Determining precise stellar ages and masses for evolved giants is crucial for Galactic archaeology but challenged by spectral degeneracies. Gaia's low-resolution XP spectra offer a unique opportunity to infer these parameters on a massive scale using data-driven methods. We extend a transformer-based astronomical foundation model to evolved stars, establishing a unified framework to simultaneously predict atmospheric parameters (T_eff, log g, [M/H]) and evolutionary labels (mass, age) with physical consistency. Treating spectra as token sequences, we integrated mass and age into the model's vocabulary. The model is trained on Gaia XP spectra cross-matched with the APOGEE DR17 DistMass catalog. Our generative approach enables flexible input handling, including spectral inpainting and parameter-to-spectrum generation. On an independent test set, the model achieves a prediction scatter of σ≈ 0.114 M_⊙ for mass and σ≈ 1.334 Gyr for age. Beyond numerical accuracy, it successfully reproduces the giant branch's mass-luminosity relation and autonomously disentangles interstellar extinction from intrinsic temperature variations without explicit physical priors. It also robustly recovers missing spectral data and estimates reliable uncertainties. Validating that foundation models can internalize stellar physics from data, this physically-aware, probabilistic framework offers a powerful tool for unraveling Milky Way history using large-scale spectroscopic surveys.

The Large Sky Area Multi-Object Fiber Spectroscopic Telescope (LAMOST) has collected tens of millions of spectra, providing an unprecedented resource for large-scale spectroscopic studies. Efficient retrieval techniques are therefore essential for exploring such massive datasets. Existing approaches often rely on predefined templates or manually labeled training samples, which can limit their applicability in large and diverse spectral archives. In this work, we present a general similarity-retrieval framework that combines self-supervised contrastive learning based on a convolutional neural network with Facebook AI Similarity Search (FAISS) for efficient large-scale spectral retrieval. The framework learns spectral representations directly from unlabeled data and enables flexible retrieval from user-defined wavelength regions based on feature similarity. We evaluate the framework on several stellar populations in LAMOST DR8. For late-type M8-star retrieval, 90.5% of the top 1000 retrieved spectra are later than M6. For M0–M5 giants, the mean retrieval accuracy across six subtypes reaches 94.8%. Using a C-H star spectrum as the query spectrum, 90.8% of the top 1000 retrieved candidates are classified as carbon stars by the LAMOST pipeline. Cross-matching with SIMBAD further confirms 255 C-H stars and 47 C-R stars among the retrieved candidates. These results demonstrate that the proposed framework can efficiently identify spectrally similar objects across large spectroscopic databases and can serve as a useful tool for searching for rare or spectrally distinctive stellar populations.

We present a homogeneous catalog of stellar atmospheric parameters and chemical abundances for 20,032 members across 800 open clusters, derived from Gaia DR3 XP sampled spectra. To extract detailed chemical information from low-resolution spectrophotometry, we developed a deep learning framework incorporating a novel boosting multihead attention mechanism. This architecture iteratively reweights attention heads to enhance sensitivity to subtle spectral features often masked by noise. We constructed a robust training set using APOGEE DR17 labels, employing a histogram distribution matching strategy to strictly align the training sample with the target cluster population in the color-magnitude diagram. This approach mitigates domain shift and ensures reliable inference for red giant branch, red clump, and lower main-sequence stars. The final catalog provides estimates of Teff, logg , [M/H], and individual abundance ratios for [alpha/M], C, N, O, Mg, Si, Ca, and Ti, along with epistemic uncertainties. Validation against independent datasets demonstrates high precision, with typical mean absolute errors of similar to 37 K for Teff, similar to 0.10 dex for logg , and similar to 0.05 dex for [M/H]. This work highlights the potential of physics-aware deep learning in the Gaia era and offers a valuable resource for Galactic archeology.

