Kepler-90 i

What checks led researchers to recognize Kepler-90 i as a planet?

Researchers searched Kepler data again for new transit signals around 670 stars that already had known planets, then kept 513 signals that passed their basic filters. A neural network gave the signal from i a score of 0.942, placing it among the first candidates to examine. The researchers then checked which star produced the signal, whether an instrument error could explain it, and whether two orbiting stars could mimic a planet. They first calculated a 0.5% chance of falsely calling the signal a planet. After accounting for the fact that seven planets were already known around the star, that false-positive probability fell below 0.01%, and i was statistically validated as a planet.

• 670 stars Stars searched again for new transit signals
• 513 signals Signals that passed the basic filters and were ranked by the neural network
• 0.942 The neural network's score for candidate i
• Below 0.01% Final probability that i was falsely classified as a planet

How 513 candidate signals were found among 670 stars

Four years of Kepler starlight data contained the transits of the seven known planets, instrument noise, and changes in the star's own brightness. The researchers first removed the known transits, then applied a separate search to 670 multiplanet systems to find brightness dips repeating at regular intervals. They rejected signals with fewer than three complete transits, physically implausible durations, and other basic problems. Even after those filters, 513 “signals for review” remained.

Not every signal left at this stage was a planet. Small spacecraft errors, changes in a star, and eclipses between two orbiting stars can all make similar repeating patterns. The neural network ranked the 513 signals so researchers could inspect the ones most similar in shape to a planetary transit first. Thirty signals scored above 0.5, nine scored above 0.8, and the Kepler-90 signal repeating every 14.449 days received 0.942.

Artificial intelligence decided which of 513 signals to examine first. Researchers decided whether it was a planet by checking its location, instrument errors, and possible binary-star explanations.

Why is a score of 0.942 not the probability that it is a planet?

The neural network was trained on about 15,000 Kepler signals that people had already labeled as planetary or false detections. It received both the brightness change across a full orbit and a close-up view of the transit. On 1,523 test signals that were not used for training, it classified about 96% correctly. Its AUC—a measure of how well it ranks planetary signals above false detections—was 0.988.

An AUC of 0.988 means that, when one planetary signal and one false detection were chosen from the test data, the model ranked the planetary signal higher 98.8% of the time. It does not mean that Kepler-90 i itself had a 98.8% chance of being a planet. The score of 0.942 also ranked candidates within this neural network; it was not a final planet probability that included every possible source of error.

The data given to the neural network did not include pixel-location information showing where in the image the signal originated. The model also sometimes gave high scores to transits around background stars and to some binary-star signals. Separate validation was therefore still necessary after a high score.

Could the signal have come from another star or the instrument?

The researchers first inspected the brightness dip repeating every 14.449 days. They checked whether the amount of blocked starlight changed between observing seasons and whether the signal repeated at the same time as other instrument effects. They also examined which pixels became dimmer during the transit. If the apparent position shifted, the signal might have come from a nearby star blended with Kepler-90 rather than from Kepler-90 itself.

High-resolution images from ground-based telescopes were used to look for a bright companion hidden inside the point that Kepler saw as one star. The researchers also looked for a shallower secondary eclipse between the main dips, which would support a binary-star explanation. The signal remained after instrument effects were corrected in different ways, so they concluded that the repetition came from a real astronomical source rather than a spacecraft error.

1. Search Remove the known planets' signals, search 670 stars for repeating dips, and keep 513 signals.
2. Ranking Compare the full-orbit and close-up transit shapes; the neural network gives candidate i a score of 0.942.
3. Additional checks Examine pixel location, instrument errors, a possible secondary eclipse, and high-resolution images.
4. Statistical validation Model possible binary-star explanations and the known multiplanet system, reducing the false-positive probability below 0.01%.
5. Planet properties Use the validated transits to find a 14.449-day orbit and a radius 1.32 times Earth's, while leaving mass and interior composition unknown.

Why the false-positive probability fell from 0.5% to below 0.01%

Using a validation tool called vespa, the researchers compared the transit shape with several explanations: a real planet, two stars orbiting each other around Kepler-90, or two background stars orbiting each other. The calculation included the star's properties, the range of companion-star brightness that high-resolution images would have missed, and the absence of a detectable secondary eclipse. It found a probability of about 0.005, or 0.5%, that a nonplanetary object made the signal. This was already below the 1% threshold often used to statistically validate Kepler planets.

The researchers then added the fact that seven planets were already seen crossing the same star. When a star has several planet-like signals, one additional signal is less likely to be a random false detection than a signal found alone. With this “multiplicity” information, the false-positive probability fell below 0.0001, or 0.01%, and i was validated as a planet. This was not a gravitational mass measurement or an independent transit seen by another telescope. The known alternative explanations had instead been made unlikely enough for statistical validation.

How much can a transit signal tell us on its own?

The transits of i repeat every 14.44912±0.00020 days and last about 2.80±0.31 hours. From the amount of starlight blocked and the radius of the host star, researchers calculated a radius 1.32±0.21 times Earth's. If i followed a circular orbit in the same plane as the other planets, its transit was expected to last about five hours, but the observed transit was shorter. The researchers suggested that i's orbit may be slightly more tilted than those of the other seven planets, making it cross closer to the star's edge.

A small radius makes a rock-rich planet more plausible, but it does not measure mass, average density, or the presence of an atmosphere. A NASA public planet card currently lists about 2.3 Earth masses, but the discovery paper did not measure a mass, and the default research row in the Exoplanet Archive leaves the mass blank. The value 2.3 must therefore not be treated as an independent gravitational measurement or used to establish the planet's composition.

The discovery study's temperature of 709±75 K, about 436±75 degrees Celsius, is also a calculated value. It assumes that heat spreads evenly around the planet and that the planet reflects between 0% and 70% of the starlight. It is not an observed surface or atmospheric temperature, and no clouds were detected. i received the last letter because it was the eighth planet discovered, but it is the third planet from the star after b and c.

The colors and surface shown on this page are an illustration, not an observed photograph. Kepler could not resolve the round disk of i. What has been measured is the timing and duration of its repeating transits, the amount of starlight blocked, and the results of the false-positive analysis.

AI set the review order; researchers determined whether it was a planet

The transit signal of Kepler-90 i was not newly observed in 2017. It was already present in the existing Kepler data. What was new was the search method used to recover the weak repeating signal, the neural network used to rank 513 candidates, and the validation process that checked the signal's location, binary-star explanations, and instrument errors. No single step was enough to identify i as a planet.

In this case, artificial intelligence did not replace an astronomer or make the final decision. It decided which signals people should inspect first in a data set too large to examine every candidate in equal detail. To reproduce the result, researchers must record how the search produced the candidates, what the neural-network score means, and which separate evidence was used for the final decision.

Sources

• The Astronomical Journal — Deep-Learning Search and Statistical Validation of Kepler-90 i
• NASA — Artificial Intelligence and the Eighth Kepler-90 Planet
• NASA Exoplanet Archive — Kepler-90 i Published Parameters
• NASA Exoplanet Archive — Planet Parameter and Mass-Provenance Definitions
• The Astrophysical Journal — Vespa False-Positive Probabilities and Statistical Validation
• Publications of the Astronomical Society of the Pacific — Kepler Data-Validation Tests
• NASA Science — Kepler Mission and Transit Data