astronomy/ai produced the result/Monthly Notices of the Royal Astronomical Society 2023 · v2
Neural network sifts Gaia data for 160,146 stars born outside the Milky Way
Astronomers trained a neural network to judge, from a star's position and motion alone, whether it was born in the Milky Way or swallowed from a smaller galaxy. It flagged 160,146 such stars among more than 27 million.
spectrum · one line per step, placed by what the step does · bright lines used AI
Exploring theex-situcomponents withinGaiaDR3
Monthly Notices of the Royal Astronomical Society, 2023
doi:10.1093/mnras/stad3817 · record aix-00075 v2 · checked 2026-10-08
- AI was for
- Classification
- Model family
- Multilayer perceptron
- Checked by
- Held-out
- Code
- not reported
The finding the paper is about came from the AI.
What this research was about
The Milky Way did not form in one piece. Smaller galaxies fell in over billions of years and their stars were absorbed, so today the Galaxy holds both stars born in place, called in-situ, and stars born elsewhere, called ex-situ. Telling the two apart is awkward. An accreted star looks much like a local one, and the usual way to spot it is through chemistry: measuring the proportions of elements such as magnesium, manganese, aluminium and iron in its light. That needs a detailed spectrum, which exists for only a small slice of the stars whose positions and motions are known.
The European Space Agency's Gaia mission has mapped positions and velocities for a vast number of stars without supplying those chemical fingerprints. The researchers set out to decide a star's origin from its movement alone, so that the judgement could be made across the whole catalogue rather than the chemically surveyed fraction of it.
Where AI came in
The network is a plain feed-forward classifier: five hidden layers of nodes that take a star's three-dimensional position, its velocity and its orbital actions, which are quantities summarising the shape of its path around the Galaxy, and return a probability that the star is ex-situ. It was trained in two stages. First on synthetic stars from a simulated galaxy, where each star's true origin is recorded. Then a second network reused the first with its settings fixed, adding a branch for the orbital actions, trained on real stars whose origins had been labelled by their chemistry.
So the network stands in for the spectrum. Where origin was previously read off from element abundances, it is inferred from motion, which Gaia measures for far more stars. Applied to 27,085,748 stars with a cut-off probability of 0.5, it flagged 160,146 as accreted. On held-back test data it scored an area under the curve of 0.98, a summary of how well a classifier separates two groups. Everything after the classification — picking out members of dwarf galaxies, clusters and streams, and estimating accreted fractions of 0.1 per cent in the thin disc, 1.6 per cent in the thick disc and 63.2 per cent in the halo — used ordinary cuts on the network's output.
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 trained a feed-forward neural network to decide, from 3D position, velocity and orbital actions alone, whether a star in Gaia DR3 was accreted from outside the Milky Way. A base network was trained on a mock Gaia catalogue derived from the m12i FIRE-2 simulation, where ex-situ origin is known, and a second network reused the frozen base network together with an action sub-network trained on stars labelled by their [Mg/Mn] and [Al/Fe] abundances from APOGEE DR17 and LAMOST DR8. Applied to a target sample of 27 085 748 stars, the model flagged 160 146 as ex-situ; the metallicity distribution of that sample peaks between -1.3 and -1.4 dex, and member stars of the Magellanic Clouds, Sagittarius, 20 globular clusters and six substructures were picked out of it. Average ex-situ percentages were estimated as 0.1% for the thin disc, 1.6% for the thick disc and 63.2% for the halo.
How AI was used
A fully connected network with five hidden layers (three of 128 nodes, two of 64), ReLU activations, batch normalisation and a single sigmoid output was built in Keras with a TensorFlow backend and trained with the Adam optimiser and a focal loss to handle extreme class imbalance. The first training phase used synthetic stars from a mock Gaia catalogue built on the m12i galaxy of the FIRE-2 Latte suite, with ex-situ labels taken from a per-particle flag, inputs being Galactocentric Cartesian positions and velocities and the data split 6:2:2 into training, validation and test sets. In the second phase all base-network weights were frozen and its output was concatenated with that of a sub-network taking the radial, vertical and azimuthal actions, with two further dense layers producing the classification; normalisation switched from a StandardScaler to a RobustScaler. Labels for this phase came from a segmental cut in the [Mg/Mn]-[Al/Fe] plane applied to Gaia DR3 stars cross-matched with APOGEE DR17 and the LAMOST DR8 value-added catalogue. Dynamic parameters for the observational sample were obtained by Monte Carlo sampling of each observable followed by action and orbit computation in AGAMA. The trained model was then run over the full target sample, with a threshold of 0.5 converting its continuous output into a binary ex-situ label; all subsequent membership selection and fraction estimates used conventional cuts on the resulting sample.
The shape of the work
Structural · the record, drawn
no AI
Build labelled synthetic training set
Cleaning, filtering, normalising or labelling data already obtained.
We labeled each star with the binary flag corresponding to the index in the text filewhere the paper describes this · verbatim
AI
Train kinematic base network NN_FIRE
Fitting model parameters, including fine-tuning an existing model.
The NN model was developed and trained using the Keras library, with TensorFlow serving as the backend framework.where the paper describes this · verbatim
no AI
Assemble Gaia DR3 target sample and derive actions
Cleaning, filtering, normalising or labelling data already obtained.
we exclude stars with positive total energy, leaving a target sample of 27 085 748 starswhere the paper describes this · verbatim
no AI
Chemically tag observational training labels
Cleaning, filtering, normalising or labelling data already obtained.
According to these criteria, 11 478 ex-situ stars were selected in our dataset.where the paper describes this · verbatim
AI
Train parallel dynamics-based network NN_parallel
Fitting model parameters, including fine-tuning an existing model.
The output of the base model was concatenated with the output of a sub-network that processes JR, Jz and JΦ.where the paper describes this · verbatim
AI
Classify the Gaia DR3 target sample
Running a trained model over new data to predict, classify or score. The AI stood in for conventional algorithm.
the NN_parallel model successfully identified 160 146 ex-situ starswhere the paper describes this · verbatim
no AI
Select members of dwarf galaxies, clusters and substructures
Reducing a candidate set by filtering or ranking, in a single pass.
We then selected 59 989 member stars from six substructures according to the order and criteria outlined in Table 4.where the paper describes this · verbatim
no AI
Estimate ex-situ fractions by location
Extracting understanding from model behaviour.
The average ex-situ percentages of the thin disc, the thick disc, and the halo are 0.1%, 1.6%, and 63.2%, respectively.where 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 paper's central product is the catalogue of ex-situ stars produced by the neural network classifier; all downstream analysis (substructure membership, ex-situ fractions) operates on the network's output
As a supervised learning algorithm, the performance of the NN is greatly influenced by the labels of the training set.where the paper describes this · verbatim
On the test set, NN_parallel achieved an AUC of 0.98, which is comparable to that of the NN_FIRE.where the paper describes this · verbatim
The catalogue of NN classification scores is available at Zenodowhere 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.
- How many were testedThe paper gives no count of what was tested.
- Version of NN_FIREWhich version of the model was used is not stated.
- Version of NN_parallelWhich version of the model was used is not stated.
- What step 2 replacedThe paper gives no basis for what the AI stood in for.
- What step 5 replacedThe paper gives no basis for what the AI stood in for.
About this article
Record aix-00075, 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