Ultra-diffuse galaxies (UDGs) are a subset of low-surface-brightness galaxies (LSBGs), showing mean effective surface brightness fainter than 24 mag arcsec(-2) and a diffuse morphology, with effective radii larger than 1.5 kpc. Due to their elusiveness, using traditional methods over large sky areas is challenging. Here we present a catalog of UDG candidates identified in the full 1350 deg(2) area of the Kilo-Degree Survey (KiDS) using deep learning. In particular, we used a previously developed network for the detection of low-surface-brightness systems in the Sloan Digital Sky Survey (SBGnet) and optimized for UDG detection. We trained this new UDG detection network for KiDS (UDGnet-K), with an iterative approach, starting from a small-scale training sample. After training and validation, the UGDnet-K has been able to identify similar to 3300 UDG candidates, of which, after visual inspection, we selected 545 high-quality ones. The catalog contains the independent rediscovery of previously confirmed UDGs in local groups and clusters (e.g., NGC 5846 and Fornax), and new discovered candidates in about 15 local systems, for a total of 67 bona fide associations. Besides the value of the catalog per se for future studies of UDG properties, this work shows the effectiveness of an iterative approach to training deep learning tools in presence of poor training samples, due to the paucity of confirmed UDG examples, which we expect to replicate for upcoming all-sky surveys such as those by the Rubin Observatory, Euclid, and the China Space Station Telescope.

Metal-poor stars (MP, [Fe/H]<−1.0) retain the chemical signatures of the early Universe, making their α-element invaluable for tracing the Galactic chemical evolution. Traditional optical spectroscopic methods and machine learning approaches often struggle at low metallicities and neglect the diagnostic power of ultraviolet (UV) spectral features (e.g., OH bands, Mg II/Mg I lines) that are accessible to the forthcoming Chinese Space Station Telescope (CSST). Using over 1.8 × 105 simulated CSST spectra, we developed the spectral transformer (SPT)-GloCal model to quantify the impact of UV (2550−4000 Å) low-resolution (R ≈ 200) spectra on the precision of [O/Fe] and [Mg/Fe] estimates in MP stars. The model improves local attention through score-aware competitive filtering and integrates it with global attention via a learnable gating mechanism. We compared models trained on full spectra versus optical-only spectra. Incorporating UV spectra halved the mean absolute error (MAE) of [O/Fe] predictions from 0.0785 to 0.0367 dex and reduced the scatter (σ) from 0.135 to 0.063 dex. For [Mg/Fe], MAE decreased from 0.0010 to 0.0006 dex and σ from 0.054 to 0.0068 dex. Error analyses demonstrated that UV data lead to more stable and accurate estimates, especially at a high log g and low Teff. SPT-GloCal outperforms both its predecessor (SPT) and tree-based regressors, confirming the effectiveness of its global-local attention design. Furthermore, tests on spectra with simulated noise show that the model maintains consistent performance across a range of signal-to-noise ratios (S/N =30−50), with marginal gains over the SPT model at the lower S/N end. UV spectral features, even at low resolution (R ≈ 200), enhance α-element abundance determinations. The SPT-GloCal framework provides a scalable solution for upcoming space-based surveys. Future work will apply this model to real CSST data, thus extending it to other elements.

Context. Hot subdwarf stars are important tracers of stellar structure and evolution, while their binary systems provide key constraints on their formation channels. Nevertheless, the number of confirmed hot subdwarf binaries is still limited. Aims. We aim to identify hot subdwarf binaries efficiently and reliably in large candidate samples by combining heterogeneous observational information. Methods. We developed a two-stage classification framework. A Bayesian neural network was first used to perform preliminary classification and uncertainty quantification from Gaia photometric and astrometric data. We then constructed a multimodal deep learning model, HsdB-FusionNet, which combines Gaia photometric and astrometric features with SDSS spectra through a cross-attention mechanism. We applied this framework to a catalog of 61 585 hot subdwarf candidates. The resulting candidates were further examined through spectral energy distribution fitting, and atmospheric parameters were derived for high-quality systems. Results. We identified 1369 hot subdwarf binary candidates from the parent sample. Among them, 1057 objects were strongly supported as binaries by spectral energy distribution fitting. Atmospheric parameters were obtained for 579 high-quality binary systems. The interpretability analysis further indicates that the model decisions are consistent with known physical priors. Conclusions. This work provides a substantially enlarged sample of hot subdwarf binary candidates and demonstrates the potential of multimodal deep learning for binary identification. The released catalog offers a useful basis for future studies on the formation and evolution of hot subdwarf binaries.

