Awesome AI for Astronomy

Review of classical and machine learning source finding and morphology classification for SKAO surveys.

Chapter on deep learning for source finding, RFI mitigation, imaging, anomaly detection and scheduling for the SKA.

Review of machine learning for transit identification, candidate vetting, detrending and atmospheric retrieval, including neural and flow matching posterior estimation, in the context of JWST and Ariel.

Review of deep learning methods that fuse images, spectra, time series and catalogues in astronomy.

LSST DESC roadmap covering photometric redshifts, transient classification, weak lensing inference, foundation models and infrastructure.

Review of statistical, physics-based and deep learning flare forecasting including CNN, LSTM and transformer models.

Annual Review covering inductive biases, simulation-based inference, anomaly detection, foundation models, reinforcement learning and LLM agents in astronomy.

Domain-oriented survey of autonomous scientific discovery including agentic astronomy and cosmology research.

Survey of scientific LLMs and their training data, including astronomy datasets and models.

Review of how machine learning, citizen science and expert inspection found about 500 high-quality lens candidates in Euclid Q1.

Survey of LLM-based scientific agents with a section on astronomy and cosmology applications.

Review of deep learning methods for Imaging Atmospheric Cherenkov Telescope data, covering gamma-hadron separation, event reconstruction and fast simulation.

Use cases, survey results and recommendations on LLM adoption in astronomy research.

Recommendations on baselines, validation, reproducibility and reporting for machine learning papers in astronomy.

Review of CNN, U-Net, R-CNN and transformer crater detection algorithms for the Moon, Mars and Mercury.

Review of machine learning applications in time-domain astronomy, galaxy surveys and emulation for cosmology.

Historical review of neural networks in astronomy with a primer on deep learning and an outlook on astronomical foundation models.

Review of deep learning applications to galaxy surveys covering classification, property estimation, discovery and cosmological inference.

Review of machine learning for LIGO and Virgo data quality, signal modeling, searches and parameter estimation.

Review of simulation-based inference methods and the role of machine learning in inference with intractable likelihoods.

Lecture notes from the IAC Winter School 2018 covering supervised and unsupervised learning for astronomical datasets.

Combines frozen embeddings of 22 foundation models from eight families by whitening and randomised SVD, outperforming every member on 19 of 21 metrics for HSC, JWST and Legacy Survey galaxies.

Autoregressive transformers from 1M to 2.1B parameters pretrained on 8.6M DESI Legacy Survey galaxy images.

Supervised scaling study on 840k galaxy images with over 100M Galaxy Zoo annotations.

Searches Zoobot representation space for rare morphological classes.

Pretrained galaxy morphology encoders and finetuning utilities, deployed in the Euclid morphology pipeline.

Pretraining galaxy morphology encoders on Galaxy Zoo volunteer responses for transfer to downstream tasks.

Self-supervised representations of 42M DESI Legacy Imaging Survey galaxies with a public similarity search tool.

Representations learned from Galaxy Zoo labels reused for similarity search, anomaly finding and new classification tasks.

Contrastive pretraining on about 1.2M SDSS galaxy images, evaluated on morphology classification and photometric redshift estimation.

Masked-modeling transformer pretrained on 5.5 million spectra from eight surveys at native resolution.

Transformer tokenizer that reads spectra on their native wavelength grids, pretrained on SDSS, DESI, GALAH and APOGEE and adapted to object classification and stellar parameter regression.

Contrastive alignment of LAMOST low-resolution and Gaia XP spectra foundation models for parameter estimation and cross-spectrum prediction.

Converts one-dimensional spectra into images for pretrained ViT backbones in classification and regression.

Transformer pretrained on synthetic stellar spectra and transferred to real spectra across wavelength ranges.

Transformer trained on cross-survey APOGEE, Gaia and 2MASS data that performs discriminative and generative stellar tasks without task-specific finetuning.

Attentive convolutional autoencoder that models galaxy spectra in the rest frame with explicit redshifting and resampling, trained on SDSS-II.

Domain adaptation between synthetic and observed stellar spectra with cycle-consistent networks.

Variational autoencoder compression of SDSS galaxy spectra into a low-dimensional latent space.

Measures the information that self-supervised light-curve encoders add beyond engineered periodogram and variability features.

Domain-informed JEPA combines raw, periodogram and phase-folded light-curve views with uncertainty-aware tokenization and multi-view self-distillation, trained on LEAVES and evaluated on StarEmbed and 12 irregular time-series datasets.

Self-supervised transformer on Kepler light curves for variability classification, surface gravity estimation and flare identification.

Contrastive alignment of supernova photometry and spectroscopy, pretrained on 0.5M synthetic events and finetuned on ZTF.

Conformer model for Kepler light curves that infers surface gravity and frequency of maximum oscillation power.

Self-supervised light-curve encoder pretrained with masked magnitude prediction on MACHO R-band data.

