NASA's AI Found 12 Potentially Habitable Exoplanets in Old Kepler Data

A neural network re-analyzed archived Kepler Space Telescope data and identified 12 previously overlooked exoplanets in the habitable zone of their stars.

A neural network developed at NASA's Jet Propulsion Laboratory has uncovered 12 new potentially habitable exoplanet candidates hidden in archived data from the Kepler Space Telescope, including three rocky, Earth-sized worlds that human analysts had previously missed. The discovery represents both a significant addition to the catalog of potentially habitable worlds and a compelling demonstration of AI's ability to extract new science from existing data. The system, called ExoMiner-3, is the third iteration of JPL's exoplanet detection neural network.

It was trained on the complete catalog of confirmed Kepler exoplanets to recognize transit signals — the minute, periodic dips in starlight that occur when a planet passes between its host star and our line of sight. Previous versions of ExoMiner had already proven effective at confirming planet candidates, but ExoMiner-3 was specifically designed to probe deeper into the noise floor of Kepler's photometric data. The central technical achievement was detecting transit signals that fell below the signal-to-noise thresholds used in traditional detection pipelines.

Conventional analysis applies conservative thresholds to minimize false positives, which inevitably means some genuine but weak signals are discarded. ExoMiner-3 learned to distinguish authentic planetary transits from false positives caused by eclipsing binary stars, stellar variability, systematic instrumental effects, and other sources of confusion — even when those genuine signals were buried deep in noise.

Of the 12 new habitable-zone candidates, the three rocky, Earth-sized planets have generated the most scientific excitement.

Two of these orbit K-type stars, which many astrobiologists consider the most favorable stellar hosts for life. K-type stars are dimmer and cooler than our Sun but significantly more stable, with lifespans ranging from 15 to 45 billion years — providing far longer windows for biological evolution than the

Sun's approximately 10-billion-year lifespan. They also produce fewer intense flares than M-type red dwarfs, reducing the radiation threat to planetary atmospheres.

The third rocky candidate orbits an M-type red dwarf star. While M-dwarfs are the most common stars in the galaxy, their habitable zones are much closer in, often resulting in tidal locking. The high flare activity of red dwarfs also poses challenges for atmospheric retention.

This candidate is considered less promising but still warrants investigation. NASA has prioritized follow-up observations using the James Webb Space Telescope. Transit spectroscopy — analyzing starlight as it filters through a planet's atmosphere during transit — could reveal the presence of biosignature gases including oxygen, methane, water vapor, and carbon dioxide.

While characterizing the atmospheres of small, distant rocky planets pushes JWST to its operational limits, the K-type star candidates are considered feasible targets due to the favorable star-to-planet size ratio. The discovery carries implications beyond the specific planets found. The Kepler mission collected photometric data on over 150,000 stars during its primary mission and extended K2 mission.

If 12 habitable-zone candidates were hiding below conventional detection thresholds in this data, the total number of overlooked planets across all orbital configurations could be substantially larger. NASA has confirmed plans to apply ExoMiner-3 to data from the Transiting Exoplanet Survey Satellite (TESS), which has surveyed a far larger portion of the sky. More broadly, this result exemplifies a growing pattern in scientific research where AI systems find meaningful discoveries in archival data sets.

Particle physics, genomics, materials science, and climate research have all seen similar examples of machine learning models identifying signals that conventional analysis missed. The implication is that the scientific community may be sitting on a wealth of undiscovered findings in data that has already been collected and cataloged — waiting only for the right analytical tools to extract them.

---

NASA's AI Found 12 Potentially Habitable Exoplanets in Old Kepler Data

Related Reading