astronomy/ai produced the result/The Astronomical Journal 2026 · v2
Kepler planets sorted by host star's Galactic orbit show differing eccentricities
Researchers measured how stretched the orbits of 2465 Kepler planets and candidates are, then compared planets around stars belonging to two populations of the Milky Way's disk. A clustering model assigned each host star to one population or the other.
spectrum · one line per step, placed by what the step does · bright lines used AI
The Orbital Eccentricities of Planets in the Kinematic Thin and Thick Galactic Disks
The Astronomical Journal, 2026
doi:10.3847/1538-3881/ae71bf · record aix-00034 v2 · checked 2026-10-08
- AI was for
- Classification
- Model family
- Clustering
- Checked by
- None stated
- Code
- not reported
The finding the paper is about came from the AI.
What this research was about
Planets do not all travel on neat circles. An orbit can be stretched into an oval, and astronomers measure that stretch with a number called eccentricity: zero is a perfect circle, and larger values mean a more elongated path. Eccentricity carries a memory of a planetary system's past, because gravitational scattering and close encounters between young planets tend to leave orbits more stretched. Measuring it is awkward, though. Most of these planets were found by the Kepler telescope, which watched stars for tiny dips in brightness as a planet crossed in front. A single dip says little about the shape of an orbit on its own.
The Milky Way's disk is usually described as two overlapping populations. A thin disk of stars on fairly circular, well-behaved paths, and a thick disk of older stars moving on more unruly orbits and carrying a different mix of chemical elements. The researchers set out to ask whether planets around thick-disk stars have differently shaped orbits from planets around thin-disk stars. To do that they needed two things: an eccentricity estimate for each planet, and a disk label for each host star.
Where AI came in
The machine learning sat at the labelling step. Chemical abundances, which would identify a thick-disk star directly, are not available for most Kepler host stars. So the team built a smaller calibration set of stars that do have measured magnesium and iron abundances, split them chemically into thick and thin groups, and then fitted a two-component Gaussian mixture model to how fast those stars move through the Galaxy. A mixture model is an unsupervised clustering method: it is given no labels and works out how to describe a cloud of data points as a blend of two overlapping groups.
Applied to the full Kepler sample, that fitted model returned a thick-disk probability for every star, and stars above a probability of 0.5 were called thick disk, giving 1515 thin-disk and 378 thick-disk planet hosts. The model stood in for a chemical measurement nobody had made for most of these stars, using their motions instead. Everything after that used conventional statistics rather than learned models: the transit light curves were fitted to extract eccentricities, and the population-level distributions were inferred with hierarchical Bayesian methods.
Written by AIxSci from the checked record below, to give context for readers outside the field. It is not part of the record.
The work
Technical · from the record
The authors constrained orbital eccentricities for 2465 Kepler planets and candidates orbiting 1888 stars by fitting transit light curves with stellar density priors, using the photoeccentric effect. Each host was assigned a kinematic thin- or thick-disk probability by a two-component Gaussian Mixture Model fitted to Galactocentric cylindrical velocities of a Kepler–APOGEE calibration sample that had been split chemically on [Mg/Fe] and [Fe/H]. Population-level eccentricity distributions were then fitted hierarchically for singles and multis in each disk. Single-transiting planets around kinematic thick disk hosts were found to follow a distinct distribution with higher eccentricities than thin disk singles, while multi-transit systems in the two disks were consistent with one another.
How AI was used
Machine learning entered at the host-star classification step. Because most Kepler stars lack homogeneous chemical abundances, the authors built a calibration sample by crossmatching Kepler planet hosts with APOGEE DR17 ASPCAP abundances, cut it to main-sequence stars, and divided it into chemical thick- and thin-disk subsamples with a linear [Mg/Fe]–[Fe/H] boundary. They then fitted a two-component Gaussian Mixture Model to these stars in Galactocentric cylindrical velocity space, using Gaia DR3 velocities where radial velocities exist and published inferred velocities otherwise, and applied the trained mixture model to the Kepler sample to obtain a thick-disk probability for every star. Stars were then labelled thin or thick disk by thresholding that probability at 0.5. All downstream steps are non-learned: transit detrending and fitting with the ALDERAAN package and dynesty nested sampling, importance sampling to extract eccentricity and longitude of periastron posteriors, and hierarchical Bayesian fitting of Rayleigh, Beta, monotonic Beta and half-Gaussian population distributions with numpyro.
The shape of the work
Structural · the record, drawn
no AI
Select Kepler planet and candidate sample
Cleaning, filtering, normalising or labelling data already obtained.
we first remove any planets marked as False Positives according to the NASA Exoplanet Archivewhere the paper describes this · verbatim
no AI
Build chemically labelled calibration sample
Cleaning, filtering, normalising or labelling data already obtained.
We apply stellar parameter cuts to ensure only main-sequence stars are included in the calibration samplewhere the paper describes this · verbatim
AI
Fit mixture model to calibration velocities
Fitting model parameters, including fine-tuning an existing model. The AI stood in for conventional algorithm.
fit a two-component Gaussian Mixture Model (GMM) to the samplewhere the paper describes this · verbatim
AI
Score thick disk probability for Kepler stars
Running a trained model over new data to predict, classify or score. The AI stood in for unresolved measurement.
We apply the trained GMM to the entire Kepler samplewhere the paper describes this · verbatim
no AI
Assign hosts to kinematic thin or thick disk
Cleaning, filtering, normalising or labelling data already obtained.
In total, we have in our sample 1515 thin disk planet hosts and 378 thick disk planet hosts.where the paper describes this · verbatim
no AI
Fit transit light curves for eccentricity posteriors
Running a trained model over new data to predict, classify or score.
We detrend and fit the Kepler transit light curves in a homogeneous manner using the ALDERAAN transit fitting packagewhere the paper describes this · verbatim
no AI
Fit population eccentricity distributions
Running a trained model over new data to predict, classify or score.
we infer the underlying population-level eccentricity distribution within a Bayesian hierarchical frameworkwhere the paper describes this · verbatim
no AI
Model comparison and leave-N-out robustness check
Testing outputs against ground truth.
we also perform a leave N out validation testwhere the paper describes this · verbatim
What the record says
Technical · every part carries its own basis
+ in the paper~ our reading− not reported
How to read the quotations. A quotation shows where the paper describes something. It does not quote every value beside it: one passage locates a part of the work, and values without their own quotation are our reading of that passage.
A fitted two-component Gaussian Mixture Model supplies the thin/thick kinematic disk label for every host star, and the central comparison is between those two model-defined populations. The eccentricity posteriors and the population-level distribution fits themselves use non-learned methods (nested-sampling transit fitting, hierarchical Bayesian inference), so 'instrument' reflects dependence of the claim on the learned classifier rather than AI producing the eccentricities.
fit a two-component Gaussian Mixture Model (GMM) to the samplewhere the paper describes this · verbatim
What this paper did not report
Technical · absence is published deliberately
- CodeWhether the code is available is not stated.
- Trained model weightsWhether the trained model is available is not stated.
- DataWhether the data are available is not stated.
- ComputeThe hardware or time used is not stated.
- ValidationNo validation of the AI is described.
- Version of two-component Gaussian Mixture Model on Galactocentric cylindrical velocitiesWhich version of the model was used is not stated.
About this article
Record aix-00034, version 2, checked by a person on 2026-10-08. The record describes the paper; it does not assess whether the paper's findings are right. The paper is published under CC-BY; quotations are at most 25 words. How we work · Report an error