Low-surface-brightness galaxies (LSBGs) play an important role in studies of galaxy formation and evolution, yet accurate measurements of their structural and photometric parameters remain challenging for conventional analysis pipelines due to their diffuse light distributions and low signal-to-noise ratios. In this work, we present an automated deep learning framework, LSBGPENet, for the robust estimation of structural and photometric parameters of LSBGs from wide-field imaging data. The framework directly operates on galaxy image cutouts and simultaneously infers key parameters, including total magnitude (m), effective radius (Reff), ellipticity (& varepsilon;), and S & eacute;rsic index (n), together with associated uncertainty estimates. Mean and central surface brightnesses (mu eff, mu 0) are subsequently derived from the inferred parameters. We assess the accuracy and reliability of the inferred parameters through comparisons with traditional profile-fitting measurements. On both simulated data and observational data from the Dark Energy Survey, the framework achieves high predictive accuracy, with mean coefficients of determination of 0.91 and 0.94, respectively, and well-calibrated uncertainty estimates, characterized by mean uncertainty calibration errors of 0.004 and 0.009. The inferred parameters are statistically consistent with those obtained from GALFIT. The proposed framework provides a scalable and reproducible solution for structural and photometric parameter estimation of LSBGs and is well suited for application to current and forthcoming wide-field surveys, including the China Space Station Telescope.

Ultra-diffuse galaxies (UDGs) are a class of galaxies characterized by an extremely low surface brightness and large effective radius. The significance of UDGs lies in their unique properties, such as their diverse dark-matter content, which collectively challenge existing theories of galaxy formation and evolution. However, their low surface brightness and diffuse stellar distributions make UDGs particularly difficult to detect. To address this challenge, this study introduces a deep learning-based object detection model for the Dark Energy Survey (DES), named UDGnet-DES. Combined with an iterative training strategy, this model is designed to conduct a large-scale search for UDG candidates in the DES Data Release 2 (DES DR2) imaging data. Using our model, we searched UDGs from more than 500,000 DES images and further filtered the results based on surface brightness and visual inspection. As a result, we obtained a catalog of 2991 UDG samples, 39 of which have spectroscopic redshifts. An analysis of our UDG samples reveals that most objects exhibit nearly circular morphologies. Among them, blue UDGs tend to have higher surface brightnesses, while red UDGs show significant spatial clustering, in contrast to the more uniformly distributed blue UDGs.The method developed in this study can improve the efficiency of UDG searches and will be applied to the search for UDGs in the China Space Station Telescope (CSST) survey project.

The blue horizontal-branch (BHB) stars are horizontal-branch stars bluer than the RR Lyrae instability strip in the Hertzsprung–Russell diagram, serving as ideal tracers for studying the structure and evolution of the Milky Way. With the accumulated photometric image data from the Sloan Digital Sky Survey (SDSS), we attempt to use object detection techniques to directly locate the position of BHB stars from the images. Given that BHB stars appear extremely tiny in images captured by the SDSS telescope, many existing object detection algorithms are unsuitable for detecting such tiny objects. In this study, we propose a blue horizontal-branch star detector (BHBDet), the first object detection algorithm with six detection heads designed for stellar detection. BHBDet achieves a precision of $80.4\%$, a recall of $91.1\%$, and an $F_{1}$ score of $85.4\%$, significantly outperforming popular algorithms with three detection heads. Among 109696 images containing at least 11 million objects, we detected 49852 BHB candidates. These candidates exhibit reduced proper motions primarily ranging from 5–$20$ mas yr$^{-1}$, and colors primarily within $-1.5 < u - g < 2$, $-2 < g - r < 0.5$, $-2 < r - i < 1$, and $-2 < i - z < 2$. We further estimated the atmosphere parameters for these candidates, which primarily fall within $7000 < T_{\mathrm{eff}} < 12000$ K, $3 < \log g < 5$ dex, and $-3 < {\rm [Fe/H]} < -1$ dex. We also find that approximately $30\%$ of the candidates are located beyond $35$ kpc, with some exceeding $100$ kpc. By applying Balmer line profile cuts, we confirmed 2093 BHB stars from those with the spectra provided by the Large Sky Area Multi-Object Fiber Spectroscopic Telescope. We have published this catalog online to further enrich the BHB population and advance the research on the structure of the Milky Way.

