Back to Bigger Picture

BIGGER PICTURE Scientific discovery

301 planets were waiting in an archive

Kepler had already recorded their faint signals. ExoMiner helped astronomers distinguish genuine planets from convincing imposters, extracting new discoveries from observations already paid for.

The covered Kepler spacecraft rests on a wheeled support in a clean room, with a technician beside it.Original photograph
Kepler, still under its protective covering, at the Astrotech processing facility in January 2009. Years later, ExoMiner analysed archived observations from this telescope to validate 301 additional planets.NASA / Chris RhodesNASA media-use guidelines
Newly validated in 2021
301 planets
Test-set recall
93.6% at 99% precision
Best comparison classifier
76.3% recall at the same precision

How the idea works

  1. Measure a stellar dip

    Kepler light curves flag candidate transits.

  2. Compare diagnostic evidence

    Expert-inspired network branches distinguish planets from imposters.

  3. Validate candidates statistically

    Strict thresholds turn supported candidates into catalogue entries.

A tiny dip can have several explanations

A planet passing in front of its star briefly reduces the light reaching a telescope. But not every repeated dip is a planet. Other astronomical configurations and false signals can imitate a transit. Kepler’s archive therefore contained candidates that had been detected but still needed convincing statistical validation before joining the planetary census.

Sources: [1]

A network built around astronomers’ checks

ExoMiner learns from previously confirmed planets and known false positives. Its separate diagnostic components imitate the tests experts use when examining a candidate, rather than producing an unexplained score from a single picture. The researchers also developed a way to show which evidence influenced a classification, helping astronomers inspect the machine’s reasoning.

Sources: [2]

More planets at the same strict threshold

In the published test, ExoMiner recovered 93.6% of the planets at a fixed precision of 99%; the strongest comparison classifier recovered 76.3%. These are recall figures, not overall accuracy. Applying the method to previously unvalidated Kepler candidates allowed the team to validate 301 additional exoplanets, announced by NASA in November 2021.

Sources: [2][1]

A discovery engine for existing observations

This was statistical validation, not 301 newly photographed worlds, and none was considered Earth-like or within its star’s habitable zone in the announcement. Its value is foundational: a better catalogue helps researchers understand how planetary systems form and how unusual our own may be. Better analysis can make an existing telescope archive scientifically productive again.

Sources: [1]

The original sources

  1. NASA JPLNew Deep Learning Method Adds 301 Planets to Kepler’s Total Count
  2. Study authors / Astrophysical Journal manuscriptExoMiner: A Highly Accurate and Explainable Deep Learning Classifier that Validates 301 New Exoplanets

What to explore next

A protein’s shape became something any lab could look upSweden tested AI in breast screening—and found 29% more cancers
Explore all 99 applications