Compared to stellar-mass black holes ($\lt100 M_{\odot}$) and supermassive black holes ($\gt10^5M_{\odot}$), intermediate-mass black holes (100–$10^5 M_{\odot}$) have been little known in observations. Globular clusters are a possible site for a few thousand solar mass intermediate-mass black holes to exist. We performed globular cluster formation simulations using a smoothed-particle hydrodynamics simulation code combined with an N-body simulation code, which can integrate stars without gravitational softening. We found that 3–5 % of cluster stellar mass joined the runaway collisions of stars and formed very massive stars with a maximum mass of $\sim 10^4 M_{\odot}$. Such very massive stars evolve to a few thousand solar-mass intermediate-mass black holes.

This short paper presents the activities of PARSEC, a research project that aims to develop and apply a machine-assisted methodology for constructing and curating semantically enriched astrophysical catalogues in the form of knowledge graphs (KGs). The ultimate objective of this methodology is to automate the process of constructing astrophysical catalogues and to simplify their analysis and integration, allowing astrophysicists to more easily generate new knowledge and scientific insights.

We present a deep learning-based framework for automated quasar spectral analysis, addressing two key challenges: continuum estimation and absorption feature detection. Using autoencoders, convolutional neural networks (CNNs), and U-Net architectures, we demonstrate improved performance in reconstructing the quasar continuum and identifying absorption lines. These models are trained to learn complex, non-linear correlations across wide spectral regions, enabling robust predictions even in the presence of noise or missing data. Our methods offer greater flexibility and accuracy, particularly in low signal-to-noise regimes.

The U-Net architecture, in particular, shows strong potential for precise absorption feature detection, leveraging its encoder-decoder structure with skip connections to capture both local and global spectral patterns. This enables the identification of absorption signatures associated with metal-enriched gas in the circumgalactic medium. Our results highlight the scalability of these techniques for application to large-scale surveys such as DESI and WEAVE, making progress for new insights into the distribution and dynamics of baryons in the Universe.

Photometric redshift estimation (photo-z) offers a scalable alternative to spectroscopic measurements, at the expense of reduced accuracy and increased uncertainty. In this work, we investigate machine learning approaches that explicitly incorporate uncertainty quantification into photo-z estimation using Sloan Digital Sky Survey (SDSS) data. Starting from a baseline neural network, we extend to Gaussian regression, quantile regression, and conformalized quantile regression (CQR). Our experiments demonstrate that, while the baseline model achieves competitive accuracy, uncertainty-aware methods provide more reliable and interpretable predictions.

We present GS$^{3}$ Hunter (Galactic-Seismology Substructures and Streams Hunter), a novel deep-learning method that combines Siamese Neural Networks and K-means clustering to identify substructures and streams in stellar kinematic data. Applied to Gaia EDR3 and GALAH DR3, it recovers known groups (e.g., Thamnos, Helmi, GSE, Sequoia) and, with DESI dataset, reveals that GSE consists of four distinct components (GSH-GSH1 through GSE-GSH4), implying a multi-event accretion origin. Tests on LAMOST K-giants recover Sagittarius, Hercules-Aquila, and Virgo Overdensity, while also uncovering new substructures. Validation with FIRE simulations shows good agreement with previous results. GS$^{3}$ Hunter thus offers a powerful tool to understand the Milky Way’s halo assembly and tidal history.

Information field theory (IFT) is the application of probabilistic reasoning to fields. Physical fields are mathematical functions over continuous spaces that exhibit certain properties of regularity, such as limited variance and finite gradients. Inferring a field from an observational dataset should exploit these regularities. However, the finite number of constraints that the data provides is insufficient to determine the infinite number of degrees of freedom of a field. IFT enables us to derive optimal field inference algorithms that explicitly exploit domain knowledge. These algorithms can be implemented via Numerical Information Field Theory (NIFTy). In NIFTy, neural operator forward models can be written and inverted probabilistically. NIFTy thereby infers fields and their remaining uncertainties. This is achieved using novel variational inference schemes that scale quasi-linearly, even for ultra-high dimensional problems. This paper introduces the basic concepts of IFT and NIFTy, highlights a few of their astrophysical applications, and discusses their artificial intelligence (AI) perspective. Finally, UBIK (the Universal Bayesian Imaging Kit), an emerging customisation of NIFTy for a suite of astrophysical telescopes, is presented as a central tool to the topic of the UniversAI conference.

We present ProMage, a feed-forward neural network that emulates the computation of observer- and rest-frame magnitudes from the SED generation package ProSpect. The network predicts magnitudes conditioned on input galaxy physical properties, including redshift, star formation history, gas and dust parameters. ProMage accelerates magnitude computation by a factor of $10^4$ compared to ProSpect, while achieving per-mille relative accuracy for 99% of sources in the test set across the g,r,i,z,y Hyper Suprime-Cam bands. This acceleration is key to enabling fast inference of galaxy physical properties in next-generation Stage IV surveys and to generating large catalogue realisations in forward-modelling frameworks such as GalSBI-SPS.

