astronomy/ai produced the result/arXiv 2026 · v2
Neural network measures rotation periods for Kepler's main-sequence stars
Researchers trained a neural network on brightness measurements from the Kepler space telescope to work out how fast stars spin. The model produced a catalogue of rotation periods, each with a calibrated uncertainty range.
spectrum · one line per step, placed by what the step does · bright lines used AI
The Maunder Model and Catalog: Stellar Rotation, Bimodal Activity, and Magnetic Braking in Kepler Main-Sequence Stars
arXiv, 2026
doi:10.48550/arxiv.2608.06604 · record aix-00221 v2 · checked 2026-10-09
- AI was for
- Property prediction
- Model family
- Transformer, Convolutional neural network
- Checked by
- Held-out4187 tested
- Code
- not reported
The finding the paper is about came from the AI.
What this research was about
Stars spin, and the speed matters. Rotation drives a star's magnetic field, which in turn shapes the radiation its planets receive, and a star's spin slows over time as its magnetic field drags on escaping gas. That slowing makes rotation a rough clock for stellar age. But you cannot watch a distant star turn. Instead, astronomers track its brightness: dark patches on the surface, much like sunspots, sweep in and out of view and make the light dip and rise in a repeating pattern. The period of that pattern is the rotation period. In practice the signal is faint, the patches come and go, and standard methods often disagree, including on whether a measured period is the true one or half of it.
The researchers used light curves from NASA's Kepler mission, which stared at one patch of sky and recorded stellar brightness every thirty minutes. They restricted the sample to main-sequence stars, those in the long stable phase of life, and set out to produce a single rotation-period catalogue with an uncertainty attached to every star.
Where AI came in
A neural network, named The Maunder, did the measuring. Each light curve was cut into overlapping 450-day stretches and turned into six parallel inputs: smoothed brightness, two measures of how variable the star was, and two standard ways of summarising the repeating patterns in a signal. The network read these through components built for sequence data and returned five quantiles of the rotation period, which together describe a range rather than a single value. It learned in two ways at once: from stars where published catalogues already agreed on a period, and, without any labels, from the full sample by comparing two different stretches of the same star.
The model stood in for the conventional signal-processing recipes astronomers use to pick a period out of a light curve. Sliding it across each star's full record produced a period estimate per window; where those estimates fell into two separate groups, the star was flagged as having two candidate periods. On a held-out set of 4,187 stars with agreed periods, the predicted ranges covered the true values as often as advertised, and comparisons with independent measurements of stellar spin speed indicated which of the two candidates was the real rotation period.
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
A neural model called The Maunder was trained on Kepler light curves to predict stellar rotation periods, using a self-supervised objective on all 148,746 main-sequence stars and a supervised quantile-regression objective on the 41,650 stars whose periods agree between at least two published catalogs. On the held-out consensus test set it reaches an RMSE of 2.36 days and a median absolute error of 0.46 days, with conformally calibrated predictive intervals whose empirical coverage matches the nominal levels. Running inference on rolling windows, 31,953 stars (21.5%) show two well-separated period modes; a forward model of APOGEE projected rotation velocities indicates that for non-harmonic cases the longer mode corresponds to the rotation period. Filtering on the normalised 80% interval leaves 119,428 stars, whose period distribution has a median of 17.94 days, and the catalog is used to examine metallicity dependence of rotation at fixed mass, equatorial velocity and specific angular momentum across the Kraft break, gyrochronology sequences and synchronized binaries.
How AI was used
Kepler long-cadence PDCSAP light curves were cut into 450-day windows and converted into six channels: two normalisations of Savitzky-Golay-smoothed flux, two percentile-difference activity proxies whose window lengths are scaled by a per-star reference period, and the autocorrelation function and Lomb-Scargle periodogram. The reference period came from a first four-channel version of the same model, whose median prediction and interval width were frozen per star and fed back only as the geometry of the activity-proxy windows. The full model shares an encoder across two random windows of the same star, passing time-domain channels through an AstroConformer module and frequency-domain channels through a CNN, concatenating the token sequences into a conformer mixer, and combining the views in a DualFormer module with duality, variance and covariance terms. A small MLP head predicts five quantiles of log10(Prot) under a multi-quantile pinball loss applied to the consensus-labelled stars, while the self-supervised term is applied to all stars; intervals were calibrated post hoc by split conformalized quantile regression on the validation set. Final predictions were produced by rolling inference with a 90-day stride, after which per-window periods were aggregated, partitioned by a one-dimensional 2-means split into unimodal, distinct bimodal and 2:1 harmonic classes, and released with per-star reliability metrics.
The shape of the work
Structural · the record, drawn
no AI
Obtain Kepler light curves
Obtaining raw data, whether by measurement, download or retrieval.
We use long-cadence light curves from the Kepler mission Data Release 25, corrected for instrumental systematics with the PDC-MAP pipeline.where the paper describes this · verbatim
no AI
Select main-sequence sample and build consensus labels
Cleaning, filtering, normalising or labelling data already obtained.
Supervised labels are assigned only to stars with a consensus rotation period, which we define as agreement to within 20%where the paper describes this · verbatim
no AI
Build multi-scale six-channel inputs
Encoding data into features, descriptors, embeddings or graphs.
The full input consists of a 6-channel light curve, with 4 time-domain channels and 2 frequency-domain channelswhere the paper describes this · verbatim
AI
Train scaffold model for reference period
Fitting model parameters, including fine-tuning an existing model. The AI stood in for conventional algorithm. Its result feeds back into an earlier step.
we first train a four-channel version of the same model, using only the channels that require no period scaffoldwhere the paper describes this · verbatim
AI
Train full model with hybrid objective and conformal calibration
Fitting model parameters, including fine-tuning an existing model. The AI stood in for conventional algorithm.
The model used to predict the periods is a machine learning model trained with both self-supervised and supervised objectives.where the paper describes this · verbatim
AI
Rolling-window period inference
Running a trained model over new data to predict, classify or score. The AI stood in for conventional algorithm.
We evaluated the model on consecutive windows with a stride of 90 days and aggregated the per-window predictions.where the paper describes this · verbatim
no AI
Aggregate windows, flag bimodality, assign reliability metrics
Cleaning, filtering, normalising or labelling data already obtained.
A star is labelled bimodal when the two modes are separated by at least 0.20 dexwhere the paper describes this · verbatim
no AI
Evaluate on held-out set and adjudicate modes against spectroscopy
Testing outputs against ground truth.
we bring in an independent spectroscopic indicator, the projected rotation velocitywhere 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.
The published product is a model-produced rotation-period catalog; the bimodality result and all downstream population analyses rest on the model's per-window predictions
The time-domain channels are processed through an AstroConformer module, which combines transformers and CNN blocks.where the paper describes this · verbatim
Self-supervised objective comparison on the held-out main-sequence consensus test set (n=4,187)where the paper describes this · verbatim
The data are available at MAST.where 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.
- ComputeThe hardware or time used is not stated.
- Version of The Maunder (six-channel full model)Which version of the model was used is not stated.
- Version of The Maunder scaffold model (four-channel)Which version of the model was used is not stated.
- Version of VICReg-objective variant (self-supervised baseline)Which version of the model was used is not stated.
- Version of Time-only modelWhich version of the model was used is not stated.
- Version of Frequency-only modelWhich version of the model was used is not stated.
About this article
Record aix-00221, version 2, checked by a person on 2026-10-09. 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