Context. Hot subdwarf binaries provide stringent constraints on envelope stripping and binary mass transfer. However, robust identification of these systems in large spectroscopic surveys is hindered by variable data quality and the subtle signatures of their companions. Aims. We aim to develop a robust automated framework for identifying binaries comprising a hot subdwarf and a main-sequence star within the diverse datasets of the LAMOST low-resolution survey, utilizing deep learning to overcome challenges posed by complex noise and class imbalance. Methods. We propose a Bayesian hybrid model that combines convolutional neural networks for local feature extraction with a Bayesian transformer encoder to model long-range dependencies. This architecture incorporates variational inference to improve classification accuracy and facilitate robust candidate identification. Results. The framework attains an accuracy of 95.3% on the test set, with a precision of 97.2% for the binary class. Applying the model to the LAMOST dataset yields 1161 binary candidates. Further certification via spectral energy distribution fitting confirms 968 binaries among the 1001 objects with reliable fits. Multi-epoch radial velocity measurements of 352 candidates identified 119 systems with significant variability. Conclusions. A subset of the candidates show no significant radial velocity changes over baselines exceeding 1000 days despite exhibiting infrared excess, consistent with the behavior of long-period binaries. The proposed Bayesian framework effectively quantifies predictive uncertainty to filter out data artifacts, yielding a sample suitable for constraining evolutionary pathways such as common envelope ejection and stable Roche lobe overflow.

Context. High-resolution theoretical stellar spectral modeling provides essential support for the analysis and interpretation of observed spectra, underpinning studies of stellar atmospheric physics and the processing of large-scale spectroscopic surveys. Aims. This work is aimed at developing a physically self-consistent modeling approach for theoretical stellar spectrum generation and parameter inversion under the assumptions of local thermodynamic equilibrium (LTE) and plane-parallel atmospheres. It is designed to mitigate the high computational cost of traditional models and the limited physical interpretability of purely data-driven approaches, as well as the training instability and reduced adaptability to sparse grids encountered by physics-informed neural networks in high-dimensional spectral modeling. Methods. We present PhysFormer, a framework that incorporates key stellar atmospheric processes, including the radiative transfer equation and the flux conservation constraint, directly into the network as part of the generative mechanism, rather than merely as external loss terms. A physically consistent autoencoder is used to learn a low-dimensional latent representation, enabling physically consistent forward spectral generation and robust parameter inversion through a three-stage modeling framework. Results. Experiments based on the PHOENIX theoretical stellar spectral dataset show that PhysFormer achieves a spectral generation root mean square error (RMSE) of 0.0054 and R2 of 0.9997, outperforming existing approaches. The model improves stability in degenerate inversion tasks across different signal-to-noise ratios and reproduces the characteristic responses of diagnostic spectral lines (e.g., Hα, Hβ, Ca II K, and Fe I), consistent with stellar atmosphere theory under LTE. PhysFormer provides a physically consistent and efficient framework for theoretical stellar spectral modeling and parameter inversion, offering a potential pathway for integrating physical constraints into data-driven models in astrophysical applications.

Chemically peculiar (CP) stars, particularly CP1 (metallic-line) and CP2 (magnetic Ap/Bp) stars, are crucial for studying microscopic processes in stellar atmospheres, such as atomic diffusion and magnetic interactions. In this study, we present ODCNNnet, a deep learning-based classification model designed to identify CP1 and CP2 stars from low-resolution spectra. The model achieves an accuracy of 99.81% and a recall of 99.43% for CP1 stars, and an accuracy of 98.82% with a recall of 99.41% for CP2 stars. Applying ODCNNnet to B- to early F-type stars in LAMOST DR11, we identified 33,049 CP1 stars and 5901 CP2 stars, including 5287 newly identified CP1 stars and 1153 newly identified CP2 stars, validated through cross matching with the MKCLASS tool. We further analyzed the stellar parameters of CP stars. The effective temperature ( T eff ) of CP1 stars ranges from 6000 K to 9000 K, peaking at 7500 K, while CP2 stars span 6000 K to 13,000 K, with peaks near 8900 K and 11,900 K. The surface gravity ( log g ) of CP1 stars is concentrated between 3.4 and 4.6 dex, peaking at 3.9 dex, while CP2 stars range from 3.5 to 4.6 dex. The [Fe/H] of CP1 stars is generally between −1.0 and 0.8 dex, peaking around −0.1 dex, whereas CP2 stars exhibit a broader distribution from −1.5 dex to 1.0 dex, with peaks around −0.25 dex and 0.75 dex. This study provides a valuable data set for investigating the physical properties of CP stars and their underlying formation mechanisms.

