New computational method reveals a planet hidden in Webb telescope's own data

A team of astronomers and computer scientists led by researchers at Northwestern University and the California Institute of Technology has pulled a previously undetected giant planet out of archival images from the James Webb Space Telescope, using a computational method that succeeded where a standard analysis of the same data had failed. The technique, called FLIP, for Physics-based Learning for Imaging Planets, is described in a paper published Nov. 4 in the journal Nature, and strips residual starlight out of Webb images with far greater precision than existing software; its developers say it could extend the telescope's planet-hunting reach without a single additional night of observing time.

The method improved the detection sensitivity for known planets by a factor of roughly two to 15 compared with conventional processing, according to the Caltech announcement of the findings. In its most striking demonstration, the team reprocessed existing Webb NIRCam observations of the star system 51 Eridani and recovered a clear detection of its known gas-giant planet, 51 Eridani b, a world that standard glare-subtraction techniques had missed entirely in that particular dataset. The planet orbits an F-type star roughly 97 light-years from Earth and was first photographed in 2014 using a ground-based instrument, the Gemini Planet Imager, before Webb's more sensitive infrared eye examined the system again more than a decade later.

Starlight as the obstacle

Photographing a planet next to its star is, as the researchers describe it, an exercise in picking out a faint candle beside a floodlight. A Sun-like star outshines an Earth-sized planet orbiting it by a factor of roughly 10 billion, and even a bright young gas giant can be millions of times fainter than its star. Space telescopes use coronagraphs to block the bulk of a star's light, then subtract whatever glow remains using a reference image of a different, planet-free star observed under similar conditions.

That subtraction is where things usually go wrong. Tiny shifts in a telescope's optics, caused by temperature changes or mechanical settling of only a few nanometers, change the pattern of scattered starlight between when the reference star and the target star were observed. Webb's optical system is engineered to be stable to the nanometer level, among the steadiest ever flown, but the Caltech-led team's findings indicate that even that stability still imposes a hard limit on how faint a companion object can be detected using the standard approach. Of the thousands of exoplanets confirmed so far, the great majority were found indirectly, through the dimming of starlight during a transit or the slight gravitational wobble of a star, rather than through a direct photograph; only a small fraction, mostly young, wide-orbiting giant planets, have ever actually been imaged.

A digital twin of the telescope

FLIP sidesteps the mismatch by building what its creators call a digital twin of Webb's Near Infrared Camera: a physics-based computer model that simulates how the telescope's optics would scatter a given star's light, accounting for millions of possible small instrumental imperfections. The model is fitted directly to the actual data from the observation in question, rather than to a separate reference star, and the predicted starlight pattern is then subtracted in its place. Because the technique works on images Webb has already captured, astronomers can apply it retroactively to the telescope's growing public archive rather than competing for new, scarce observing time.

Rodrigo Ferrer-Chavez, a graduate student at Northwestern University and the paper's lead author, said the result demonstrates that small instrument instabilities need not be the limiting factor they have long been assumed to be. "We now have a tool that lets us discover and study planets that we were not able to 'see' before," he said, according to the Caltech-based release carried by ScienceDaily. Jason Wang, an astronomer at Northwestern and a co-author on the study, described the result more bluntly: "I feel we just unlocked a new frontier." Ferrer-Chavez has also said his first attempt at the problem, training machine-learning models to predict stellar glare directly from scratch, did not work, and that the eventual approach instead combined physical modeling of the telescope's optics with data-driven fitting to real observations.

"What's exciting is that we're showing how computation can help overcome limitations of the telescope itself," said Katie Bouman, a Caltech professor of computing and mathematical sciences and a co-author of the study.

Who stands to benefit

The most immediate beneficiaries are astronomers who already have access to Webb's public data archive, which holds years of accumulated observations of nearby stars, many taken for other purposes and never searched exhaustively for faint companions. Because FLIP requires no new telescope time, researchers elsewhere can, in principle, reprocess existing datasets on their own computers to check for planets, disks or other faint structures that earlier analyses may have overlooked. Additional co-authors on the paper came from MIT and the Space Telescope Science Institute in Baltimore, which operates Webb's science program, reflecting the project's reliance on both computational expertise and institutional access to the telescope's calibration data. The work was supported by the Space Telescope Science Institute, the Sloan Foundation, an Amazon AI4Science Discovery Award and the National Science Foundation.

Looking further ahead, the team says the approach is not specific to Webb. They intend to apply FLIP to NASA's Nancy Grace Roman Space Telescope, which is expected to begin science operations within the next couple of years, and to ground-based observatories that face similar glare problems. The results are also being weighed as input to the design of NASA's proposed Habitable Worlds Observatory, a still-conceptual mission intended to directly image roughly two dozen potentially habitable planets around nearby stars and search their atmospheres for chemical signs of life. Easing the stability requirements such a telescope would otherwise need could simplify its design and lower its cost, a consideration NASA has flagged as central to the mission's feasibility during early planning.

Caution alongside the enthusiasm

Outside coverage of the paper has been largely positive but not uncritical. A report from the science news site Universe Today noted that FLIP's gains varied depending on the target, the instrument configuration and how close a candidate planet sat to its star in the image, meaning the technique is not a uniform multiplier across all observations. The method also did not reach the theoretical photon-noise limit, the point at which detection is constrained only by the number of light particles collected rather than by instrument artifacts, in any of the datasets tested so far. Extending the approach to the much fainter signal of an Earth-sized, rocky planet remains an open question that has not yet been demonstrated, and the authors have been careful to frame this result as a proof of concept rather than a finished tool for that purpose.

The results likewise have not yet been independently replicated by teams outside the collaboration, which is typical for a newly published technique and does not undercut the specific claim at the heart of the paper: that a known planet invisible to standard processing became visible once a more detailed model of the telescope was used in its place.

What comes next

The authors say they plan to rerun FLIP across a broader swath of Webb's archive to look for additional overlooked planets and debris structures, and to make the underlying code available so other research groups can apply it to their own datasets rather than waiting on the original team. Mission planners for the Habitable Worlds Observatory, still years from a final design, are expected to factor the technique's demonstrated gains into ongoing studies of how stable the mission's optics will actually need to be, a question with direct bearing on engineering cost and schedule. For now, the clearest result is narrower but concrete: a planet that a space telescope's standard software looked straight past has been found sitting in data NASA had already collected, processed and released to the public years earlier.