Northwestern-led team demonstrates FLIP exoplanet imaging method

Nature published Imaging exoplanets below the telescope stability limit on Nov. 4, 2026, a Northwestern University-led study by Rodrigo Ferrer-Chavez with researchers from Caltech, MIT and the Space Telescope Science Institute that describes a computational method for imaging exoplanets below telescope stability limits.

FLIP stands for Physics-based Learning for Imaging Planets. The method pairs physical models of a telescope with machine-learning optimization to model and subtract starlight, using automatic differentiation (a technique that calculates derivatives of functions defined by computer code) to tune millions of parameters describing mirror figure, optical alignment, thermal wavefront errors and other instrument instabilities. JWST has an effectively infinite set of potential imperfections, from mirror segment tilts and pistons to wavefront phase errors and optical path differences, all computed through wave optics. The team stresses FLIP is not a text-generating AI like ChatGPT; it combines physics with computational techniques to interpret observational data.

Standard imaging blocks starlight with coronagraph masks and subtracts light measured from a reference star, but the reference star is never observed at the same time and thermal state as the target. Residual diffracted starlight creates shifting speckles (grainy patterns of residual starlight) around the star as the telescope flexes and thermally cycles, patterns that can mimic or overwhelm dim companion signals, especially young, cool, dim planets observed in infrared. Coronagraph effectiveness is further limited by noise from subtle variations in telescope configuration and pointing between observations.

A point spread function describes how a telescope distorts a single point of light into a slightly blurry, spread-out image because of hardware imperfections and pointing limits. One way to determine it is to image a known target, but hardware changes over time and wavelength variation limit that approach; modeling hardware behavior directly involves too many parameters to be computationally tractable.

The team demonstrated FLIP on James Webb Space Telescope Near-InfraRed Camera coronagraph observations, achieving sensitivity two to fifteen times superior to standard techniques on actual on-sky coronagraphic data rather than simulations alone. In archival Cycle 1 observations of 51 Eridani, standard reference-star subtraction did not reveal the planet; its expected location was marked with an empty dashed circle. Reprocessing with FLIP showed a clear point source at precisely the predicted location. The planet appeared at 4.3 micrometers with a signal-to-noise ratio of 6.3, improving to about 2.1 times the conventional method under different conditions.

51 Eridani b is a known gas giant, roughly three times Jupiter’s mass with a temperature around 750 Kelvin, or 890 degrees Fahrenheit. It orbits at a distance comparable to Saturn’s position around a star about 96 light-years from Earth in the constellation Eridanus. The planet’s discovery dates to 2015, and it carries the second-strongest methane signature of any directly imaged exoplanet, with water vapor, methane and carbon monoxide in chemical disequilibrium that make it a benchmark for cool atmospheric chemistry.

The up-to-fifteen-times figure comes from a simulated planet injection and recovery test at about 0.5 arcseconds from the star TYC 5899 and does not mean fifteen new planets were found or that planets are visible fifteen times farther away. The improvement depends on target geometry, filters and coronagraph configuration, and the quality of reference observations.

A factor of two means FLIP can detect planets roughly one-quarter as luminous as the faintest existing methods can find at the same angular separation, while fifteen-fold means the effective depth of a given JWST observation is vastly increased without additional telescope time. Verification covered three star systems and four observation datasets, and the authors say equivalent performance has not been proven for all planets and telescopes.

Over-adjusting calculations to remove starlight risks erasing the planet’s actual light, a problem related to overfitting or excessive signal subtraction. FLIP requires GPU-based calculations, and the processing burden cannot be ignored. Conventional image processing techniques such as KLIP have years of proven use and computational advantages, so conventional methods will not become entirely unnecessary. Complete starlight removal is impossible because light spreads out due to slight telescope misalignments or minute changes in the optical system.

A young bright giant planet can emit only about one-hundred-thousandth of its host star’s near-infrared light, so stellar glare easily buries the planet signal. Exoplanet imaging is the main method for detailed characterization of planets orbiting beyond one astronomical unit, capturing the planet’s own light. It is the only technique that can analyze a planet’s atmosphere across a wide range of orbits, forming the foundation for future searches for signs of life. Addressing starlight suppression will increase the number of planets amenable to direct imaging as a notable step toward imaging exo-Earths with future observatories.

FLIP can be extended to the Nancy Grace Roman Space Telescope and the Habitable Worlds Observatory, potentially relaxing the strict stability requirements those missions require. Roman launched on Aug. 30, 2026 aboard a Falcon Heavy rocket and is in commissioning, with its Coronagraph Instrument expected to begin initial power-on testing in the coming weeks. Roman’s Coronagraph Instrument is the first space-based coronagraph designed to demonstrate technology for imaging Earth-like planets, and its results will help shape the Habitable Worlds Observatory coronagraph design. The Habitable Worlds Observatory is NASA’s designated next flagship astrophysics mission, planned for launch in the 2040s at an estimated cost of at least $11 billion, and aims to directly image at least 25 potentially habitable Earth-sized worlds around nearby Sun-like stars.

Directly imaging an Earth-like world next to a Sun-like star requires a contrast of roughly one part in ten billion. A technique that recovers signals below what hardware alone permits could be the difference between the Habitable Worlds Observatory meeting its science goals or falling short. If FLIP-like approaches can remove starlight after images are taken, engineers may not need to solve all instability issues at the hardware level.

The full archive of existing JWST coronagraphic observations is substantial and publicly available, and FLIP could reveal planetary companions in systems where none were previously thought to exist. If even a handful of archival datasets harbor planets missed by standard methods, the archive may be considerably richer than current surveys suggest. The team plans to apply FLIP to the full JWST NIRCam coronagraph archive and extend it to the Roman coronagraph, ground-based adaptive optics on eight-to-ten-meter telescopes and the ESO Extremely Large Telescope under construction in Chile. Future use on higher-performance observation facilities is possible, but performance and reliability must first be verified with much more observational data.

The arXiv preprint version 1 was submitted on Mon, 9 Nov 2026 at 21:26:14 UTC. Ferrer-Chavez first attempted to train purely data-driven machine learning models to predict speckle patterns, but that approach failed; co-author Jason Wang described the failure as a necessary precondition that eliminated unproductive directions before the successful approach emerged. The collaboration spans Caltech, MIT and the Space Telescope Science Institute, and includes Caltech professor Katie Bouman, who helped develop core algorithms behind the Event Horizon Telescope’s first black hole image in 2019, and Dimitri Mawet.

The Space Telescope Science Institute, the Alfred P. Sloan Foundation, an Amazon AI4Science Discovery Award and the National Science Foundation supported the research. That mix of an institute, a philanthropic foundation, an industrial AI award and a federal science agency reflects the interdisciplinary nature of the work.

The authors call this the first demonstration, as far as they know, of improved sensitivity on actual on-sky coronagraphic imaging using this kind of approach. The claim carries the authors’ own assessment of prior work.

The demonstration involved a known giant planet, not a discovery of small Earth-like planets or extraterrestrial life. “these little imperfections or instabilities are not a limiting factor in the kinds of planets we can image,” said Rodrigo Ferrer-Chavez.