AI finds 100+ hidden exoplanets

Exoplanet hunting just got a lot more automated. A team led by the University of Warwick ran a new AI pipeline called RAVEN through archival data from NASA’s TESS mission and came back with 118 statistically validated planets, including 31 that had not been detected before. (arxiv.org) The bigger point is not just the 31 new worlds. It’s that the software also surfaced more than 2,000 vetted candidates still sitting in the data, which means the backlog may be far larger than humans can sort by hand. ### What is TESS actually looking at? TESS — NASA’s Transiting Exoplanet Survey Satellite — watches huge patches of sky and looks for tiny dips in starlight when a planet crosses in front of its star. (science.nasa.gov) It has monitored more than 2 million stars, which is great for discovery but terrible for manual triage, because most interesting-looking signals are not planets at all. They can be noise, binary stars, or instrumental junk. ### Why is this such a hard sorting problem? The hard part is not finding dimming events. It’s deciding which dimming events are real planets. (arxiv.org) TESS produces enormous numbers of light curves, especially from full-frame images, and the planet sample is incomplete and biased toward short-period worlds. That makes demographic studies messy and leaves a lot of plausible candidates stuck in limbo. ### So what does RAVEN do? RAVEN is a vetting and validation pipeline. First it searches for transit-like signals with a box least squares method. Then it uses machine-learning classifiers trained on simulated planets and multiple false-positive scenarios to rank how likely each signal is to be a real planet. (earth.com) After that, it applies statistical validation rather than requiring follow-up observations for every case. Basically, it acts like a very fast filter for a giant inbox. ### What did the team actually find? The headline result is 118 newly validated planets in the sample RAVEN analyzed, with 31 of those being newly detected in this work rather than previously flagged candidates. (arxiv.org) The paper also says the system recovered 875 TOIs and 37 CTOIs and assembled a vetted candidate sample of more than 2,000 objects for future work. That is why some headlines say “100+ planets” while others talk about “thousands more.” Both are pointing at different parts of the result. ### Are these all Earth-like worlds? No — and that matters. The method is strongest for short-period transiting planets, meaning worlds that orbit close to their stars and cross our line of sight often enough for TESS to catch them. (arxiv.org) Some of the validated planets are described as rare or extreme, but this is not a sudden haul of Earth 2.0 candidates. It is more about scale and completeness than about one especially habitable world. ### Why does archival data matter so much? Because telescope time is scarce, but stored data keeps getting more valuable. NASA already showed this pattern with ExoMiner, an earlier AI system from Ames that validated 370 planets from Kepler data. (arxiv.org) Now ExoMiner++ is being adapted for both Kepler and TESS. RAVEN fits the same trend — better models can keep mining old observations long after the spacecraft has moved on. ### Does this change the big exoplanet picture? It helps fill in the census. NASA’s running total now sits above 6,200 confirmed exoplanets, but that count is still shaped by what is easiest to detect. (arxiv.org) Tools like RAVEN can reduce the backlog and make the sample less skewed, which is what astronomers need if they want to answer population questions instead of just adding one planet at a time. ### Bottom line The real news is not that AI found a few hidden planets. It’s that exoplanet discovery is starting to look like a data-sorting problem as much as a telescope problem — and that means there may be a lot more worlds already sitting in the archive, waiting for better software. (science.nasa.gov) (arxiv.org) (science.nasa.gov)