Astronomers use AI’s training tools, but no AI, to get more out of Webb’s planet images

Northwestern University researchers led by assistant professor Jason Wang developed FLIP (Physics-based Learning for Imaging Planets), a method that uses gradient descent — the same optimization tool used to train AI models — to model and subtract starlight glare from James Webb Space Telescope NIRCam images, without employing AI itself. FLIP fits a simulated point spread function with up to 10^6 parameters to real star data in just over half an hour on a current GPU, then fine-tunes on the specific science image, addressing the optics drift that undermines the two-decade-old reference-star subtraction technique. The work matters because planets are a million times fainter than their stars, and NASA's proposed Habitable Worlds Observatory would need picometer-level stability 100 to 1,000 times better than Webb, making software correction an alternative to costly hardware.

To take a picture of a planet orbiting another star, you mostly block out the star. The James Webb Space Telescope carries coronagraphs for exactly this. But some starlight always leaks past the block as a faint haze, and the planet is sitting in that haze. Rodrigo Ferrer-Chavez, a Northwestern University PhD student, puts the scale of the problem simply: “Planets are a million times fainter than their stars.” Even a little leftover glare is a lot of glare compared with that. For roughly two decades the standard fix has been a sort of subtraction trick. You point the telescope at a second star with no planets you care about, photograph its glare, and subtract that image from the one of your target star. Whatever is left over should be the planet. Jason Wang, the Northwestern assistant professor who runs the group, describes it as “observing a different star and using its image as a reference,” according to the university’s release https://www.eurekalert.org/news-releases/1150080 . The catch is that the telescope you used for the reference star isn’t quite the same telescope a little later. Optics shift by tiny amounts, so the glare patterns never match exactly, and the bit that doesn’t cancel out can hide a faint planet. “We try to make telescopes extremely stable, but there are extreme engineering challenges that we haven’t quite solved yet,” Wang said. So you have two options. One is to build a stiffer telescope. That is what NASA’s proposed Habitable Worlds Observatory may need: picometer-level stability, around 100 to 1,000 times better than Webb, according to Northwestern. That is a hardware problem, and hardware problems in space cost a lot of money. The other option is to accept that the telescope wobbles and model the wobble. Borrowing the GPUs The second option used to be impractical. A proper physical model of a telescope camera has a vast number of knobs. The team’s model of Webb’s NIRCam has 10^6 possible parameters, by John Timmer’s count in Ars Technica https://arstechnica.com/science/2026/11/getting-software-to-help-block-out-a-star-so-we-can-see-exoplanets/ ; Northwestern says “millions.” Turning a million knobs until the simulated glare matches the real glare is the sort of job that sounds like a career. Except that one industry has spent the past several years getting extremely good at turning billions of knobs. Training a large neural network is, mechanically, an exercise in adjusting enormous numbers of parameters with a technique called gradient descent. Because that technique is central to the AI business, a lot of effort has gone into making it run fast on GPUs. The astronomers simply borrowed it. “We used the same computational tools that researchers use to train AI models, but we did not actually use AI,” Wang said. Nothing here is a black box guessing what a planet looks like. The physics is written down, and the optimizer just fills in the numbers. Sure, it’s slightly funny that the main beneficiary of chatbot-driven GPU tuning so far might be exoplanet imaging. But that is how positive externalities work: someone else pays to build the tool, and you get to use it. The method is called FLIP, which officially stands for Physics-based Learning for Imaging Planets, although the letters don’t line up. It begins with a simulated point spread function, which is the pattern a point of starlight makes once it has gone through the telescope’s optics. It then uses gradient descent to fit that to data from a real star, which takes a bit over half an hour on a current GPU. Last, it fine-tunes the result on the specific science image being analysed. That final step matters because some variables change from one observation to the next. That drift is the whole reason the reference-star method falls short. In Ferrer-Chavez’s words, FLIP “models how light travels through the telescope and accounts for those tiny changes.” Science left on the table The team tested it on archived Webb data for three young stars. Two of them, HIP 65426 and 51 Eridani, have massive planets that are still warm from forming, which is why they could be imaged before. The third, TYC 5899, has no known planets but is known to vary a lot between observations, which makes it a good stress test. The size of the improvement depends on how you measure it. Ars reports that FLIP raised the signal-to-noise ratio by a factor of 1.5 to 2 depending on the observation, and improved sensitivity on TYC 5899 by a factor of 15. Northwestern’s summary says sensitivity improved by 2 to 15 times compared with conventional processing. Either way, from a telescope that is already up there and already paid for, that is a big gain. The headline result is 51 Eridani b at 4.3 microns, a wavelength used to look for carbon dioxide. Carbon dioxide in the planet’s atmosphere absorbs light there, which dims the planet. In Webb data roughly two years old, the standard processing method, known as KLIP, could not pull it out of the glare. FLIP detected it at better than 5 sigma, Phys.org reported https://phys.org/news/2026-11-giant-planet-emerges-year-jwst.html . Some headlines call it a hidden planet, but it isn’t a new discovery: 51 Eridani b was found years ago with ground-based telescopes. What’s new is seeing it in Webb’s data at that wavelength. “We showed that we can actually recover a planet that people previously couldn’t see,” Wang said, and he called the missed signal “science left on the table.” That phrase is the interesting part, I think. If the noise drops, dimmer planets become visible too, which means older or lower-mass ones that nobody has spotted yet. Some of them may already be in the archive. “It’s not that they don’t exist,” Ferrer-Chavez said. “It’s that we cannot see them.” The researchers also think FLIP could be adapted to show where it is changing the model most. That could tell engineers what Webb’s hardware is actually doing, which is more or less a free diagnostic tool for the instrument. Ground-based astronomers already fire lasers into the sky to make fake stars https://rews.cc/a/chile-s-very-large-telescope-is-firing-up-to-seven-lasers-to-44f296 so they can correct for the atmosphere in real time. FLIP makes its corrections later, in software, after the photons have arrived. The work, published in Nature on November 4 https://astrobiology.com/2026/11/05/erasing-the-stars-to-find-exoplanets/ under the title “Imaging exoplanets below the telescope stability limit,” was funded by the Space Telescope Science Institute, the Sloan Foundation and the National Science Foundation. For a telescope a very long way from anyone with a screwdriver, a better model on a graphics card is the only upgrade on offer, and it’s also the cheapest.