AI and Astronomy Weekly Roundup - Week of May 5 - 9, 2025
Help us pick a better name?
First things first, I am not too keen on the name of this newsletter. I am open to suggestions. Please vote on your favorites or suggest a new one:
Last week’s batch of machine learning and methods papers in astronomy showcases a mix of supervised, unsupervised, and generative approaches applied across the cosmic landscape. With a total of 27 papers, this is the longest list so far this year. In cosmology, diffusion models and GANs are making inroads in emulating cosmic structures and supernova spectra, while hybrid techniques like Gaussian processes combined with neural networks tackle dark sector physics. In stellar and galactic contexts, RNNs and symbolic regression are used for classifying heartbeat stars and parameterizing star formation, respectively, and multiple teams continue to refine galaxy morphology pipelines and quasar selection with active learning. Anomaly detection remains a hot topic, from searching for techno-signatures to sifting through millions of Hubble cutouts. Instrumental methods feature prominently this week, with nine contributions, featuring new software for redshift estimation, CMD analysis, image reprojection, and telescope array optimization. Finally, solar flare prediction and gravitational wave signal disentangling round out a diverse set of contributions that reflect the growing integration of AI tools into many corners of astrophysics.
The full bib file is available here.
🪐 Exoplanets
Using anomaly detection to search for technosignatures in Breakthrough Listen observations, Pardo et al., arxiv:2505.03927
Unsupervised learning — Applies anomaly detection techniques to radio observations to identify potential techno-signatures. Spoiler (from the abstract): “no candidate survived basic scrutiny“.
☀️ Solar Physics
Solar Flare Forecast: A Comparative Analysis of Machine Learning Algorithms for Solar Flare Class Prediction, Bringewald, arxiv:2505.03385
Supervised learning — Compares Random Forest, k-Nearest Neighbors (KNN), and Extreme Gradient Boosting (XGBoost) for predicting flare classes. “Our findings indicate that Random Forest and XGBoost consistently demonstrate strong performance across all metrics, benefiting significantly from increased dimensionality.“
⭐ Stars and Stellar Remnants
Gaussian process representation of dispersion measure noise in pulsar wideband datasets, Susobhanan & van Haasteren, arxiv:2505.05274
Bayesian time series model — Uses Gaussian processes to model non-stationary dispersion effects.DeepHMC: a deep-neural-network acclerated Hamiltonian Monte Carlo algorithm for binary neutron star parameter estimation, Perret et al., arxiv:2505.02589
Accelerated Bayesian inference — Uses neural networks to speed up Hamiltonian Monte Carlo sampling.A Catalog of 12,766 Carbon-enhanced Metal-poor Stars from LAMOST Data Release 8, Fang et al., arxiv:2505.05024
Supervised deep learning — Utilizes a deep neural network trained on labeled spectra to estimate stellar parameters (Teff, log g, [Fe/H], [C/H]) from LAMOST DR8 data, enabling the identification of CEMP stars based on these derived parameters.Heartbeat Stars Recognition Based on Recurrent Neural Networks: Method and Validation, Li et al., arxiv:2505.04067
Recurrent neural networks (RNNs) — Classifies light curves to detect heartbeat stars in large photometric surveys.
⚡ High-Energy Astrophysics
A robust neural determination of the source-count distribution of the Fermi-LAT sky at high latitudes, Eckner et al., arxiv:2505.02906
Bayesian neural networks — Infers source-count distributions using a probabilistic deep learning approach.
🌀 Galaxies, Gravitational Lenses & QSOs
Gravitational Lenses in UNIONS and Euclid (GLUE) I: A Search for Strong Gravitational Lenses in UNIONS, Storfer et al., arxiv:2505.05032
Supervised deep learning (ResNet-based) — Utilizes a deep residual neural network, adapted from the CMU-Deeplens architecture, trained on images of real strong gravitational lenses to detect lens candidates in multi-band imaging data from the UNIONS survey. The model was applied to approximately 8 million galaxies with full coverage in the g, r, and i filters, resulting in the identification of 1,346 new strong lens candidates, which were subsequently graded through human inspection.
.
Machine Learning Workflow for Morphological Classification of Galaxies, Doser et al., arxiv:2505.04676
Supervised learning — End-to-end pipeline using CNNs trained on labeled morphology datasets.A data-driven approach for star formation parameterization using symbolic regression, Salim et al., arxiv:2505.04681
Symbolic regression — Learns interpretable functional forms for star formation relations in simulations.Classifying Radio-Loud and Radio-Quiet Quasars With Novel PCA Based Regression Classifier, Joshi & Shinde, arxiv:2505.01335
Supervised learning + PCA — Uses dimensionality reduction and regression classifiers to separate AGN types.Using Active Learning to Improve Quasar Identification for the DESI Spectra Processing Pipeline, Green et al., arxiv:2505.01596
Active learning with convolutional neural networks (CNNs) — Enhances the QuasarNET CNN classifier by implementing an active learning strategy to select the most informative spectra for labeling, incorporating an outlier rejection step using Self-Organizing Maps (SOMs) to ensure representative training data, resulting in improved classification accuracy with a smaller labeled dataset.
