Faith Hawthorn had a problem most astronomers would envy. NASA’s Transiting Exoplanet Survey Satellite, known as TESS, had spent four years photographing nearly the entire sky, generating light curves for millions of stars. Somewhere in that mountain of data, undiscovered planets were hiding in plain sight. The trouble was that no human team could vet all those signals fast enough to keep up.
So Hawthorn, a postdoctoral researcher at the University of Warwick, and her collaborators built an automated pipeline to do it for them. They call it RAVEN, short for RAnking and Validation of ExoplaNets. In a study accepted for publication in Monthly Notices of the Royal Astronomical Society, the team reports that RAVEN scanned more than 2.2 million stars in the TESS archive and flagged over 2,000 strong planet candidates. Of those, 118 worlds passed rigorous statistical tests that classify them as validated planets, and roughly 1,000 of the candidates had never appeared in any previous TESS catalog.
The result is one of the largest single-pass automated planet searches ever applied to TESS full-frame images, and it signals a broader shift in how new worlds get found: not one light curve at a time by a human reviewer, but millions at once by a machine trained to separate planets from imposters.
How RAVEN hunts for planets
RAVEN processed TESS-SPOC full-frame images from sectors 1 through 55, covering roughly the satellite’s first four years of operations. Its target list drew from a magnitude-limited sample of more than 2.2 million main-sequence stars selected using data from the European Space Agency’s Gaia observatory. The search window covered orbital periods between 0.5 and 16 days, a range tuned to catch close-in planets whose repeated transits produce the clearest, most detectable dips in starlight.
The pipeline works in three stages. First, a box-least-squares algorithm combs each star’s brightness record for periodic dimming patterns consistent with a planet crossing the stellar disk. Next, a machine-learning classifier sorts those signals, weeding out instrumental noise and astrophysical false alarms like eclipsing binary stars. The Warwick team trained gradient-boosted decision trees and Gaussian process classifiers on synthetic datasets laced with simulated planet transits and known contaminants, teaching the system to recognize subtle differences between real planetary signals and lookalikes. Finally, a Bayesian statistical validation step calculates the probability that each surviving candidate is a genuine planet rather than a false positive.
That three-layer filter is what allowed RAVEN to surface signals that earlier searches missed. Some candidates fell below older detection thresholds; others had light curves muddied by systematic noise that the new classifier handles more effectively. The roughly 1,000 entirely new candidates represent planets that were, in effect, already recorded in the TESS data but invisible to previous methods.
RAVEN is not working alone
RAVEN joins a growing roster of AI systems mining TESS observations. NASA’s own ExoMiner++ model, trained on both Kepler and TESS data, previously identified 370 exoplanets and flagged thousands of additional candidates in an initial TESS run. Those results are described in NASA’s open-science materials, which detail how deep-learning architectures can replicate the decisions of human vetters at scale.
The two systems differ in architecture and training data, and no head-to-head comparison exists in the published literature. Their results are best understood as complementary: each pipeline catches signals the other might miss, and the fact that both exist reflects just how large the vetting bottleneck has become. TESS generates far more transit-like signals than any team of human reviewers can process, and automated pipelines are now the only realistic way to close that gap.
What “validated” does and does not mean
An important distinction separates RAVEN’s 118 validated planets from fully confirmed discoveries. Statistical validation means each candidate passed probability thresholds designed to rule out common false-positive scenarios, such as background eclipsing binaries blended into the target star’s light. It does not mean these objects have undergone independent ground-based follow-up like radial-velocity measurements or high-resolution imaging, which would nail down their masses and rule out remaining ambiguities.
Until that follow-up happens, some fraction of the 118 could still turn out to be grazing binaries, background blends, or other astrophysical mimics that slipped through the statistical filters. As of June 2026, no public reporting confirms that NASA’s Exoplanet Archive has ingested any of RAVEN’s validated planets as officially confirmed worlds. The broader workflow for TESS discoveries, summarized on the mission overview page, describes how the satellite’s data products feed into candidate and confirmed exoplanet lists, but the archive team has not publicly addressed RAVEN’s outputs specifically.
That gap matters for researchers who rely on the canonical planet count for occurrence-rate calculations. RAVEN’s numbers should be treated as the team’s own assessment, grounded in a defined statistical model and peer-reviewed methodology, but still awaiting independent replication and catalog integration.
The limits of a short-period search
RAVEN’s 0.5-to-16-day orbital period window introduces a strong selection effect that shapes everything the pipeline can find. Planets in this range orbit scorchingly close to their host stars, completing a full year in less time than it takes Earth to rotate 16 times. That means RAVEN is sensitive to hot Jupiters, ultra-short-period rocky worlds, and close-in sub-Neptunes, but it cannot detect longer-period planets, including those orbiting in or near the habitable zones of Sun-like stars.
Any projection about how many total planets RAVEN might eventually find in extended TESS sectors depends on assumptions about the occurrence rate of short-period planets around the specific stellar populations in the sample, and those rates carry their own measurement uncertainties. The pipeline’s strength is depth within a defined niche, not breadth across all planetary architectures.
There are also questions about how robustly the machine-learning classifier will perform beyond its training regime. The synthetic injections and labeled contaminants used to teach the model are designed to mimic real TESS data, but they cannot capture every instrumental quirk or exotic astrophysical configuration. Unusual systems, such as multi-planet resonant chains, highly eccentric orbits, or stars with intense magnetic activity, may produce light curves that sit at the edges of the classifier’s experience, raising the risk of both missed planets and misclassified false positives.
What comes next for RAVEN’s candidates
The most consequential next steps are observational, not computational. Ground-based telescopes equipped with precision spectrographs can measure the gravitational wobble that RAVEN’s candidate planets induce on their host stars, pinning down planetary masses and confirming their existence beyond statistical probability. High-resolution imaging can check whether faint background stars are contaminating the TESS aperture and mimicking transit signals.
Equally important is transparency. The degree to which other research groups can independently reproduce RAVEN’s candidate list will depend on public access to the pipeline’s code, training sets, and intermediate data products. Peer review of the methodology paper is a strong first filter, but stress-testing by outside teams using the same tools on the same data is the gold standard for trust in automated discovery.
Cross-matching RAVEN’s candidates against ExoMiner++ and other vetting pipelines could also sharpen confidence in the strongest detections. Where multiple independent systems flag the same signal, the case for a real planet grows considerably stronger.
If those pieces fall into place, RAVEN and systems like it could transform the overwhelming flood of TESS photometry into a more complete census of nearby planetary systems. The 118 validated planets are a down payment. The real payoff will come when follow-up observations turn statistical validations into confirmed worlds with measured masses, known compositions, and a place in the broader story of how planetary systems form and evolve.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