The precise determination of stellar atmospheric parameters (effective temperature T_ eff surface gravity log g, and metallicity Fe/H ) serves as a cornerstone of Galactic studies. In this work, we develop a novel deep learning approach, the Atmospheric CSWin Framework (ACF), to measure these parameters with high precision. The ACF employs a dual-input architecture that combines astrometric data (parallaxes and their corresponding errors) from Gaia Early Data Release 3 with photometric images from the fourth data release (DR4) of the SkyMapper Southern Survey (SMSS). The framework utilizes a CSWin Transformer backbone for hierarchical feature extraction from photometric images, integrated with Monte Carlo dropout in the prediction module for robust uncertainty quantification. Trained on cross-matched stars between SMSS DR4 and the third data release of the Galactic Archaeology with HERMES spectroscopic survey, the ACF achieves parameter estimates with dispersions of 95.02 K for T_ eff 0.07 dex for log g, and 0.14 dex for Fe/H . Systematic experiments demonstrate that incorporating parallax information significantly improves the precision of all three parameters, especially log g. Our image-based methods outperform traditional approaches based on stellar magnitudes or colors, with improvements ranging from 2% to 14%. The ACF yields parameter estimates approaching those of high-resolution spectroscopic analyses, and the framework remains effective even for low-quality samples, highlighting its robustness and general applicability. Using the ACF, we compiled a comprehensive catalog of atmospheric parameters for one million SMSS DR4 stars.

White dwarfs represent the end stage for 97% of stars, making precise parameter measurement crucial for understanding stellar evolution. Traditional estimation methods involve fitting spectra or photometry, which require high-quality data. In recent years, machine learning has played a crucial role in processing spectral data due to its speed, automation, and accuracy. However, two common issues have been identified. First, most studies rely on data with high signal-to-noise ratios (SNR > 10), leaving many poor-quality datasets underutilized. Second, existing machine learning models, primarily based on convolutional networks, recurrent networks, and their variants, cannot simultaneously capture both the spatial and sequential information of spectra. To address these challenges, we designed the Estimator Network (EstNet), an advanced algorithm integrating multiple techniques, including Residual Networks, Squeeze and Excitation Attention, Gated Recurrent Units, Adaptive Loss, and Monte-Carlo Dropout Layers. We conducted parameter estimation on 5,965 poor-quality white dwarf spectra (R~1800, SNR~1.17), achieving average percentage errors of 14.86% for effective temperature and 3.97% for surface gravity. These results are significantly superior to other mainstream algorithms and consistent with the outcomes of traditional theoretical spectrum fitting methods. In the future, our algorithms will be applied for large-scale parameter estimation on the Chinese Space Station Telescope and the Large Synoptic Survey Telescope.

The formation and evolution of ring structures in galaxies are crucial for understanding the nature and distribution of dark matter, galactic interactions, and the internal secular evolution of galaxies. However, the limited number of existing ring galaxy catalogs has constrained deeper exploration in this field. To address this gap, we introduce a two-stage binary classification model based on the Swin Transformer architecture to identify ring galaxies from the DESI Legacy Imaging Surveys. This model first selects potential candidates and then refines them in a second stage to improve classification accuracy. During model training, we investigated the impact of imbalanced data sets on the performance of the two-stage model. We experimented with various model combinations applied to the data sets of the DESI Legacy Imaging Surveys DR9, processing a total of 573,668 images with redshifts ranging from z_spec = 0.01–0.20 and mag _r < 17.5. After applying the two-stage filtering and conducting visual inspections, the overall precision of the models exceeded 64.87%, successfully identifying a total of 8052 newly discovered ring galaxies. With our catalog, the forthcoming spectroscopic data from DESI will facilitate a more comprehensive investigation into the formation and evolution of ring galaxies.

Double-lined spectroscopic binaries (SB2s) hold significant importance for understanding stellar formation and evolution. With the release of massive spectra, deep learning techniques have achieved major breakthroughs in the search of spectroscopic binaries. However, traditional deep learning models are unable to provide predicted uncertainty, which raises concerns about the reliability of the results. In this study, we propose a model called SB ^2 Net to identify SB2s in the Large Sky Area Multi-Object Fiber Spectroscopic Telescope Medium-resolution Survey (LAMOST-MRS) and estimate predicted uncertainty. SB ^2 Net integrates multiple techniques, enabling the efficient extraction of critical features from the cross-correlation function and providing high-confidence classification results. On the test set, our model achieved a precision of 99.53%, with 99% of the samples having epistemic uncertainty less than 0.0017 and aleatoric uncertainty less than 0.0681. By applying SB ^2 Net to 2,526,141 spectra from LAMOST-MRS data release 11, we identified 7711 SB2 candidates, of which 3037 were newly discovered. Additionally, we estimated radial velocities (RVs) and RV errors for all candidate spectra, while validating the consistency of the candidates properties with other catalogs through the color–magnitude diagram. Finally, by fitting the RV curves, we obtained 246 well-fitted orbital solutions with most systems having mass ratio converge to 1. The period–eccentricity relationship reveals that short-period binaries tend to exhibit more circular orbits. In the future, we plan to separate the composite spectra of SB2s to study the properties and atmospheric parameters of the component stars.

