astronomy/ai produced the result/Astronomy and Astrophysics 2025 · v2
Classifying stars by type from a single wide-band telescope image
Researchers simulated Euclid-like images of stars in one broad colour band and trained machine learning models to sort each star into one of 13 spectral types. The classifiers did the sorting; a support vector machine worked from how well modelled blur patterns matched each image.
spectrum · one line per step, placed by what the step does · bright lines used AI
Breaking the degeneracy in stellar spectral classification from single wide-band images
Astronomy and Astrophysics, 2025
doi:10.1051/0004-6361/202452224 · record aix-00125 v2 · checked 2026-10-08
- AI was for
- Classification, Structure determination
- Model family
- Support vector machine, Convolutional neural network, Multilayer perceptron
- Checked by
- Held-out1000 tested
- Code
- available
The finding the paper is about came from the AI.
What this research was about
Stars differ in temperature and composition, and those differences show up in their spectra — the way their light is spread across colours. Astronomers summarise this with a spectral class. Getting it normally means splitting a star's light into a spectrum, or at least imaging it through several different filters. With only one wide filter, the picture collapses all those colours into a single number of counts per pixel, so very different stars can look much the same. That is the degeneracy in the paper's title: one image, many possible explanations.
There is a second reason to care. A telescope smears every point of light into a small blurred shape, the point spread function, and the exact shape depends on the star's colour. So modelling the blur well requires knowing the spectra of the stars used to measure it. The researchers set out to recover spectral class from a single wide-band image, and then to feed those recovered spectra back into fitting the blur model.
Where AI came in
All the classifying was done by trained models. The team made 13 000 simulated star images using a blur simulator called WaveDiff together with 13 spectral templates across eight wavelength slices. On 10 000 of these they trained two classifiers that look only at the pixels: one, reimplemented from earlier work, compresses each image into a small set of components and passes them to a committee of 48 small neural networks; the other uses a six-layer convolutional network to produce a compact feature vector for a network with 13 possible outputs.
The third approach adds knowledge of the optics. WaveDiff was fitted to 2 000 star images to give approximate blur models of varying accuracy. Each model's blur shape at a single wavelength was compared with the observed star, giving one similarity number per wavelength slice — a stand-in for the spectrum that a spectrograph or multi-band photometry would otherwise supply. A support vector machine, a standard method that draws boundaries between categories, was fitted to those numbers and used to predict classes for held-out stars. The assigned spectra were then used to refit the blur model.
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 simulated 13 000 Euclid-like single-band star images with the WaveDiff PSF simulator and used them to train classifiers that assign one of 13 stellar spectral templates to each star. Two pixel-only classifiers were compared with a PSF-aware method that computes similarity features between a star image and a fitted approximate PSF model at eight wavelengths and classifies those features with a support vector machine. On 1 000 held-out test stars the PSF-aware classifier reached 91% top-two accuracy using the ground-truth PSF and 76% top-two accuracy with the least precise PSF model tested, which was fitted from 50 stars; the paper reports it gave higher top-two accuracy than both pixel-only classifiers at every PSF error level considered. In a proof-of-concept test, refitting the PSF model using 2 000 stars with classifier-assigned SEDs gave a relative PSF error of 0.78%.
How AI was used
Simulated single-band star observations were produced from a ground-truth WaveDiff PSF field combined with 13 SED templates over eight spectral bins. Two pixel-only classifiers were trained on 10 000 of these images: a reimplementation of prior work that projects each image onto 24 PCA components and feeds the coefficients to a committee of 48 two-hidden-layer MLPs, and a replacement in which a six-layer convolutional encoder produces a 32-dimensional feature vector for an MLP with a 13-way softmax output. For the PSF-aware method, WaveDiff was fitted to 2 000 star observations divided into six nested datasets, yielding approximate PSF models of differing accuracy; monochromatic PSFs from each model were evaluated at every star position and compared with the star image using a Frobenius-norm similarity metric to give one feature per wavelength bin. An RBF-kernel C-support vector classifier from scikit-learn was fitted to these similarity features, separately for each approximate PSF model and for the ground-truth PSF, and then used to predict the spectral class of held-out stars. In a further test, SED templates assigned by the classifier to 2 000 stars were combined with 50 stars having ground-truth SEDs to retrain WaveDiff PSF models.
The shape of the work
Structural · the record, drawn
no AI
Simulate single-band star observations
Numerical or physics simulation, including where a learned surrogate replaces it.
Using the WaveDiff PSF simulator we generate a total of 13 000 star observations from a single ground truth PSF modelwhere the paper describes this · verbatim
AI
Train pixel-only classifiers
Fitting model parameters, including fine-tuning an existing model.
Both methods were trained with the same dataset of 10 000 simulated star observationswhere the paper describes this · verbatim
AI
Fit approximate PSF models
Fitting model parameters, including fine-tuning an existing model.
We employ 2 000 simulated star observations for training the approximate PSF modelwhere the paper describes this · verbatim
no AI
Compute similarity features
Encoding data into features, descriptors, embeddings or graphs.
From the monochromatic PSFs and the star observations we compute the similarity features following Eq. 4where the paper describes this · verbatim
AI
Train PSF-aware SVM classifier
Fitting model parameters, including fine-tuning an existing model.
We use the similarity features as the input to our SED classifierwhere the paper describes this · verbatim
AI
Assign spectral class and SED template to stars
Running a trained model over new data to predict, classify or score. The AI stood in for unresolved measurement.
make the predictions for each star in the test datasetwhere the paper describes this · verbatim
AI
Retrain PSF model with classified stars
Fitting model parameters, including fine-tuning an existing model. Its result feeds back into an earlier step.
the 2 000 newly classified stars are used together with the original 50 stars to train new WaveDiff PSF modelswhere the paper describes this · verbatim
no AI
Evaluate classification and PSF model error
Testing outputs against ground truth.
For each PSF model in Fig. 9 the relative error is averaged over the 1 000 test starswhere 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 result is the performance of its own classifiers; the spectral classes assigned to stars are produced entirely by the trained models
applying a support vector machine to these similarity features to classify the observed starwhere the paper describes this · verbatim
We evaluate both methods on the 1 000 test dataset using the aforementioned classification metricswhere the paper describes this · verbatim
The code is available here: https://github.com/CentofantiEze/sed_spectral_classificationwhere 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.
- DataWhether the data are available is not stated.
- ComputeThe hardware or time used is not stated.
- Version of WaveDiff PSF modelWhich version of the model was used is not stated.
- Version of PCA + MLP classifier committee (reimplemented from prior work)Which version of the model was used is not stated.
- Version of CNN + MLP classifierWhich version of the model was used is not stated.
- Version of SVM + PSF classifier (sklearn C-Support Vector Classification, RBF kernel)Which 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 3 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.
- What step 7 replacedThe paper gives no basis for what the AI stood in for.
About this article
Record aix-00125, 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