Joint representation of F150W imaging and 1 to 2 micron PRISM spectra that recovers stellar mass at 1 < z < 3 with R2 of about 0.9, with CLIP alignment improving single-modality inference.

Token-based masked modeling over 39 modalities from Legacy Survey, HSC, SDSS, DESI and Gaia, released at 300M to 3.1B parameters.

Trimodal CLIP-style pretraining on variable-star light curves, spectra and metadata.

CLIP finetuned on Hubble Space Telescope observations paired with observing proposal abstracts.

Contrastive alignment of DESI galaxy images and spectra in a shared embedding space.

Autoregressive pretraining of a single surrogate model across heterogeneous physical simulation systems.

Evaluates whether continued pretraining and instruction tuning on astronomy corpora help open-ended scientific reasoning.

Tool-using vision-language agent trained with supervised fine-tuning on expert inspection trajectories and outcome-based reinforcement learning, which vets rare-object candidates by iteratively inspecting spectral regions, raising macro-F1 on five LAMOST tasks from 28.3 to 76.5 with transfer to SDSS and DESI.

Llama-3.1-70B with continued pretraining and reasoning-chain SFT. Scores 89.0% on a validated 3,846-question subset of AstroMLab 1.

LLaVA finetuned on about 30k captioned images from APOD, ESO and Hubble for astronomical visual question answering.

Llama-3.1-8B with continued pretraining on about 250k astro-ph and gr-qc preprints, SFT and model merging. Scores 80.9% on AstroMLab 1.

BERT model adapted to astrophysics and high energy physics literature for concept analysis.

Releases AstroLLaMA-3-8B and AstroLLaMA-2-70B and reports that the 7B AstroLLaMA series underperforms its base model.

Cosmology assistant built by continued pretraining on cosmology papers and textbooks followed by question-answer finetuning.

Continual pretraining on abstracts, introductions and conclusions followed by domain conversational finetuning.

LLaMA-2-7B finetuned on over 300k astro-ph abstracts.

BERT model with 110M parameters trained on NASA ADS papers for named entity recognition and classification.

GPT-5.5 agents extract source-linked discussions from arXiv papers and coordinate matching groups the evidence into 13,616 galaxy entries.

Multi-agent pipeline that retrieves HEASARC data and writes its own tools, tested by reproducing IXPE polarimetry results.

Cmbagent multi-agent workflow with human intervention that builds a weak lensing inference pipeline and took first place in the FAIR Universe Weak Lensing Uncertainty Challenge.

Multi-agent tree search over CIGALE SED models with a self-updating knowledge base, applied to JWST Little Red Dots.

Modular multi-agent research assistant for idea generation, analysis and paper drafting, using cmbagent as its analysis backend.

Agents that generate and validate CTAO control software and Gammapy analysis scripts with automated self-correction.

Agents for planning, coding, execution, analysis and synthesis that run machine learning experiments on cosmological datasets.

Multi-agent system that proposes astrobiology hypotheses from mass spectrometry data.

LLM agent prototypes for CTAO control software and Gammapy analysis code generation with retrieval augmentation.

Planning and control system of about 30 LLM agents for autonomous cosmological data analysis.

LLM agent that plans observations, controls a 10-telescope network and runs transient detection for the Nearby Galaxy Supernovae Survey.

Retrieval-augmented literature search and synthesis over about 350k ADS papers with time and citation weighting.

Simulated HSC-like weak lensing convergence maps and a competition scoring cosmological parameter estimates together with their uncertainty calibration.

Fifteen-year flare forecasting dataset from 2011 to 2025 built from SDO/HMI magnetograms, NOAA region summaries and flare lists, with active region level splits and configurable horizons.

Lunar remote sensing foundation model released with a benchmark of six downstream tasks.

Instance retrieval benchmark of about 25,000 Mars CTX crater identities with verified multi-view queries, used to evaluate 30 vision backbones.

Eight-year dataset of 429,620 all-sky images with star-aware cloud masks and per-pixel altitude-azimuth calibration, supporting DINOv3-based segmentation and short-horizon cloud nowcasting benchmarks for observatory operations.

Seven simulated neutrino telescope datasets used to compare ParticleNeT, DynEdge, GRIT and DeepIce on energy, direction, topology, vertex and inelasticity reconstruction.

20 orbital and rover image datasets for classification, segmentation and object detection used to compare natural-image, Earth observation and vision-language backbones.

Paired dataset of 134,533 HSC galaxy images and DESI spectra with a masked autoencoder baseline for cross-modal reconstruction and redshift estimation.

Curated SDO core dataset with application benchmarks on active region emergence, flares, magnetic field extrapolation, EUV irradiance and solar wind prediction.

About 34,000 point clouds from Quijote, CAMELS-SAM and CAMELS together with 25,000 merger trees for cosmological parameter inference and tree tasks.

2,800 radio and infrared images with 4,155 expert-annotated galaxies in COCO format for multi-component radio galaxy and host detection.

