NASA’s AI planet hunter has helped validate 370 worlds beyond our Solar System

Hundreds of planets were waiting inside observations that astronomers had already collected. NASA’s ExoMiner helped validate 301 of them in a result announced in November 2021. A second study added 69 in 2023, bringing those two sets of discoveries to 370 worlds beyond our Solar System.[1][3]

The achievement came from revisiting signals recorded by the Kepler space telescope. Kepler’s processing pipeline had detected the original 301 signals, and scientists had already designated them as candidates. Machine learning supplied a further step: enough evidence to validate them statistically as planets.[1]

Reading a planet in a dip of starlight

A planet passing across its star blocks a small amount of light. Repeated dips provide a timetable for its orbit; their depth helps determine the planet’s size relative to the star. Astronomers call the changing brightness a light curve. A useful discovery can therefore begin with a pattern in measurements rather than a photograph of a distant world.[6]

ExoMiner learns from examples of established planets and misleading signals. Its architecture divides the evidence into diagnostic components inspired by the checks astronomers make. The original research also introduced a way to show experts which inputs influenced a classification. This makes the system useful for examining a candidate as well as ranking it.[2]

On the original paper’s test set, at a precision of 99%, ExoMiner recovered 93.6% of the planets, compared with 76.3% for the best existing classifier in that comparison. Precision describes the share of selected signals that are right; recovery describes how many of the available planets are found. Holding the first measure fixed makes the improvement in the second meaningful.[2]

A discovery can depend on its neighbours

The 2023 advance added information about the planetary system itself. A signal around a star with other convincing planetary detections carries different evidence from an isolated dip. The researchers combined that configuration information with ExoMiner’s score using a method called a multiplicity boost. It enabled the additional 69 validations in systems with multiple Kepler Objects of Interest.[3]

Opening up a larger part of the sky

The TESS version combines training data from Kepler with observations from its successor mission. It also examines additional diagnostics, including changes in brightness over time and images that help locate the source of a dimming event. In the 2025 analysis, the selected candidate records amounted to about 5% of the unlabelled input records, calculated from the figures in the table. That creates a much more focused queue for follow-up investigation.[4]

The January 2026 preprint takes that approach into TESS’s full-frame images. Those images support measurements of millions of stars beyond the smaller set selected for two-minute observations. The new model receives a map of known neighbouring stars and their relative brightness. That context helped reject some misleading signals, although the overall comparison recovered fewer real-planet signals. The added information changes how the system weighs a candidate.[5]

Observations that keep producing discoveries

NASA publishes ExoMiner’s code, and its current repository includes a pipeline that takes TESS catalogue identifiers through to model scores. The underlying mission observations and catalogue products are public too. Other researchers can examine candidates, rerun the processing and adapt the software to their own investigations.[7]

Finding a planet also creates questions that other measurements can answer. A star’s gravitational wobble provides information about a planet’s mass; starlight passing through an atmosphere can reveal molecular signatures. A stronger catalogue gives those investigations more targets to choose from.[6]

This is an especially productive role for AI in astronomy: making expensive, irreplaceable observations yield more discoveries over time. ExoMiner’s progression from Kepler validations to a wider TESS search shows how better analysis can expand the scientific return from a telescope’s existing work.[1][7][5]

Report an error or suggest a correction