An AI Pipeline Named RAVEN Just Pulled 31 New Planets Out of NASA TESS Data and Counted the Neptunian Desert

Astronomers at the University of Warwick just announced that an AI pipeline named RAVEN dug through four years of NASA TESS observations and pulled out 31 brand new planets, plus more than 100 confirmed worlds in total, plus around 1,000 fresh candidates that nobody had spotted yet. The data had been sitting in NASA archives the entire time. Humans missed them. The AI found them. And the part nobody is saying out loud is the funny one: the bot is also more honest about what it does not know.

The study was published on May 3 in Monthly Notices of the Royal Astronomical Society, three papers at once. RAVEN scanned 2.2 million stars looking for the tiny dip in brightness that happens when a planet crosses in front of its host. Spotting one is easy. Proving the dip is actually a planet and not a binary star, a sunspot, or an instrument glitch is the hard part. Most pipelines handle one slice of that workflow. RAVEN handles the whole thing in one shot.

The Neptunian Desert Just Got a Census

If you have not heard of the Neptunian desert, here is the short version: it is a region of orbital space, hot and close to the host star, where Neptune-sized planets almost never exist. Smaller planets are common there. Bigger ones too. Mid-sized worlds in tight, fast orbits are weirdly missing. The leading theory is that their atmospheres get stripped by the star until they shrink into something else.

Until now nobody had counted them. Kaiming Cui, who led the desert sub-study, put it bluntly: “For the first time, we can put a precise number on just how empty this ‘desert’ is.” The number is 0.08 percent of Sun-like stars. Eight in ten thousand. The previous range had been a guess based on Kepler data with uncertainty up to ten times larger. RAVEN squeezed it down because it processed the entire TESS catalog the same way, every star, every signal, no human deciding which ones were “interesting enough” to bother with.

This is the part that should land. We have been talking about AI doing science for years, and most of it has been benchmarks and overclaimed press releases. (See also the AI startup whose coding agent wiped a production database in nine seconds and confessed in all caps.) RAVEN is not that. It produced reproducible occurrence rates that map onto the real galaxy, and the team trained it on simulated TESS data so it would learn what fakes look like before going near anything real.

What RAVEN Actually Does That Older Tools Did Not

Andreas Hadjigeorghiou, who built the pipeline, summarized his daily problem: “The challenge lies in identifying if the dimming is indeed caused by a planet in orbit around the star or by something else.” That something else has a long list. Two stars eclipsing each other behind a foreground star. A grazing transit that mimics a smaller planet. Detector noise. Cosmic rays. Each used to need a separate paper, a separate human eyeball, a separate vetting committee.

RAVEN does the detection, the false-positive rejection, and the statistical validation as one continuous pipeline. David Armstrong, also at Warwick, described the upside in a sentence that should be on every machine learning textbook: “RAVEN allows us to analyze enormous datasets consistently and objectively.” Consistency and objectivity are the boring superpowers. A graduate student looking at a transit signal at 2 a.m. on a Tuesday will not vet it the same way a different student vets it at 10 a.m. on a Friday. RAVEN does. Every single time.

The 31 new confirmed planets include some genuinely strange worlds. Some whip around their host stars in less than 24 hours, meaning a year on those planets is shorter than a single Earth day. A kitten born on one of these things would have lived through five years of seasons by lunchtime. Several of the new finds sit inside the Neptunian desert itself, which makes them rare twice over.

The Honest Uncertainty Bit

Here is the part that does not usually make headlines. RAVEN does not return a yes-or-no answer for each star. It returns a probability, with error bars. Marina Lafarga Magro, who led the planet sample team, framed the result as a starting point: “This represents one of the best characterized samples of close-in planets and will help us identify the most promising systems for future study.” Translated: we have a clean shortlist now, go look at these with bigger telescopes. The James Webb crowd is presumably already updating their target lists.

What is sneakily impressive is that the team did not just publish the planets. They published the candidate catalog, the rejection rates, the simulated training set, and the methodology. If you want to know how AI science should look from the outside, this is the template. Compare it to the closed-shop approach we saw earlier this year when Peter Thiel poured 140 million dollars into floating ocean AI data centers with vague promises about “frontier compute.” Open methodology beats vague promises every Tuesday.

Why TESS Data Was Hiding This Stuff

NASA’s Transiting Exoplanet Survey Satellite has been operating since 2018, generating a fire hose of data. Early TESS analyses focused on the obvious candidates, big planets around bright nearby stars where the signals are loud. The Warwick team essentially asked, what if we go back through the boring stars too. The boring stars turned out to host most of the new finds. There is a lesson in there about not throwing away your seemingly empty datasets. We made a similar argument when we covered Gen Z buying 25 dollar wired earbuds like vinyl records. Old stuff is not necessarily done with you.

The next phase is pointing RAVEN at the rest of the TESS archive plus data from PLATO when ESA’s planet hunter starts returning observations in 2027. By the late 2020s, the bottleneck for exoplanet science will not be discovery, it will be follow-up: who gets time on JWST and the Habitable Worlds Observatory to actually study these things.

The Pudgy Cat Take

Most “AI does science” stories are press releases dressed up as discovery. RAVEN is the rare case where the AI did the actual work that humans had not done, where the team published their methods openly, and where the result is a more honest picture of the galaxy, not just a longer list of planets. The Neptunian desert was always going to get measured. AI just made it happen six years sooner.

The cat verdict: when an AI is built to admit it might be wrong, fed simulated fakes until it learns the difference, and then asked to find new things, it can pull off real science. When an AI is built to sound confident, fed scraped slop, and asked to write a press release, it produces the floor. Same technology, very different culture. Astronomy got the good version this week.