1,256 NVSS and FIRST radio-loud AGN labelled by Fanaroff-Riley class with confidence flags, distributed as a batched dataset for deep learning libraries.

Detection and characterisation of 233,245 simulated HI sources in a 2,000 h SKA-MID spectral line cube, with reproducibility awards for open pipelines.

Common datasets and metrics for comparing machine learning gravitational-wave search algorithms.

Open challenge in which participants classified 100,000 simulated space-based and ground-based cutouts as lenses or non-lenses, comparing CNNs, SVMs, arc finders and visual inspection.

Evaluates 13 multimodal LLMs on transient alert review through metadata grounding, scientific reasoning and hierarchical classification over five classes.

More than 4,100 expert-verified instances over five tasks covering HSC imaging, FIRST and NVSS radio maps, AKARI photometry, PLAsTiCC light curves and optical spectra.

Benchmark of vision-language models on retrieval over Mars orbital imagery at planetary scale.

621 expert-reviewed multiple-choice questions on astronomical images across six subfields, used to evaluate 25 MLLMs.

AstroChart provides 1,690 question-answer pairs over 482 astronomy charts and is used to evaluate 21 MLLMs on perception, numerical reasoning and domain knowledge.

500 expert-reviewed binary questions from 346 astronomy papers asking which machine learning method performs better for a stated problem.

Reproduction framework that separates execution from verification, applied to one ApJ and thirteen Nature astronomy papers, eleven of which leave the reproduction path ambiguous.

Evaluates thirteen LLMs on translating natural-language requests into SQL queries over the ALeRCE alert broker database, comparing direct and step-by-step generation.

Controlled study of 144 synthetic researchers on 2,592 astrophysics assignments, comparing solo work with four styles of AI assistance over 12,960 scored episodes.

Evaluates five LLMs on 49 theory and 8 data analysis problems from the 2022 to 2025 IOAA exams.

111 tasks derived from 20 astrophysics papers that test whether agents can reproduce published results, with the best models scoring about 20 percent.

105 expert-written question-answer pairs from cosmology papers for comparing retrieval-augmented generation agents against human evaluation.

864 processing and visualization tasks from 110 NOIRLab and STScI notebooks, scored by execution and a VLM judge.

3,082 English and Chinese questions of six types across astrophysics, astrometry, celestial mechanics, history and techniques, evaluated on 27 LLMs.

4,425 multiple-choice questions curated from the Annual Review of Astronomy and Astrophysics for evaluating LLM knowledge recall.

Distils a general-purpose VLM answering Galaxy Zoo decision tree questions into Zoobot through side tuning, with teacher uncertainty guiding which galaxies are sent for human labels.

Euclid Q1 morphology study listed among Zoobot applications.

First Euclid visual morphology catalogue produced with Galaxy Zoo volunteers and Zoobot.

Bar fraction measurement in Euclid Q1 using machine learning morphology.

Dwarf galaxy census in Euclid Q1 listed among Zoobot applications.

Uses Galaxy Zoo DESI deep learning morphologies to study bars and AGN.

Uses deep learning bar classifications to study the link between bar strength and quenching.

Classification of faint tidal features in DECaLS images using pretrained morphology models.

Zoobot trained to predict detailed morphology for Euclid imaging.

Object detection of star-forming clumps in galaxy images with transfer learning.

Hybrid transformer and convolutional classifier trained in the low-data regime on Galaxy10 DECals.

Zoobot-based search of the HST archive for interacting galaxies.

Universal domain adaptation between SDSS, DECaLS and LSST mock images that also clusters unknown classes.

Compares deep learning architectures for optical galaxy morphology classification using the Zoobot library.

7.5M volunteer classifications and Bayesian CNN posteriors for 314k DECaLS galaxies.

CNN pretrained on simulated galaxies and transferred to SDSS and CANDELS bulge-to-total classification.

DINOv2 light-curve embeddings, isolation forests and multimodal LLM agents that vet anomaly candidates.

BYOL-based embedding pipeline applied to ALMA protoplanetary disks and infrared dark clouds.

Sparse autoencoders on Zoobot and masked-autoencoder embeddings of Euclid Q1 galaxies to surface interpretable features.

FixMatch with EfficientNet and a user-in-the-loop active learning interface, integrated into ESA Datalabs.

Applies AnomalyMatch to the full Hubble Legacy Archive and reports 1,176 new anomalies across 19 classes.

Isolation-forest based active anomaly detection with expert feedback, used by the SNAD team on ZTF data.

Self-supervised features with clustering, artefact removal, anomaly detection and similarity search on survey images.

Applies Astronomaly to about 4M galaxies using pretrained Zoobot representations.

Studies the robustness of galaxy morphology classifiers to image perturbations.

Wasserstein GAN anomaly scores for HSC galaxy images.

Active learning framework that combines anomaly scores with user feedback to rank interesting objects.

Unsupervised ordering algorithm that recovers one-dimensional trends in spectra and other datasets.

