TESS AI finds 10,000-plus exoplanets

Exoplanet hunting is a data problem now. TESS has been staring at the sky since 2018, watching stars for tiny dips in brightness that can mean a planet crossed in front. The hard part is that most of those dips are weak, messy, or buried in instrumental noise. What changed this month is that a team behind the T16 Planet Hunt pushed a machine-learning-assisted search through TESS’s first full-frame survey and came back with 10,091 new planet candidates in one shot. ### What did they actually find? They searched TESS Cycle 1 full-frame images and reported 11,554 candidate planets total. Of those, 10,091 are new candidates, 1,052 were already known TESS candidates, and 411 are single-transit events where the team could not yet pin down full orbital parameters. The paper says those candidates orbit stars as faint as TESS magnitude 16, with periods between 0.5 and 27 days. ### Why is that a big deal? (arxiv.org) Because TESS usually shines brightest on bright stars, and that leaves a lot of fainter stars underworked. The T16 project built a uniformly detrended set of 83,717,159 light curves from those fainter targets, then searched them at scale. The result, the team says, more than doubles the number of known TESS exoplanet candidates. That is not “10,000 new confirmed planets.” It is 10,000-plus leads that now need vetting and follow-up. ### Where does the AI part come in? The key trick is not a robot astronomer doing everything alone. It is a pipeline that cleans the light curves, searches for transit-like dips, and uses machine-learning-assisted triage to sort the promising signals from the junk. NASA is pushing a similar idea with ExoMiner++, an open-source model adapted from Kepler to TESS data. On an initial run, ExoMiner++ flagged 7,000 TESS targets as exoplanet candidates. Basically, the field is moving toward automation because humans cannot manually inspect this volume forever. (arxiv.org) ### Did they prove the pipeline works? Yes — at least enough to show it is finding real planets, not just artifacts. The T16 team followed up one candidate host, TIC 183374187, with radial-velocity measurements from Magellan/PFS and confirmed a newly identified hot Jupiter. One confirmed planet does not validate all 10,091 candidates, but it does show the pipeline can surface genuine worlds that earlier searches missed. ### Is TOI-715 b part of this news? (arxiv.org) No. TOI-715 b is real, but it is older news. NASA lists it as a super-Earth announced in 2023, about 1.55 Earth radii and 3.02 Earth masses, orbiting an M dwarf every 19.3 days. So if you saw posts tying this week’s “10,000 candidates” story to TOI-715 b, that is mixing two separate exoplanet threads together. ### So are we about to get 10,000 new worlds? Not quickly. (arxiv.org) Candidate is the operative word. Each signal still has to survive false-positive checks — eclipsing binaries, instrumental systematics, blended stars, all the usual troublemakers. Then the most interesting targets need follow-up from ground telescopes and, for a select few, bigger observatories that can study atmospheres or masses. (science.nasa.gov) ### Why does this matter beyond the headline? Because the bottleneck in exoplanets has shifted. Finding possible planets used to be the scarce step. Now the scarce step is confirming and characterizing them. A catalog this large gives astronomers a much richer menu of short-period worlds around faint stars — especially the kinds of targets that can feed future mass measurements, demographic studies, and atmospheric follow-up. ### Bottom line? (science.nasa.gov) The real story is not that AI “discovered 10,000 planets.” It is that automated searches are getting good enough to turn TESS’s giant archive into a much bigger candidate factory — and that changes what exoplanet science works on next. (arxiv.org)