Hot subdwarf stars are important celestial objects in the study of stellar physics, but the population remains limited. The LAMOST DR12-V1, released in 2025 March, is currently the world's largest spectroscopic database, holding great potential for the search of hot subdwarf stars. In this study, we propose a two-stage deep learning model called the hot subdwarf network (HsdNet), which integrates multiple advanced techniques, comprising a binary classification model in stage one and a five-class classification model in stage two. HsdNet not only achieves high precision with 94.33% and 94.00% in the binary and the five-class classification stages, respectively, but also quantifies the predicted uncertainty, enhancing the interpretability of the classification results through visualizing the model's key focus regions. We applied HsdNet to the 601,217 spectra from the LAMOST DR12-V1 database, conducting a two-stage search for hot subdwarf candidates. In stage one, we initially identified candidates using the binary classification model. In stage two, the five-class classification model was used to further refine these candidates. Finally, we confirmed 1008 newly identified hot subdwarf stars. The distribution of their atmospheric parameters is consistent with that of known hot subdwarf stars. These efforts are expected to significantly advance the research on hot subdwarf stars.

White dwarfs (WDs) are the ultimate stage for approximately 97% of stars in the Milky Way and are crucial for studying stellar evolution and galaxy structure. Due to their small size and low luminosity, WDs are not easily observable. Traditional search methods mostly rely on analyzing photometric parameters, which need high-quality data. In recent years, machine learning has played a significant role in astronomical data mining, due to its speed, real time, and precision. However, we have identified two common issues. On the one hand, many studies are based on high-quality spectral data, while a large amount of image data remain underutilized. On the other hand, existing astronomical algorithms are essentially classification algorithms, with sample incompleteness being a critical weakness. In our study, we propose the WD Network (WDNet) algorithm, which is a new object detection algorithm that integrates multiple advanced technologies and can directly locate WDs in images. WDNet overcomes the degradation issue of WDs and detected 31,065 candidates in 80,448 images. The candidates exhibit a wide range of types, including DA, DB, DC, DQ, and DZ, with surface gravity within 7.8 dex ∼ 8.4 dex, effective temperatures within 10,000 K ∼ 56,000 K, colors within −1 < u − g < 1 and −0.8 < g − r < 0.4, and reduced proper motion within 20∼35 mag. In the future, WDNet will conduct large-scale searches using the Chinese Space Station Telescope and Sloan Digital Sky Survey V.

Ultra-diffuse Galaxies (UDGs) are a subset of Low Surface Brightness Galaxies (LSBGs), showing mean effective surface brightness fainter than 24 mag arcsec^-2 and a diffuse morphology, with effective radii larger than 1.5 kpc. Due to their elusiveness, traditional methods are challenging to be used over large sky areas. Here we present a catalog of ultra-diffuse galaxy (UDG) candidates identified in the full 1350 deg^2 area of the Kilo-Degree Survey (KiDS) using deep learning. In particular, we use a previously developed network for the detection of low surface brightness systems in the Sloan Digital Sky Survey and optimised for UDG detection. We train this new UDG detection network for KiDS (UDGnet-K), with an iterative approach, starting from a small-scale training sample. After training and validation, the UGDnet-K has been able to identify ∼3300 UDG candidates, among which, after visual inspection, we have selected 545 high-quality ones. The catalog contains independent re-discovery of previously confirmed UDGs in local groups and clusters (e.g NGC 5846 and Fornax), and new discovered candidates in about 15 local systems, for a total of 67 bona fide associations. Besides the value of the catalog per se for future studies of UDG properties, this work shows the effectiveness of an iterative approach to training deep learning tools in presence of poor training samples, due to the paucity of confirmed UDG examples, which we expect to replicate for upcoming all-sky surveys like Rubin Observatory, Euclid and the China Space Station Telescope.