Unsupervised random forest outlier scores for about two million SDSS galaxy spectra.

Semi-supervised network that combines contrastive, color and redshift losses on CANDELS images.

Multimodal network with cross-attention and Bayesian layers that integrates SED templates on HSC PDR3 data.

Self-organizing map analysis showing lower scatter and attenuation bias for image-based photo-z on SDSS.

LSST DESC framework for training, evaluating and running photometric redshift estimators.

Inception CNN on four-band images combined with an MLP on nine-band magnitudes for 1.2 million KiDS-Bright galaxies.

Supervised contrastive learning with adaptive KNN to calibrate photometric redshift probability densities.

Diffusion model with contrastive networks that predicts optical spectra and photometric redshifts from SDSS images.

Normalizing flows for modeling galaxy catalogs and photometric redshift posteriors.

Neural posterior estimation for galaxy SED fitting that handles missing data and out-of-distribution noise.

Graph neural networks on dark matter merger trees that emulate semi-analytic galaxy properties.

Bayesian CNN that predicts posteriors of bulge-to-total ratio, effective radius and flux.

Amortized neural posterior estimation of 12 SED parameters in about one second per galaxy.

Bayesian SED modeling pipeline for DESI BGS that uses neural emulators for speed.

Neural network emulator of stellar population synthesis for fast SED fitting.

Gaussian process representation of galaxy star formation histories for SED fitting.

CNN that predicts redshift probability densities directly from multiband SDSS images.

Gaussian process photometric redshifts that combine templates with data-driven flux modeling.

Ensembles of neural networks and boosted decision trees for photometric redshift PDFs.

A convolutional autoencoder trained on HST COSMOS galaxy images supplies morphologies for shear inference on 10,000 simulated radio interferometric visibilities.

Iterative fine-tuning of a modified VGG16 from 27 seed lenses to 30,686 training images, yielding 441 grade A or B candidates in Euclid Q1, 130 of them new.

Mask R-CNN instance segmentation of cluster-scale lensed arcs trained on Euclidised HST simulations.

ResNet lens finder based on CMU-DeepLens applied to 8 million UNIONS galaxies.

Overview of the Euclid Q1 strong lens search combining volunteers and machine learning.

Machine learning lens finders applied to Euclid Q1.

Double-source-plane lens candidates from the Euclid Q1 search.

Ensemble classification of lens candidates and lessons for Euclid DR1.

Bayesian neural network that predicts lens mass and light parameters and initialises classical lens modelling.

Five CNN lens finders applied to 13 Euclid ERO fields, followed by visual inspection and spectroscopic confirmation.

Differentiable, GPU-accelerated lensing simulator for machine learning workflows.

Score-based generative priors for lensed source reconstruction.

Simulation-based inference of subhalo mass function parameters from many lenses.

GPU-accelerated lens modeling with gradient-based optimization, variational inference and HMC.

Self-supervised representations of DESI Legacy Survey images used to find new strong lens candidates.

Configuration-driven simulation of strong lensing training sets for deep learning.

Bayesian neural network lens modeling for time-delay cosmography.

Bayesian neural networks for population-level lens parameter inference.

Classifies strong lensing images by dark matter substructure type with deep learning.

Likelihood-ratio estimation with neural networks for subhalo population inference from lensed images.

U-Net mass mapping of Dark Energy Survey Science Verification weak lensing data.

Residual network strong lens finder.

Pixel-space flow matching for super-resolution that weights training by observational uncertainty and source regions, with a training-free Wiener-regularised correction against spurious sources; introduces 19,500 co-registered DESI to HST image pairs.

DDPM trained on JWST JADES imaging to generate Roman-like multi-band galaxy stamps for weak lensing.

Synthesises paired training data from a CCD noise model covering shot noise, response non-uniformity, dark current, readout effects, cosmic rays and hot pixels on stacked high-SNR bases; releases multi-band raw frames from two twin ground-based telescopes.

JAX scene modeling for multi-band, multi-epoch and multi-instrument data with neural network priors.

Stable Diffusion fine-tuned on Galaxy Zoo 2 morphology prompts to augment classifiers and find rare dusty early-type galaxies.

Class-conditioned diffusion model with residual attention blocks trained on Galaxy10 DECaLS categories.

DDPM trained on CANDELS images that generates Euclid-like galaxies for shear calibration.

Simulation and metrics framework for benchmarking detection and deblending algorithms.

Scalable Gaussian process classification of blended stellar images.

Differentiable wavefront-based PSF model that captures chromatic variations for Euclid-like telescopes.

Denoising diffusion models that generate realistic galaxy images.

Gaussian process classifiers that flag blended galaxies in simulated LSST images.

Scalable local Gaussian process classification of star and galaxy image cutouts.

Random forest star and galaxy separation for Pan-STARRS1 sources.

Gaussian process classification of stars and galaxies with dimensionality reduction.

Conditional generative models of galaxy light profiles integrated with GalSim.

