Machine-learning searches are helping astronomers identify rare hot subdwarfs in large survey datasets, according to a narrative review. One image-based search reported 263 new hot subdwarfs among 29,695 candidates found across 216,014 images, showing the scale at which automated screening can be applied.
The review covers selected studies on machine-learning applications to hot subdwarfs and white dwarfs. The literature was chosen mainly from the past decade, with emphasis on the past five years, and was not intended to be exhaustive. The paper introduces no new algorithm, benchmark dataset or quantitative comparison of methods.
The reported figures are summaries of cited work rather than a new independent validation. The review treats machine-learning outputs as discovery or priority-ranking tools, meaning a computer-generated candidate is not the same as a physically confirmed star.
From image searches to stellar atmospheres
Alongside image searches, cited studies applied spectral classifiers to the patterns in starlight. A hybrid model built around a convolutional neural network was reported to reach 96.17% testing accuracy on LAMOST spectra, while a kernel support-vector machine using ALS baseline correction was reported at 87.0%.
Other models used changes in a star’s apparent position and brightness to estimate which hot subdwarfs might be in binary systems. The predicted binary ratios varied sharply between groups: about 18% for targets showing neither astrometric nor photometric variability, more than 60% for targets with only photometric variability, about 77% for those showing both, and roughly 82% for those showing only astrometric variability.
Those figures are model predictions, and the review does not provide uncertainty intervals or complete subgroup denominators for them. An unsupervised Gaia pipeline, which looks for patterns without fixed labels, processed approximately 1,500 light curves and 49 statistical and physical parameters before using t-SNE and Gaussian Mixture Model clustering to explore possible groups. The review cautions that this embedding is qualitative and can change with the chosen features, settings and sample.
Machine learning was also used to estimate atmospheric properties from spectra. One model trained on 11,396 synthetic spectra and 945 observed LAMOST spectra reported mean absolute errors of 730 kelvin for effective temperature, 0.09 dex for surface gravity and 0.03 dex for helium abundance. The review notes that observed spectra remain scarce.
White-dwarf surveys reveal both scale and blind spots
The same approach appears in white-dwarf research, where models combine different kinds of survey evidence. Systems integrating Gaia astrometry with Sloan Digital Sky Survey spectroscopy were reported with accuracy above 90%, while multimodal convolutional neural networks, designed to combine several data types, were reported to approach 99% accuracy.
Unsupervised methods were used to separate magnetic and non-magnetic white-dwarf populations, estimate magnetic field strengths when direct measurements were unavailable and identify new candidates with signs of metal pollution. Other approaches combining neural networks with Gaussian-process classification identified tens of thousands of white-dwarf–main-sequence-star candidates in Gaia data, including systems in which the companion star dominated the observed spectrum.
But high accuracy does not automatically mean a complete or unbiased census. The review warns that training sets built from spectroscopically confirmed SDSS or LAMOST objects may not represent white-dwarf candidates selected photometrically in Gaia. Estimated class fractions and occurrence rates are therefore conditional on how the training sample was assembled, rather than unbiased measures of the full population.
The review also warns that rare-object searches are especially vulnerable to misleading accuracy scores. When the class of interest is much rarer than the contaminants, a model can appear accurate while still missing positives or admitting many false candidates. The authors recommend reporting precision, recall, F1 scores and the area under the precision–recall curve instead of relying on accuracy alone.
The telescope work is not finished
Across both groups of stars, machine learning is presented as a way to discover and rank targets for attention, not as the final word on their physical nature. Dedicated spectroscopy and observations that track changes over time are still needed to confirm candidates.
That means the review does not independently establish particular binary-formation pathways from the cited results. Its central message is that machine-learning classifications and clusters need physical confirmation before they can support stronger conclusions about stellar populations and evolution.
No new data were generated for the paper. The discussed data are publicly available in the cited original publications, and the Gaia DR3 data used for its figure are available through the Gaia Archive.
Paper data and sources
Original title: Exploring Late Stellar Evolution in the Era of Large Surveys: Machine Learning Prospects for Hot Subdwarfs and White Dwarfs
Authors: Princy Ranaivomanana, Murat Uzundag
Journal/Repository: arXiv
Status: Peer-reviewed
First online: 2026-08-26
DOI: 10.3390/universe12090257
Original paper · Full text