🌌 Cosmology
nuGAN: Generative Adversarial Emulator for Cosmic Web with Neutrinos, Kaushal et al., arxiv:2505.03936
Generative adversarial networks (GANs) — Emulates cosmic web structure including neutrino effects using GANs.In search of an interaction in the dark sector through Gaussian Process and ANN approaches, Abedin et al., arxiv:2505.04336
Hybrid (Gaussian Processes + ANN) — Uses both GPs and artificial neural networks to probe dark sector interactions.Stochastic analysis of finite-temperature effects on cosmological parameters by artificial neural networks, Hatefi et al., arxiv:2505.02223
Supervised learning — Models the impact of thermal fluctuations on cosmological parameters with ANNs.Variational diffusion transformers for conditional sampling of supernovae spectra, Shen & Gagliano, arxiv:2505.03063
Generative diffusion model — Uses diffusion-based transformers for conditional generation of supernova spectra.JFlow: Model-Independent Spherical Jeans Analysis using Equivariant Continuous Normalizing Flows, Lim et al., arxiv:2505.00763
Normalizing flows (unsupervised density estimation) — Introduces an equivariant normalizing flow model to solve the spherical Jeans equations for dynamical modeling, enabling flexible, non-parametric inference of galaxy mass profiles without assuming a fixed functional form.
🌊 Gravitational Waves
Can Transformers help us perform parameter estimation of overlapping signals in gravitational wave detectors?, Papalini et al., arxiv:2505.02773
Transformer networks — Applies transformer models to disentangle overlapping signals in GW data.
🛠️ Methods and Instruments
An Unsupervised Learning Method for Radio Interferometry Deconvolution, Yu et al., arxiv:2505.04887
Unsupervised learning — Introduces a deconvolution method using neural networks trained without ground truth.dfreproject: A Python package for astronomical reprojection, Rhea et al., arxiv:2505.03932
Software tool — Presents a utility for consistent reprojection of astronomical images. Builds on the functionality of the existingastropy.reprojectpackage but speeds up some of the computationally-intensive functions withpytorch.PANCAKE: Python bAsed Numerical Color-magnitude-diagram Analysis pacKagE, Zheng et al., arxiv:2505.04534
Bayesian inference — CMD fitting tool with MCMC and nested sampling for stellar population studies.
AnomalyMatch: Discovering Rare Objects of Interest with Semi-supervised and Active Learning, Gómez & O’Ryan, arxiv:2505.03509
Semi-supervised + active learning — Combines clustering and user-guided exploration to find anomalies in survey data.Identifying Astrophysical Anomalies in 99.6 Million Cutouts from the Hubble Legacy Archive Using AnomalyMatch, O’Ryan & Gómez, arxiv:2505.03508
Anomaly detection — Scales up AnomalyMatch to the full Hubble cutout archive for anomaly mining. This is an application of the method described in the previous paper to the Hubble image archive. Anomalies here are rare types of objects and the objects found fall into known astrophysical classes.
Redshift Assessment Infrastructure Layers (RAIL): Rubin-era photometric redshift stress-testing and at-scale production, RAIL Team et al., arxiv:2505.02928
Meta-framework — Infrastructure for testing and deploying photometric redshift estimation algorithms at scale. One of two papers this week on redshift measurements, following on the one last week: improvements in redshift estimation is a very active area of work.RVSNUpy: A Python Package for Spectroscopic Redshift Measurement Based on Cross-Correlation, Kim et al., arxiv:2505.01710
Template fitting — Implements a classical cross-correlation approach in Python for redshift estimation.An Optimization Framework for Wide-Field Small Aperture Telescope Arrays Used in Sky Surveys, Xiang et al., arxiv:2505.03094
Optimization and simulation — Evaluates survey strategies using statistical models and design constraints.Power Laws Associated with Self-Organized Criticality: A Comparison of Empirical Data with Model Predictions
Markus J. Aschwanden, Felix Scholkmann, arxiv:2505.00748
Statistical modeling — Compares observed power-law distributions in astrophysical systems to predictions from self-organized criticality (SOC) models, focusing on empirical consistency. I flagged this paper because I was curious to find out what SOC is. Self-organized criticality (SOC) is a concept describing how complex systems naturally evolve into a critical state where minor events can lead to significant, scale-invariant outcomes. In this state, systems exhibit power-law distributions in event sizes, durations, or energies, indicating that small and large events follow the same statistical patterns. “In the astrophysics section in Table 1 we are quoting statistics of solar flare peak fluxes, but the universality of SOC systems has also been identified in stellar flares (Aschwanden and G¨udel 2021), and in galactic, extragalactic, and blackhole systems (Aschwanden and Gogus 2025). What is observed in astrophysical SOC systems is generally photon fluxes in various wavelength regimes (gamma rays, hard X-rays, soft X-rays, extreme-ultraviolet (EUV), ultra-violet (UV), and white-light).“ These seem to fit the distribution expected by SOC. Lunar craters, on the other hand, do not.