VAE-based deblender trained jointly on ground and space-based simulated images.

Instance segmentation for joint detection, deblending and star-galaxy classification.

Constrained matrix factorization deblender for multiband survey images.

Fine-tunes a large language model on LAMOST low-resolution spectra in two stages to infer stellar parameters and elemental abundances, validated against Gaia-ESO and GALAH DR4.

BiLSTM, attention and CNN classifier applied to about 2.8 million TESS light curves.

Single transformer trained across APOGEE, GALAH, DESI and Gaia RVS that outputs atmospheric parameters, 20 abundances, distances and ages on one consistent scale.

CNN-LSTM models trained on Kepler red giants recover delta nu, nu max and period spacing from Gaia XP spectra and provide predictions for more than 2.5 million giants.

Extends a transformer stellar foundation model to infer mass and age of giants from Gaia XP spectra.

CNN that predicts detached eclipsing binary parameters from phase-folded TESS light curves to initialise JKTEBOP fits.

CVAE that generates stellar spectra over a grid of atmospheric parameters, rotation and instrumental resolution.

Neural spectral emulator of single stars used to identify more than 14 million binary candidates from XP spectra.

Uncertainty-aware cost-sensitive network that estimates metallicities down to [Fe/H] about -4 for giants.

APOGEE-trained network that corrects XP flux systematics before model-atmosphere fitting.

Heteroscedastic neural network ensemble that predicts [Fe/H] and [C/Fe] for more than 10 million XP stars.

LSTM and transformer model pretrained contrastively on Kepler light curves and trained on spot-model simulations.

CNNs that infer stellar rotation and surface gravity from Kepler light curves.

CNN flare detection in TESS two-minute light curves.

Bayesian neural networks for stellar parameters and abundances from APOGEE spectra.

Neural process that jointly reconstructs and classifies sparse multiband light curves on ELAsTiCC2 and MALLORN.

Physics-motivated light-curve augmentation for rare classes combined with gradient boosted ensembles.

Two-stage framework uses a Vision Transformer to assess aperture-photometry feasibility and a ResNet to predict aperture size for optical survey photometry, achieving a test ROC AUC of 0.96 for feasibility assessment.

Active learning loop on the Fink stream that selects spectroscopic targets to improve early SN Ia classifiers.

Shows that transient classifiers pretrained on simulations or ZTF light curves can be adapted to real data and to LSST with small labelled samples.

Forecasts stellar flares from Kepler light curves by combining stellar physical properties and historical flare records through a soft prompt module and a residual record fusion module.

Hierarchical recurrent classifier for real-time LSST alert streams.

SNGuess, FollowMe and FinalBet classification channels for ELAsTiCC alerts.

Fully convolutional networks with a NaN mask and masked Dice loss detect stellar flares point by point in variable-length Kepler light curves, yielding a catalogue of 30,285 flares on 1,426 stars.

Tree-based and deep learning classifiers in the Fink broker evaluated on ELAsTiCC, including the CATS architecture.

Compares Conv1D-BiLSTM and Swin Transformer classifiers with fine-tuned LLM, multimodal LLM and audio LLM models on Kepler and K2 variable-star light curves; the Swin Transformer reaches 99% accuracy and the LLM-based models about 90%.

Multimodal CNN that selects bright transients for automated spectroscopic follow-up.

Object-detection based classifier on light curves and power spectra from Kepler, TESS and ZTF.

Scalable Gaussian processes for filling gaps in and forecasting light curves.

Multi-label deep learning classifiers for ZTF alert phenomenology.

CNN on wavelength-time heatmaps of supernova light curves.

Transformer classifier for multiband transient light curves evaluated on PLAsTiCC.

Modular machine learning framework for classifying variable sources in ZTF light curves.

ResNet-GRU featurizer and normalizing flow that infer binary microlensing parameters in real time.

Phase-folded light-curve classifiers that are invariant to cyclic permutations.

CNN real-bogus classifier deployed in the ZTF alert pipeline.

CNN ensemble that identifies streaks from fast-moving near-Earth objects in ZTF images.

Recurrent and Bayesian neural networks for light-curve classification.

Feature extraction and classifier pipeline for photometric supernova typing.

Learns a sparse precision matrix for correlated stellar residuals and tests likelihood-ratio exoplanet detection on 13 VLT/SPHERE datasets.

Transformer world model trained on 16,499 Kepler light curves predicts stellar flux, with matched filtering on residuals reaching AUC 0.938 on DR25 and recovering 47 TESS planets zero-shot.

Late fusion of light-curve CNN features with attention and stellar parameters, calibrated by temperature scaling.

Machine learning classifiers trained on injected simulations combined with statistical validation for TESS FFI candidates.

Multi-branch network for transit vetting integrated into the SHERLOCK pipeline.

Quantum extreme learning machine retrieval executed on IBM quantum hardware.

CNN, bidirectional LSTM and attention classifier for Kepler DR25 transit signals.

