Machine learning confirms 118 exoplanets

Exoplanet hunting has a data problem. NASA’s TESS telescope watches millions of stars for tiny dips in brightness, and every dip might be a planet — or just noise, a binary star, or some other astrophysical fake-out. That bottleneck is why this result matters. A University of Warwick team used a machine-learning pipeline called RAVEN to comb through TESS data and validate 118 exoplanets, including 31 that had not been detected before. (warwick.ac.uk) ### What actually changed? (warwick.ac.uk) The new thing is not a telescope launch or a one-off planet announcement. It’s a workflow. RAVEN ran through TESS full-frame image light curves, searched for repeating transit-like signals, classified likely planets against several false-positive scenarios, and then statistically validated the strongest cases. (science.nasa.gov) That process yielded 118 newly validated planets and a much larger backlog of strong candidates for follow-up. ### What is RAVEN doing differently? Basically, it is not just saying “this dip looks planet-ish.” RAVEN combines a standard transit search with machine-learning classifiers and then a Bayesian validation step. The models were trained on simulated planets and multiple astrophysical impostors, so the system can score whether a signal is more likely to be a real transiting planet or something like an eclipsing binary. (warwick.ac.uk) ### Why does TESS need this? TESS is built for scale. It surveys huge patches of sky and tracks brightness changes for enormous numbers of stars, which is great for discovery but brutal for manual checking. NASA has already framed AI vetting as a necessary next step for exoplanet science, because the missions are now producing more candidates than small teams can inspect one by one. (arxiv.org) ### How big is this haul? The headline number is 118 validated planets, but the bigger number may be the candidate pile behind them. The Warwick team says RAVEN also identified more than 2,000 high-quality planet candidates, with nearly 1,000 entirely new. So this is less like finding one treasure chest and more like mapping a whole new mine. (arxiv.org) ### What kinds of planets showed up? This search focused on short-period transiting planets around Sun-like stars — roughly worlds with orbital periods from 0.5 to 16 days in the press summary, and short-period systems more broadly in the paper. That means many of these planets are close-in worlds, the kind TESS is especially good at catching because they transit often and produce repeated signals quickly. (science.nasa.gov) ### Why is “validated” different from “discovered”? That distinction matters. Some of the 118 were already known as candidates, but had not cleared the statistical bar to count as confirmed planets. Validation means the odds now strongly favor the planet explanation over the false-positive alternatives. (warwick.ac.uk) In this batch, 31 were newly detected by the pipeline itself, while the rest were upgraded from candidate status. ### Does this mean astronomers are handing discovery to AI? Not really. The catch is that machine learning is becoming the front end, not the whole scientific process. Humans still decide how to train the models, what false positives to simulate, which candidates deserve telescope time, and how to interpret the resulting planet population. (warwick.ac.uk) But the screening step is now too large and too repetitive to do efficiently by hand. ### So why should anyone outside astronomy care? Because this is what modern science looks like when the instrument gets too productive for the old workflow. TESS is not short on signals. It is short on time, attention, and follow-up capacity. Tools like RAVEN turn giant archives into usable planet catalogs — and that is what makes the next atmospheric study, formation model, or weird outlier planet possible. (warwick.ac.uk) (science.nasa.gov) (arxiv.org)