The proliferation of data has led to a considerable challenge in identifying pertinent subjects for study, given the vast amount of information available. The identification of outliers in a given dataset can facilitate the direction of investigations towards a more limited group of objects that may offer insights as they deviate from the trends exhibited by the underlying population. Here, we propose a methodology for the identification of outliers in voluminous datasets. Our approach is founded on the utilisation of k-nearest neighbours and spectral clustering algorithms, which are well-established machine learning algorithms that have been employed for a range of applications, including outlier detection. These algorithms are implemented through the memory-distributed tensor framework library developed by Helmholtz Analytics (Heat). We have applied our methodology to a catalogue of X-ray sources, namely the Chandra Source Catalogue. We will present and discuss the results of our study.

AI is reshaping society and science. Astrophysics is, and has been since the 1980s, at the bleeding edge of AI adoption due to the size and complexity of its data. The upcoming Vera C. Rubin Observatory Legacy Survey of Space and Time will scan an unprecedented volume of the universe: 20 orders of magnitude in distance scales, 60 orders of magnitude in energy scales, 20Tb of data every night, opening regions of the parameter space unexplored until now. With bigger than ever data come big challenges, and the astrophysical community has risen up to the challenge. Scientists around the world have developed innovative data analysis solutions. But what challenges remain open at the dawn of LSST? And as we continue to embrace AI in astrophysics, can we be thoughtful and conscientious enough to shape our communities to realize the promise of AI, without contributing to the threats that it poses?

The Gaia mission has observed over 2 billion stars repeatedly across the entire sky over 10 years, revealing the many astronomical objects that vary on human timescales from seconds to years. Its repeated astrometric, photometric, spectrophotometric and spectroscopic measurements create an unprecedented dataset to probe the variable celestial sources down to $G\approx 21$ mag. To extract meaningful results from these many time series for so many sources, we have used machine learning techniques for crossmatching, variability detection, and variability classification. This approach has now led to the largest catalogue of classified variable sources ever produced over the entire celestial sphere.

We investigate the impact of machine learning algorithm combinations on the identification of Gaia–Enceladus (GSE) stellar populations in the solar neighborhood. Using high-quality chemical abundances from APOGEE DR17 and precise astrometry from Gaia, we apply unsupervised learning pipelines combining PCA or t-SNE for dimensionality reduction with K-means, GMM, or HDBSCAN clustering algorithms. Our aim is to assess consistency and performance in selecting the GSE component. The optimal cluster is then used to estimate the initial gas mass of GSE through a one-zone chemical evolution model. We find that the t-SNE + HDBSCAN pipeline yields the most chemically and dynamically coherent identification, and our best-fit chemical evolution model predicts an initial gas mass of $4.93^{+0.32}_{-0.72} \times 10^9~M_\odot$ for the progenitor.

We combine APOGEE DR17 spectroscopy with Gaia DR3 astrometry to investigate the chemo-dynamical structure of the Milky Way using unsupervised machine learning. After applying strict quality cuts, we selected a high-quality sample of 135,080 giant stars and derived their orbital parameters in the McMillan potential. To mitigate disk dominance, a cut of $e_{p}$ $>$ 0.25 was applied. Dimensionality reduction with PCA and clustering with K-means were tested across three input spaces: chemical abundances, kinematic parameters, and their combination. The most robust separation of stellar populations is obtained when using chemical abundances alone. The resulting clusters correspond to the thin disk, thick disk, halo, and an accreted component, as revealed in chemical-abundance and Lindbland diagrams. Our results highlight the effectiveness of abundance-driven clustering in disentangling Galactic stellar populations and tracing ex-situ stars, providing a data-driven framework for future studies of Galactic archeology.

Modern astronomical surveys such as the Sloan Digital Sky Survey (SDSS) provide extensive astronomical databases enabling researchers to access vast amount of diverse data. However, retrieving data from archives requires knowledge of query languages and familiarity with their schema, which presents a barrier for non-experts. This work investigates the use of Microsoft Phi-2, a compact yet powerful transformer-based language model, fine-tuned on natural language–SQL pairs constructed from SDSS query examples. We develop an interface that translates user queries in natural language into SQL commands compatible with SDSS SkyServer. Preliminary evaluation shows that the fine-tuned model produces syntactically valid and largely semantically correct queries across a variety of astronomy-related requests. Our results show that even small-scale models, when carefully fine-tuned, can provide effective domain-specific natural language interfaces for large scientific databases.