1D CNN retrieval of composition and planetary properties from HWO-like reflected-light spectra.

Benchmark of PLS, SVM, KNN, tree ensembles and XGBoost on Ariel-like transmission spectra.

Compact CNN on phase-folded light curves that triages TESS and Kepler threshold crossing events.

Neural ratio estimation for marginal posteriors and amortized Bayesian model comparison, applied to WASP-39b.

Convolutional embeddings and transformer encoder applied to flux, centroid and background series of 4.1 million light curves.

Flow matching posterior estimation with importance sampling to verify retrievals and estimate the evidence.

Sequential neural posterior estimation with neural spline flows for self-consistent atmospheric retrievals.

Self-supervised spectrum autoencoder and RV estimator that separate Doppler shifts from activity-induced line-profile changes.

Normalizing-flow posterior estimation for exoplanet atmospheric retrieval.

Data-driven PSF subtraction for high-contrast imaging based on half-sibling regression.

Multi-branch network that mimics expert vetting diagnostics and validates 301 Kepler planets.

Bayesian neural network that predicts instability times of compact planetary systems.

Neural surrogate for radiative transfer that accelerates exoplanet atmospheric retrieval.

Gradient boosted classifier on short N-body integrations that predicts multiplanet stability.

Neural networks for triage and vetting of TESS transit signals.

AstroNet adapted to K2 light curves.

Random forest atmospheric retrieval for exoplanet transmission spectra.

Generative adversarial network for atmospheric retrieval.

CNN transit classifier for Kepler threshold crossing events.

Neural networks that detect transit signals in simulated and Kepler light curves.

An actively trained neural emulator approximates the joint cosmological and nuisance-parameter log posterior using roughly 300,000 likelihood evaluations per dataset.

Amortised network that maps multiband images directly to a variational posterior over tomographic shear and convergence, validated on LSST DESC DC2 images.

Point-cloud diffusion that generates galaxy positions and star formation rates from IllustrisTNG dark matter fields.

Hybrid power spectrum and neural summaries that compress DES Y3 lensing maps to seven statistics.

Pretraining on lognormal mocks and fine-tuning on 60 N-body simulations for KiDS-Legacy field-level inference.

Set-based networks and autoregressive flows on eRASS1-like cluster catalogues.

Pixel-space diffusion transformer that treats cosmological parameters as tokens and samples convergence maps and parameters jointly from noisy shear fields.

Couples a correlator convolutional network whose summaries map to given N-point orders with simulation-based inference on DES Y3-like convergence maps.

Emulates the likelihood of a data set directly with neural networks, giving an auto-differentiable surrogate that matches true-likelihood credible intervals to within 0.1 sigma using fewer than 20,000 evaluations.

Graph CNN compression of spherical lensing and clustering maps with normalizing-flow likelihoods.

Combines analytical blind separation of foregrounds with neural network parameter inference.

Two-channel UNet with frequency-differenced and PCA-cleaned inputs for HI reconstruction under realistic beams.

Maps observed shear fields directly to cosmological parameters, roughly doubling the figure of merit of convergence-based inference on CSST-like mocks.

LSTM and Kolmogorov-Arnold network emulators of the global 21-cm signal for Bayesian inference.

Benchmark of deep learning models that recover the neutral hydrogen fraction from 2D 21-cm power spectra.

Spherical persistent homology summaries with neural compression and likelihood estimation on DES Y3 lensing maps.

Differentiable neural emulator of EFTofLSS galaxy power spectra.

Neural network that predicts the global 21-cm signal from power spectrum measurements under SKA-like noise.

Compares simulation budgets of explicit and implicit full-field inference for LSST Y10-like convergence maps.

Compares neural compression losses against an HMC full-field reference with a differentiable lognormal simulator.

Neural emulator of 21cmFAST power spectra and global signals.

Julia neural emulator of CMB power spectra for gradient-based samplers.

Neural network that extracts global 21-cm parameters from antenna temperatures with foregrounds and ionospheric distortion.

JAX version of CosmoPower for gradient-based samplers.

Iteratively trained emulators of CLASS observables for MCMC.

Graph neural networks on galaxy catalogues for cosmological parameter inference.

Deep learning analysis of combined weak lensing and galaxy clustering maps.

Neural network emulation of the likelihood for accelerated cosmological parameter inference.

Variational autoencoder emulator of the global 21-cm signal.

Graph neural networks trained on CAMELS to infer halo masses from galaxy properties.

Learned harmonic mean estimator of the Bayesian evidence from posterior samples.

Moment networks trained on a single foreground image for CMB B-mode inference.

Information maximising neural network compression of Gaussian and lognormal fields followed by simulation-based inference.

Neural emulators of CMB and matter power spectra.

Neural network emulator of the global 21-cm signal.

Graph convolutions on spherical sky maps such as HEALPix.

Neural ratio estimation for dark matter constraints from stream density variations.

