AI finds 118 new exoplanets
- University of Warwick astronomers used the RAVEN AI pipeline on NASA TESS data and validated 118 exoplanets, including 31 planets not previously detected. (phys.org) - RAVEN searched 2.2 million main-sequence stars from TESS’s first four years, then produced more than 2,000 strong candidates, nearly 1,000 entirely new. (phys.org) - It matters because TESS still has nearly 7,931 project candidates, and better triage tools can turn public archive data into confirmed worlds faster. (exoplanetarchive.ipac.caltech.edu)
Exoplanet hunting just got a lot more automated. A team at the University of Warwick used a new AI-driven pipeline called RAVEN to comb through NASA’s TESS archive and validate 118 planets in one sweep, including 31 that had not been detected before. (phys.org) That matters because the bottleneck in exoplanet science is no longer raw data — it’s sorting real planets from all the fakeouts hiding in the same light curves. This week’s news is that one pipeline seems to do that at scale. ### What is RAVEN actually doing? TESS watches stars for tiny dips in brightness. A planet crossing in front of its star can cause one of those dips, but so can lots of other things — eclipsing binary stars, noise, or instrumental quirks. (exoplanetarchive.ipac.caltech.edu) RAVEN starts with a standard transit search, then runs machine-learning classifiers trained on simulated planets and false positives, and finally applies statistical validation to decide which signals are likely real planets. Basically, it tries to handle detection, vetting, and validation as one pipeline instead of three separate chores. ### Why is that hard? Because the sky is full of lookalikes. A dimming event can resemble a planet even when it is really two stars eclipsing each other or some other astrophysical impostor. (phys.org) Human vetting works, but it is slow, and TESS produces a huge volume of targets. That is exactly the kind of problem where machine learning helps — not by replacing astronomers, but by ranking and filtering the pile so humans spend time on the most promising cases. ### What data did they search? The team focused on more than 2.2 million well-characterized main-sequence stars observed in TESS full-frame images during the mission’s first four years, covering sectors 1 through 55. (phys.org) They looked for short-period planets with orbits from about 0.5 to 16 days. That means these are mostly close-in worlds, not slow, Earth-like planets on year-long orbits. But short-period planets are exactly where TESS is strongest, so this is the right place to do a large uniform census. ### Why does 118 matter? It is a big batch, but the more interesting number may be 31. Those are planets the pipeline newly detected rather than merely reclassifying from an existing list. (phys.org) On top of that, RAVEN produced over 2,000 high-quality candidates, with nearly 1,000 entirely new. So the headline is not just “AI found some planets.” It is “AI built a much larger queue for future confirmations.” ### What kinds of planets showed up? Several of the validated worlds sit in especially interesting categories. The sample includes ultra-short-period planets that whip around their stars in under 24 hours, planets in the so-called Neptunian desert where theory says they should be rare, and compact multi-planet systems with close-orbiting planetary pairs. (phys.org) Those are useful because oddball populations are where formation theories get stress-tested. ### Is this replacing telescopes and follow-up work? No — and that is the catch. Validation is powerful, but many candidates still need follow-up observations from other telescopes to lock things down. (phys.org) NASA’s own ExoMiner++ effort makes the same point: AI is great at pulling likely planets out of giant public archives, but candidate status and confirmation are not the same thing. The real win is speed. AI moves astronomers from “where do we even look?” to “which targets deserve scarce telescope time first?” ### Why now? Because the archive is getting too large to handle the old way. NASA’s Exoplanet Archive listed 6,286 confirmed planets as of May 7, 2026, along with 7,931 TESS project candidates. (phys.org) TESS has already found nearly 700 confirmed exoplanets, but the public data still holds far more signals than teams can inspect manually. RAVEN is part of a broader shift — the discovery problem is becoming a software problem as much as an observing problem. ### Bottom line? The big change is not just 118 more planets. It is that exoplanet surveys are starting to get industrial-scale sorting tools. If RAVEN and similar systems keep working, the next decade of planet discovery may come less from brand-new telescopes than from smarter ways of mining data we already have. (science.nasa.gov) (phys.org) (exoplanetarchive.ipac.caltech.edu)
