia mission expected to reveal thousands of giant planets on wide orbits, there will be no shortage of targets for a sharper eye.

Reference: Ferrer-Chavez et al., “Imaging exoplanets below the telescope stability limit,” Nature, 2026. DOI: 10.1038/s41586-026-11213-z

Tags: astronomy, James Webb Space Telescope, exoplanets, 51 Eridani b, direct imaging, NIRCam

More Like This

James Webb found a planet where rock clouds form each morning and clear by night

James Webb mapped 164,000 galaxies to chart the cosmic web back a billion years

A lava planet that should be bare rock reflects 74% of its starlight, Webb finds

Webb watched water clouds thicken and thin on a world 7.4 light-years away

Webb caught a galaxy several times the Milky Way’s mass — and it isn’t spinning

Planet formation predicted plenty of small Kuiper Belt objects — Webb found fewer