Networks that estimate marginal posterior moments for high-dimensional inference.

GAN that adds small-scale non-Gaussian structure to Galactic dust foreground maps.

Differentiable particle-mesh N-body simulation for gradient-based inference and hybrid ML models.

Learned displacement layers that map dark matter simulations to baryonic fields.

3D UNet ensemble that separates cosmological HI signal from radio foregrounds.

Graph networks distilled into symbolic expressions, including a dark matter halo example.

Scalable 3D GAN for N-body density fields.

Pixel-based convolutions on HEALPix maps applied to CMB analysis.

Neural compression that maximises Fisher information for downstream inference.

GAN that generates weak lensing convergence maps.

Unified interface over sbi, pydelfi and lampe with validation and calibration tools.

Neural posterior estimation package with featurizers for light curves and spectra.

Importance nested sampling with neural network emulation of the likelihood contours.

Coverage test that checks generative posterior estimators using samples alone.

Active-learning Gaussian process surrogate of the posterior for expensive cosmological likelihoods.

Sequential Monte Carlo sampler preconditioned with normalizing flows.

Normalizing flows and KDEs that learn nuisance-free marginal posteriors for marginal KL divergences and emulation.

Truncated marginal neural ratio estimation applied to CMB cosmological inference.

Simulation-efficient marginal posterior estimation with prior truncation.

Density-estimation likelihood-free inference with active simulation acquisition.

Applies the gamma-PhysNet CNN to LST-1 Crab Nebula observations, improving sensitivity over the random forest analysis below 1 TeV.

SIREN neural field that learns the IceCube photon Green function, predicting photon yield and arrival-time distributions for a point energy deposit and sensor.

CNN-LSTM classifiers that separate GRBs, TGFs, solar flares and SGRs in Fermi-GBM light curves and flag outliers.

Unsupervised clustering of GRBs into types using rest-frame prompt emission properties.

U-Net detection with CNN classification and deep-ensemble flux estimation on Fermi-LAT and CTAO simulations.

Neural ratio estimation with spherical convolutions on 14 years of high-latitude Fermi-LAT photon maps.

IceCube Collaboration CNNs that reconstruct energy, direction, vertex and flavour signature of GeV neutrinos in DeepCore for oscillation measurements.

Predicts hits on virtual optical modules between real ones to super-resolve neutrino telescope events, improving muon angular reconstruction.

Neural background estimation with FOCuS-Poisson trigger search on Fermi-GBM count rates.

Point-cloud GNN that classifies and reconstructs GeV-scale neutrino events in IceCube.

Normalizing-flow posterior estimation on spherical Fermi-LAT maps with graph neural network summaries.

Neural networks that estimate source-count distributions and flux fractions of gamma-ray templates.

Bayesian graph CNNs on HEALPix maps that separate point-source and smooth contributions to the Fermi Galactic Center Excess.

Deep learning acceleration of Bayesian parameter inference for EMRI signals.

Normalizing-flow stacking of single-event posteriors with GPU sequential Monte Carlo for EOS inference.

Adds neural posterior estimation to Bilby, analysing 38 high-mass GWTC-3 events with a median importance-sampling efficiency of 7 percent.

DINGO-based GNPE model for echo searches that analyses an O4a event in under 20 minutes on one GPU and finds no significant echo evidence.

Neural network BNS search with heterodyned inputs that runs in real time on a single GPU.

Fuses waveform features with neural posterior samples of common-source parameters to rank lensed multi-image signals against unrelated overlapping mergers in O4 data.

Neural posterior estimation with heterodyning, tailored coordinates and neural rescaling built on cogwheel, covering long signals with secondary masses below 10 solar masses.

Independent normalizing flows for each ppE deviation parameter give posteriors in about 0.4 s per event, a speedup of about 9 x 10^4 over MCMC.

Extends DINGO to lensed signals, reducing parameter estimation and lensing Bayes factor computation from weeks to seconds while matching Bilby.

CNN on matched-filter SNR time series that detects BNS mergers and is robust to glitches.

Gradient boosted trees trained on numerical-relativity simulations predict BNS merger outcomes from inspiral parameters.

Prior-conditioned neural posterior estimation that infers all BNS parameters in about one second, including before merger.

Low-latency neural network search pipeline used in O4.

Semi-supervised anomaly detection for unmodeled gravitational-wave transients.

Truncated marginal neural ratio estimation for compact binary parameter estimation.

MCMC with normalizing-flow global proposals, used for gravitational-wave parameter estimation.

Normalizing-flow posteriors for binary black hole events with inference times of about a minute.

Nested sampler that uses normalizing flows to propose new live points.

Neural network for low-frequency noise subtraction in gravitational-wave detectors.

Conditional VAE for rapid compact binary parameter estimation.

Critical evaluation of CNN-based gravitational-wave detection.

A two-stage k-means pipeline groups 17,422 DKIST Ca II flare profiles into 19 spectral clusters and maps chromospheric kernel–halo structure.

