The hardest problem in direct imaging of exoplanets is not the planet. It is the star, which is millions of times brighter and sits, from the telescope's point of view, in almost exactly the same place.
Coronagraphs handle the bulk of it: a mask in the optical path blocks the star's light so that faint things next to it become visible at all. What they cannot remove is the residue — starlight that scatters past the mask because the telescope's optics are not perfectly stable. The mirror flexes by nanometres as the observatory warms and cools. The pointing drifts. The result is a field of speckles, roughly as bright as the planet you are looking for.
Two ways to know your own blur
To subtract speckles you have to know the point spread function: the mathematical description of how the telescope smears a single point of light. There are two ways to get it.
You can image a known star and derive the function from that reference — good at removing noise, but limited, because the telescope's behaviour changes over time and the distortion depends on wavelength. Or you can build a physical model of the hardware and compute the function from the model. That second route is the more robust one and the one nobody could make work, because a modern telescope is a machine with far too many parameters to tune by hand.
That is the sentence worth pausing on. Too many parameters to tune by hand is not a description of a problem unique to astronomy. It is a description of a neural network before gradient descent.
FLIP
A team of researchers built a physical model of NIRCam, Webb's near-infrared camera, and gave it on the order of a million tunable parameters. Then they did what you do with a model that has a million knobs: they started from a simulated point spread function and optimised the whole parameter set by gradient descent against data from an actual star, exploiting the same GPU tooling that made large-model training cheap. On a current GPU the fit takes a little more than half an hour.
The optimised model is then fine-tuned on the specific scientific image being analysed, because some parameters differ from observation to observation — pointing, thermal state, the particular dither pattern. The method is called FLIP, short for Physics-based Learning for Imaging Planets, which the authors present as a name rather than an acronym.
What it bought
The team tested on three targets: HIP 65426 and 51 Eridani, both young stars with known directly imaged giant planets, and TYC 5899, a star with no known planets that is known to vary strongly between observations.
Using Webb data, FLIP raised the signal-to-noise ratio by a factor of 1.5 to 2 depending on the observation. At a wavelength usable for detecting carbon dioxide, it produced the first detection of one of those planets in the Webb data. For TYC 5899 — the hard case, a star whose own variability mimics the noise you are trying to remove — sensitivity improved by a factor of 15.
That last number is the interesting one less for what it says about that star than for what it implies about the pipeline: the method works best exactly where reference-star calibration fails, which is on stars that change.
Why it matters beyond prettier pictures
The obvious consequence is a larger catalogue. Lower noise means dimmer planets become detectable, and dimmer at a given age means older or less massive — filling in the part of the mass–age diagram where direct imaging currently has almost nothing.
The less obvious consequence is diagnostic. The researchers note they can modify FLIP to report where it is making the largest corrections to the model. In a hardware context that is telemetry: the optimiser is effectively telling you which part of the optical train has drifted away from its specification, from the far side of the solar system, without sending anyone to look. For a telescope that cannot be serviced, that is close to the most valuable output a software package can produce.
What this is not
It is not new hardware, and it does not create photons. The gain comes entirely from modelling the instrument well enough to subtract a nuisance term that was previously being subtracted crudely — no additional observing time, no new instrument, a better estimate of what the telescope did to the light.
And it is validated on three targets in one instrument. Whether a model fitted this way generalises across NIRCam's filters, across observatory thermal cycles, and across instruments is the next question, and the one that decides whether this becomes standard practice at every ground- and space-based observatory with a coronagraph, or stays a well-cited method for three interesting stars.
The work is published in Nature (DOI: 10.1038/s41586-026-11213-z).