We present TUNA, a Vision-Transformer based network adapted from segmentation to flux regression for faint, diffuse radio emission. Trained on LOFAR-like mock observations derived from cosmological simulations, TUNA accurately reconstructs low surface-brightness structures, with only mild smoothing and small brightness-dependent biases. Applied to LOFAR data of the A399-A401 galaxy cluster system, it recovers the ridge not identifiable in the high resolution observation and matches the low resolution tapered map. These results indicate how TUNA can deliver automated, quantitative surface brightness estimates for diffuse extragalactic sources, enabling scalable analyses for upcoming surveys.

Interpreting the spectral energy distributions (SEDs) of astrophysical objects with physically motivated models is computationally expensive. These models require solving coupled differential equations in high-dimensional parameter spaces, making traditional fitting techniques such as Markov Chain Monte Carlo or nested sampling prohibitive. A key example is modeling non-thermal emission from blazar jets–relativistic outflows from supermassive black holes in Active Galactic Nuclei that are among the most powerful emitters in the Universe. To address this challenge, we employ machine learning to accelerate SED evaluations, enabling efficient Bayesian inference. We generate a large sample of lepto-hadronic blazar emission models and train a neural network (NN) to predict the photon spectrum with strongly reduced run time while preserving accuracy. As a proof of concept, we present an NN-based tool for blazar SED modeling, laying the groundwork for future extensions and for providing an open-access resource for the astrophysics community.

The intracluster medium (ICM) records the history of galaxy clusters through its complex dynamical properties. To effectively interpret these properties, robust methods are needed to compare observational data with theoretical models. We present a novel machine learning framework for comparing ICM line-of-sight velocity maps derived from X-ray observations. Our approach uses convolutional and Siamese neural networks to identify similarities between different kinematic fields. We outline the architecture of this framework and perform a series of sanity checks to validate its performance. These checks demonstrate the model’s ability to correctly identify and quantify kinematic features, establishing a powerful new tool for future comparative studies of the ICM.

Understanding the variability of Sagittarius A* (Sgr A*) requires coordinated, multi-wavelength observations. However, these observations suffer from irregular sampling, noise, and large data gaps, limiting our ability to identify flares and infer cross-wavelength time lags. We address this challenge using machine learning models designed for sparse, multivariate time series. In particular, we introduce a diffusion-based generative model, representing the first application of score-based diffusion models to astronomical time series. We also use a transformer-based model, and benchmark both against a multi-output Gaussian Process. The models are trained on simulated light curves that replicate the cadence, noise, and cross-band correlations of real Sgr A* data, including observations from ALMA, Spitzer, GRAVITY, and Chandra. Our results show that the diffusion model outperforms other methods in accuracy and uncertainty calibration, on both synthetic and real data. This work highlights the potential of generative models for reconstructing astrophysical variability from incomplete observations.

Weak gravitational lensing of background galaxies traces the growth of dark matter and is key to constraining the cosmological model. This proceedings paper provides an overview of the analysis presented in Voigt (2024), applying convolutional neural networks (CNNs) to the challenge of precision shear estimation. Using a shallow CNN, we investigate the sensitivity of shear recovery to point spread function (PSF) misestimation, with reference to the accuracy requirements of Euclid. To mitigate noise bias we employ a non-standard loss function, and assess performance assuming a linear relation between predicted and true shears.

Efforts are underway to measure the global 21 cm signal from neutral hydrogen, which is a powerful probe of the early universe, using NASA radio telescopes on the far side of the Moon. Physics-based models of the signal are computationally expensive to perform Bayesian multi-parameter inferences, for which we have developed novel, publicly-available neural network emulators utilizing a Long Short-Term Memory (LSTM) network and a Kolmogorov-Arnold Network (KAN). <sc>21cmLSTM</sc> is currently the most accurate emulator in the community by leveraging the signal’s temporally-correlated structure, and <sc>21cmKAN</sc> maintains similar accuracy while training 75 times faster, by learning expressive functional transformations. Each emulator can fit realistic mock signals and obtain unbiased physical parameter constraints, with <sc>21cmKAN</sc> able to complete end-to-end training and inference in under 30 minutes. The implementation of machine learning tools like these in data analysis pipelines is important to fully exploit upcoming measurements of the cosmological 21 cm signal.

Machine learning (ML) is revolutionising the analysis of large astronomical surveys such as Gaia DR3. We present a two-step ML pipeline to classify Asymptotic Giant Branch (AGB) stars: (i) cosine similarity identifies candidate AGB stars from 140 million sources, and (ii) an XGBoost classifier which separates carbon-rich (C-rich) and oxygen-rich (O-rich) stars. Our classifier achieves ROC-AUC > 99.9%, producing a catalogue of 307,191 O-rich and 4,227 C-rich stars. Feature importance analysis shows that the sixth RP coefficient (RP6) dominates classification, underlining the critical role of Gaia red spectral features. This work demonstrates how hybrid ML approaches can efficiently identify chemically distinct stellar populations in massive survey datasets, providing opportunities for new studies of Galactic structure and stellar evolution.