Attention model on 13 co-aligned SDO AIA and HMI channels that forecasts daily 6.5 nm EVE irradiance over three days for 33 significant flares.

ViT and LSTM hybrid that predicts whether upcoming flares will be eruptive or confined.

CNN, CNN-BiLSTM and transformer models with class-dependent rewards for imbalanced M-class flare prediction.

Multimodal LLM fine-tuned with LoRA on magnetograms and textual magnetic parameters for 24-hour flare forecasts.

Compares SigLIP2, VideoMAE and Moirai2 on SDO/HMI magnetograms and GOES X-ray fluxes.

366M-parameter spatiotemporal model trained on eight AIA channels and five HMI products at native resolution.

Physics-informed neural network extrapolation of force-free coronal magnetic fields from photospheric magnetograms.

Unsupervised neural multiframe blind deconvolution of solar observations.

CNN classification of solar features in H-alpha images.

Fully convolutional networks that invert Hinode Stokes maps into 3D thermodynamic and magnetic cubes.

Curated 6.5 TB AIA, HMI and EVE dataset with baselines for EUV irradiance forecasting and instrument translation.

Deconvolution and super-resolution of HMI continuum images and magnetograms.

End-to-end network that estimates horizontal velocity fields from pairs of granulation images.

Two-stage YOLO detection and characterisation network for flux, size and position angle on SKA SDC1 images.

3D CNN detector for HI sources in SKA SDC2 hyperspectral cubes, extending the method that won SDC2.

Diffusion-generated radio galaxies assembled into sky models and imaged to mimic LoTSS pointings.

Self-supervised pretraining and deep clustering that classify LoTSS-DR2 radio galaxies into 12 morphology classes.

Continuous-time diffusion model that generates radio galaxy images with control over flux and morphology.

Series of residual networks for radio interferometric imaging.

Fine-tunes multimodal LLMs on diagnostic plots, text and numerical features to classify pulsar candidates, outperforming earlier machine learning classifiers after a few training epochs.

Learned denoisers in proximal algorithms for radio-interferometric imaging with uncertainty quantification.

BYOL pretraining on unlabelled radio galaxy images for downstream classification.

Neural networks that reconstruct radio images from incomplete visibility data.

Super-resolution network for radio interferometric image reconstruction.

Attention-gated CNNs for FRI and FRII radio galaxy classification.

Semi-supervised GAN classifier for pulsar candidates with few labelled examples.

CNN classifier for fast radio burst and single-pulse candidates.

Faster R-CNN detection and classification of radio source morphologies.

TerraMind-style model pretrained on nearly two million co-registered tile bundles over 11 modalities at 1 m and 100 m scales, evaluated on craters, irregular mare patches and ice prospectivity.

Foundation model mapping 18 channels from six instruments on three lunar missions into one embedding for similarity search, resource mapping, mineral regression and geology classification.

U-Net++ crater detector coupled with an extended Kalman filter for lunar landing navigation.

Multi-sensor foundation model for Mars that combines representations from several orbital instruments for downstream remote sensing tasks.

LLaVA adapted to lunar remote sensing with LUCID, 96k panchromatic images paired with scientific captions and question-answer pairs.

Benchmarks unsupervised visual anomaly detection methods on lunar and Mars rover imagery, with PaDiM localising craters without labels.

OWLv2 vision transformer fine-tuned on LRO images for lunar crater detection.

3,000 tasks derived from lunar mission procedures with Environmental Scenario Indicators for safety, efficiency, integrity and alignment; the best of 36 models reaches 47.8% against 65.1% for human experts.

8B reasoning model trained by continued pretraining, SFT and RL, with Latent Difficulty Perception that cuts chain-of-thought tokens by 38.9%, the Lunar-QA benchmark and RK3588 NPU deployment.

DeepONet thermophysical model of water sublimation, fitted with simulated annealing to the water production curves of 28 comets to estimate nucleus sizes.

Two-stage YOLO detection and CNN or ResNet-50 size classification of craters with an annotated dataset.

Earth-Moon collaborative Chang'e large model with Yutu-Text and Yutu-VL small models, the Lunar_GenData multi-agent instruction workflow built on Chang'e mission data, and the Lunar_Eval suite.

Zero-shot crater segmentation with SAM on lunar and Martian images and DEMs.

115,970 annotated real images of 16 small bodies for keypoint detection and description benchmarks.

CNN crater detection on lunar digital elevation maps, transferred to Mercury.

Images, spectra, time series and catalogs for hundreds of millions of objects, distributed as Hugging Face datasets with benchmark tasks.

Simulated LSST multi-band light curves for transient and variable classification.

15 TB across 16 physics simulation datasets, including MHD and supernova explosion simulations.

4,233 cosmological simulations varying cosmological and feedback parameters, designed for training ML models.

44,100 N-body simulations spanning more than 7,000 cosmological models, designed for information quantification and ML training.