The Transiting Exoplanet Survey Satellite has spent four years watching stars blink. Every time a planet crosses the face of one, the starlight dims — by a fraction of a percent, for hours, and then it's gone. TESS logged those dips across 2.2 million stars. Most of them, nobody had quite gotten around to checking.
2.2 Million Stars and No One Counting
TESS launched in April 2018 with a mission that sounds simple: watch the whole sky, find planets. It carved the sky into sectors, spent 27 days on each one, and produced a river of photometric data that astronomers have been sorting through ever since. But "sorting through" has always meant something more modest than it sounds. TESS produces light curves for hundreds of thousands of stars per sector. A single sector can surface thousands of candidate signals, each of which could be a planet transit, a false alarm, a background eclipsing binary, or an instrument artifact. Confirming even one requires careful vetting. Doing it at scale has been, until recently, mostly beyond reach.
The result is a backlog. Thousands of promising signals sit in the TESS archive in a state of permanent maybe: flagged as planet candidates, never confirmed or ruled out, waiting for the kind of systematic attention that one team can't supply fast enough. Before a candidate can be called a planet, astronomers have to rule out every other explanation for the light-curve dip. That process, planet validation, involves comparing the likelihood of a planetary origin against the likelihood of every known false-positive scenario. It's careful, computationally intensive work. And there's more of it every month.
The University of Warwick team had a different question. Not "can we confirm individual candidates faster" but "what if we built something that could do all of it at once?"
What RAVEN Learned to See
RAVEN stands for RAnking and Validation of ExoplaNets. The name is workmanlike; the actual system is more interesting. Most planet-hunting pipelines separate their steps: detect a transit signal, then vet it by hand or algorithm, then validate it statistically in a separate process. RAVEN does all three in one workflow, a unified machine learning architecture trained on TESS data and tested against pre-classified candidates.
The validation problem is subtle. When TESS measures a dip in starlight, the most common culprit isn't a planet. It's another star. A background eclipsing binary, two stars orbiting each other and happening to lie in the same patch of sky as the target star, can produce a light curve that looks almost identical to a planetary transit. Instrument noise can do it too. So can a star physically orbiting the target, eclipsing it with geometry that mimics a small planet. Distinguishing a real planet from these impostors is what validation is for, and it's what makes the process slow.
RAVEN was trained to compare the probability of a planetary scenario against a list of false-positive scenarios simultaneously, then output a ranking. On an independent external test set of 1,361 pre-classified TESS candidates, the pipeline achieved 91% overall accuracy. That's not perfect. But it's good enough to be genuinely useful at scale, in a way that no prior pipeline has managed.
When TESS sees a star dim, it can't tell you why. A planet crossing the star's face is one explanation. A background binary star aligned with the target is another — and statistically more common. RAVEN's job is to assign a probability to each scenario. When the planet scenario wins by a wide enough margin, the candidate is validated. The margin required has been calibrated so that fewer than one in a hundred validated planets should turn out to be an impostor.
The catalog that came out of this is now public. Every one of the 118 confirmed planets, and the 2,000-plus candidates that didn't quite reach confirmation threshold, is available for follow-up. Ground-based radial velocity measurements, high-resolution imaging, spectroscopy — any of these could move a candidate from "probable" to "confirmed" or "ruled out." RAVEN didn't end the search. It started several hundred new ones.
The paper appeared in the Monthly Notices of the Royal Astronomical Society. The team is already in early discussions about adapting RAVEN for the European Space Agency's PLATO mission, scheduled to launch in late 2026.
The Worlds We'd Been Sliding Past
The 118 planets RAVEN confirmed are mostly close-in worlds: planets completing their orbits in under 16 days, which puts them inside the orbit of Mercury relative to our own sun. At those distances, they're too hot for life as we know it. But that's not why they matter. They matter because they're real, they were missed, and counting them correctly answers questions that astronomers have been arguing about for a decade.
"Getting a more accurate census of these close-in worlds is fundamental to understanding how planetary systems form and evolve."— University of Warwick press release, May 2026
Among the 118, the team found examples of rare planet types. Ultra-short-period planets: worlds that complete an orbit in less than a single Earth day, skimming their stars so closely that surface temperatures can exceed the melting point of iron. And a handful of Neptunian desert planets, which sounds like a contradiction in terms. The Neptunian desert is a gap in the exoplanet distribution: a region of parameter space (short orbital period, Neptune-like radius) where planets should be common but aren't, presumably because the intense stellar radiation at close range strips their atmospheres faster than they can form. Neptunian desert planets appear around only 0.08% of sun-like stars. Every one that survives long enough to be found is a data point in a still-incomplete theory of atmospheric erosion.
| Planet Type | Occurrence Rate | Key Property |
|---|---|---|
| Close-in (period <16 days) | 9–10% of sun-like stars | Confirms Kepler estimates with 10× less uncertainty |
| Ultra-short-period (<1 day) | Rare subset | Surface temperatures can exceed iron's melting point |
| Neptunian desert | ~0.08% of sun-like stars | Survive atmospheric stripping at extreme proximity |
Figure 1 — Planet types in RAVEN's confirmed catalog
One of the cleaner results from RAVEN's full census: about 9 to 10 percent of sun-like stars host at least one close-in planet. That number aligns with the earlier estimates from NASA's Kepler mission. But Kepler covered a smaller patch of sky with a less complete statistical treatment. RAVEN reduces the uncertainty in that occurrence rate by up to a factor of ten. That might sound like a technical accounting fix, but occurrence rates are how we calibrate planet-formation models. Get the census wrong and you get the theory wrong.
When the Map Gets Full
TESS is not finished. It continues to observe, sector by sector, building out a dataset that will outlast the spacecraft's operational life in its scientific usefulness. Every sector adds more candidates, more light curves, more dips in need of someone's attention. The backlog isn't going to shrink without automated tools that can keep pace with the incoming stream. RAVEN is the first system to show it's possible to do that at full scale.
PLATO changes the picture again. The European Space Agency's mission, scheduled for late 2026, is designed for something TESS wasn't optimised for: finding Earth-sized planets in Earth-like orbits around sun-like stars. That means watching individual stars for years, not weeks. The transit signals are smaller, the orbits are longer, and the data volume is enormous. The team at Warwick is already looking at whether RAVEN's architecture can be adapted to handle what PLATO will produce.
There's a quieter implication in all of this. The 118 planets RAVEN found weren't hidden in some dark corner of the data that no one ever reached. They were sitting in TESS's public archive, in sectors that had been examined before. The bottleneck wasn't access. It was the capacity to look carefully at everything, simultaneously, without missing the signals that didn't quite rise to the top of anyone's priority list. RAVEN's 2,000-plus candidates include roughly 1,000 that are entirely new: signals that hadn't been flagged by any prior analysis. Those aren't just unvalidated. They're unseen.
The universe has been broadcasting this whole time. The archive is full of it. The question was always whether we could build something patient enough to listen.
Buy me a coffee