astronomy/ai produced the result/Astronomy and Astrophysics 2026 · v2
Neural networks sort millions of stars to build a young-star candidate catalogue
Astronomers turned star measurements into pictures and trained convolutional neural networks to judge whether each source is a young stellar object. The networks' majority vote defines the resulting NGYSO catalogue of 274,408 candidates.
spectrum · one line per step, placed by what the step does · bright lines used AI
The NEMESIS general young stellar object catalogue. I. Supervised classification with deep learning methods
Astronomy and Astrophysics, 2026
doi:10.1051/0004-6361/202660312 · record aix-00278 v2 · checked 2026-11-04
- AI was for
- Classification
- Model family
- Convolutional neural network
- Checked by
- Held-out92543 tested
- Code
- available
The finding the paper is about came from the AI.
What this research was about
Stars begin as clumps of gas and dust that collapse under their own gravity. For a while afterwards they are still wrapped in leftover material, and still settling towards the steady hydrogen-burning life of a star like the Sun. Astronomers call objects at this stage young stellar objects. Finding them matters for understanding how stars and planets form, but picking them out of a sky full of other things is awkward. Distant galaxies, cool red M dwarf stars, dying stars and ordinary variable stars can all look similar in survey data. The usual clue is the dust: it absorbs starlight and re-emits it as infrared, so a young star tends to look unusually bright at long wavelengths.
The researchers set out to make that judgement automatic and repeatable. They gathered sources already labelled as young stars from two curated lists, plus labelled examples of the various non-young-star object types that commonly masquerade as them. They then pulled together published catalogues of confirmed and candidate young stars, roughly 3.9 million entries, to be re-judged by one consistent method.
Where AI came in
Rather than feed numbers to the classifier, the team drew pictures. Each source's brightness measurements across wavelengths were plotted as a spectral energy distribution, the curve that shows how much light an object gives off at each colour. Patterns of brightness change over time from the ZTF survey became histogram images, and infrared survey cutouts supplied a third kind of picture. Convolutional neural networks, the sort of image-recognition software used to tell cats from dogs, were then trained from scratch on these images to answer one question: young star or not.
Twelve different network designs were trained per picture type, and a source counted as a candidate when at least six of the twelve agreed. That vote is what decides membership of the published catalogue, standing in for the eye of an astronomer reading each curve by hand. On a held-out test set the ensemble on plain spectral energy distribution images reached an F1 score, a combined measure of hits and misses, of about 96.9, with false positives below 0.5 per cent. On 92,543 external non-young-star sources, 914, or 0.99 per cent, were labelled young anyway.
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 turned astronomical source data into images (spectral energy distribution plots, ZTF variability maps and AllWISE cutouts) and trained convolutional neural networks to decide whether each source is a young stellar object. Twelve architectures per data product were combined into a majority-vote ensemble; the ensemble on plain SED images reached an F1 score of about 96.9 with false positives below 0.5% on the held-out test set. Applying this ensemble to roughly 3.9 million literature YSOs and candidates yielded the NGYSO catalogue of 274 408 unique candidates. On an external sample of 92 543 non-YSO sources, 914 (0.99%) were labelled YSO, and the classified sources fall in the pre-main-sequence region of the Gaia colour-absolute magnitude diagram.
How AI was used
Photometry collected from VizieR within a 2 arcsec radius was rendered as fixed-range logarithmic SED plots (and a variant encoding line-of-sight reddening as background colour), ZTF light curves were converted into delta-time/delta-magnitude histogram images, and AllWISE RGB stamps were retrieved via hips2fits. Ten off-the-shelf PyTorch CNN architectures were trained from scratch with a replaced binary output layer, cross-entropy loss with label smoothing, the AdamW optimiser and a One-Cycle schedule over 30 epochs; two custom lightweight networks, SmallResNet and a residual-plus-self-attention model, were trained with binary cross-entropy, AdamW and ReduceLROnPlateau with early stopping. Labelled YSOs came from the KYSO and NEMESIS Orion catalogues, and non-YSOs from literature samples of extragalactic sources, variables, M dwarfs, main-sequence and evolved stars, split at source level into train, validation and test sets. Per-model thresholds were set by a validation-set F1 sweep, and the twelve models for a given data product were combined by majority vote, with a source accepted as a candidate when at least 6 of 12 voted YSO. The SED-image ensemble was then run over the literature candidate compilation, and duplicate detections within 2 arcsec were merged.
The shape of the work
Structural · the record, drawn
Compile labelled YSO and non-YSO samples and literature candidate list
Obtaining raw data, whether by measurement, download or retrieval.
Our YSO training sample contains sources from two well-curated lists of young stars.where the paper describes this · verbatim
Query photometry, light curves and imaging cutouts
Obtaining raw data, whether by measurement, download or retrieval.
To collect the most complete data possible for our SEDs, we relied on the VizieR database.where the paper describes this · verbatim
Build image representations of SEDs, variability and imaging
Encoding data into features, descriptors, embeddings or graphs.
For each source in the dataset, we constructed and used several different image-based representationswhere the paper describes this · verbatim
Train CNN classifiers per data product
Fitting model parameters, including fine-tuning an existing model. The AI stood in for conventional algorithm.
we trained a set of CNNs spanning a wide range of parameter counts and design principles, available in the PyTorch Python frameworkwhere the paper describes this · verbatim
Classify literature candidates with the SED ensemble vote
Running a trained model over new data to predict, classify or score. The AI stood in for manual curation.
we applied our method to a compilation of literature catalogues containing confirmed YSOs, YSO candidateswhere the paper describes this · verbatim
Evaluate on held-out test set and external contaminant catalogues
Testing outputs against ground truth.
we applied the same voting system to a large sample of non-YSO sources from completely external catalogueswhere the paper describes this · verbatim
Astrophysical characterisation of the classified sample
Extracting understanding from model behaviour.
we used the slope of the infrared spectral energy distribution between 2 and 24 μ mwhere 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 catalogue is defined by the output of the CNN ensemble; membership of NGYSO is the classifier's decision
All these standard CNNs were trained from scratch with all parameters optimisedwhere the paper describes this · verbatim
The total sample comprises 92 543 objects with sufficient photometric coverage for classification.where the paper describes this · verbatim
The DLYSO software is released as open-source and is publicly available at GitHub.where the paper describes this · verbatim
What this paper did not report
Technical · absence is published deliberately
- Trained model weightsWhether the trained model is available is not stated.
- ComputeThe hardware or time used is not stated.
- Version of ShuffleNet V2 x0.5Which version of the model was used is not stated.
- Version of MNASNet 0.5Which version of the model was used is not stated.
- Version of MobileNet V3 SmallWhich version of the model was used is not stated.
- Version of RegNet-Y 400MFWhich version of the model was used is not stated.
- Version of EfficientNet-B0Which version of the model was used is not stated.
- Version of ResNet-18Which version of the model was used is not stated.
- Version of ResNet-50Which version of the model was used is not stated.
- Version of EfficientNet-V2-SWhich version of the model was used is not stated.
- Version of ResNeXt-50 32x4dWhich version of the model was used is not stated.
- Version of SmallResNetWhich version of the model was used is not stated.
- Version of ResConvAttnClassifier (RCA)Which version of the model was used is not stated.
- What step 6 replacedThe paper gives no basis for what the AI stood in for.
About this article
Record aix-00278, version 2, checked by a person on 2026-11-04. 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
Cite this record: AIxSci (2026). Neural networks sort millions of stars to build a young-star candidate catalogue. Record aix-00278, version 2, checked 2026-11-04. https://aixsci.org/articles/aix-00278/