**Organized alphabetically by speaker’s last name**
Daniel Anglés-Alcázar, University of Connecticut
Cosmology and Astrophysics with MachinE Learning Simulations (CAMELS) to model the impact of baryonic physics on cosmological structure formation
The CAMELS project aims to overcome major obstacles limiting our understanding of the fundamental properties of the Universe by (1) providing thousands of state-of-the-art hydrodynamic simulations of cosmological structure formation covering a broad range of sub-grid model implementations of key physical processes driving galaxy evolution and (2) developing novel machine learning algorithms to maximize the extraction of information from cosmological surveys while marginalizing over uncertainties in galaxy formation physics. In this talk, I will review new second-generation simulations expanding the CAMELS public data repository and discuss recent progress towards understanding the impact of feedback from massive stars and supermassive black holes on the cosmic matter distribution and efficiently emulating their effects over large cosmological volumes.
Anirban Bairagi, Institut d’Astrophysique de Paris
PatchNet: GPU and Simulation is not the limitation anymore for Cosmological field-level inference
Extracting cosmological information from the non-linear matter distribution is essential for tightening parameter constraints with upcoming surveys like Euclid, DESI, and the Vera Rubin Observatory. Traditional statistics such as the power spectrum and bispectrum, though widely used, fail to capture the full non-Gaussian structure of the density field. Simulation-Based Inference (SBI) offers a powerful alternative by learning directly from forward-modeled simulations.
In this work, we apply SBI to both standard and primordial non-Gaussian (PNG) Quijote simulations and propose a hybrid hierarchical framework that combines small-scale features from field patches with large-scale P(k) and B(k). This patch-based strategy enables efficient CNN-based training on high-resolution data while circumventing GPU memory bottlenecks.
We demonstrate that our method achieves significantly improved Fisher information over traditional summaries and even wavelet-based statistics. This scalable approach provides a promising path toward extracting near-optimal information content from cosmological fields in both Gaussian and non-Gaussian scenarios.
Manuel Ballester Matito, Northwestern University
Accelerating Stellar Structure Modeling with Neural PDE Solvers
We develop a Physics-Informed Neural Network (PINN) architecture to solve the steady-state structure equations of single stars as a mesh-free alternative (or complement) to traditional solvers like MESA. Focusing on the hydrostatic regime, we rewrite the governing equations (mass conservation, hydrostatic equilibrium, radiative transport, and energy generation) using enclosed mass as the independent variable. Our PINN outputs temperature, pressure, radius, and luminosity as functions of mass while incorporating boundary conditions as hard constraints within the architecture. Opacity and density are supplied to the PINN through auxiliary neural networks that are pre-trained to interpolate semi-empirical tabulated data with an R² above 99.9%. We demonstrate that the PINN achieves high accuracy in both supervised mode (leveraging sparse MESA outputs) and physics-only regime (solving purely from PDE residuals), with R² scores of 99.9% and 99.6%, respectively. We utilize a Multilayer Perceptron (MLP) model with sinusoidal activation functions (SIREN) and layer normalization. Moreover, unlike traditional PINNs, the output is constrained to force the boundary conditions. While time dependence and convection are not yet included, our preliminary experiments show the promising results of neural PDE solvers in accelerating stellar modeling. Ongoing work will extend this to full evolutionary models and binary interactions.
Adrian Bayer, Flatiron Institute/Princeton University
Optimal, Interpretable, and Robust Cosmological Inference of the Multimodal Sky
Extracting maximal information from upcoming cosmological surveys is essential to understanding phenomena such as neutrino mass, dark energy, and inflation. This requires developing new analysis methodologies and leveraging synergies across datasets.
I will first present advancements for extracting information from cosmological data via differentiable forward modeling and AI-based reconstructions of the Universe’s initial conditions. To achieve this, I will present accelerated methods to sample high-dimensional distributions, such as Microcanonical Langevin Monte Carlo, alongside neural approaches using CNNs, diffusion models, and stochastic interpolants. I will compare their efficiency and show applications to DESI, yielding tighter constraints than traditional methods. I will also interpret what CNNs learn from cosmological fields and motivate the use of out-of-distribution detection to ensure robust inference.
I will then discuss strategies for combining cosmological datasets to break degeneracies and calibrate systematics. In particular, I will present the HalfDome simulations: full-sky, multimodal simulations designed for joint analysis of Stage IV surveys, including galaxies, lensing, and the CMB. I will show how these simulations are being used in conjunction with AI methods to accelerate forward modeling and obtain optimal information about our Universe.
Supranta Sarma Boruah, University of Pennsylvania
Generative machine learning solutions for weak lensing mass mapping
We present a novel diffusion-based approach for weak lensing mass map reconstruction that addresses key challenges in Stage IV cosmological surveys. Our method uses a single diffusion model trained on N-body simulations for dual purposes: generating fast simulations and reconstructing mass distributions from noisy shear data.
Using Diffusion Posterior Sampling (DPS), we train an unconditional diffusion model as a prior, then solve the inverse reconstruction problem through likelihood-guided sampling. We identify and correct systematic biases in standard DPS by down-weighting likelihood gradients at early timesteps, achieving unbiased inference.
Our reconstructed maps outperform Kaiser-Squires inversion with ~50% improvement in correlation coefficients for low signal-to-noise data. The method accurately reproduces both Gaussian (power spectra) and non-Gaussian statistics (scattering transforms, peak/void counts) at sub-arcminute resolution. We are currently applying this technique to reconstruct mass maps from actual DES-Y3 survey data.
This work demonstrates the transformative potential of generative AI in astronomy, providing computationally efficient tools for uncertainty quantification in cosmological inference while enabling robust detection of cosmic structures from survey observations.
Hong-Yu Chen, Northwestern University
StarEmbed-GPT: a foundation model for general-purpose inference on variable stars
We introduce StarEmbed-GPT (Star Embedding Generative Pre-trained Transformer), a Transformer-based foundation model tailored to variable star light curves. By fine-tuning a large pre-trained time series foundation model on 50 GB of ZTF light curves, StarEmbed-GPT offers three key features: 1) General-purpose capability: It is the first foundation model that supports a broad range of downstream tasks on variable star data. 2) Survey-agnostic pre-training and fine-tuning: The pre-training on the large-scale of cross-domain time series data and fine-tuning on the light curve data equip the model with a survey-agnostic understanding of the underlying variable stars. 3) Universal embeddings: Its learned representations achieve state-of-the-art zero-shot performance on periodic star classification, period inference, and anomaly detection. Our empirical results demonstrate the efficacy of StarEmbed-GPT’s fine-tuning and point to a promising new paradigm where time-series foundation models significantly enhance time-domain astronomy studies using data from surveys like LSST.
Samuel Dyson, The University of Chicago
Uncertainty Quantification in Time-Series Coincidence Detections
In recent decades, several significant astronomical discoveries have relied on our ability to verify that coincident events in two or more detectors are in fact associated with the same astrophysical source, whether neutrinos from supernovae or gamma-ray bursts from merging neutron stars. Forthcoming missions like LSST will provide millions of triggers of transient events per night. Correlating this time stream of alerts with time streams from other astronomical observatories will uncover a staggeringly large AI-assisted discovery space, while also increasing the potential for false detections.
We seek to strengthen the statistical foundations of coincidence detection. Our objective is to identify instances in which two or more processes exhibit temporally aligned or near-simultaneous activity beyond what would be expected by chance. We examine the (implicit or explicit) assumptions that underlie two common approaches to coincidence detection in astronomy and beyond.
We find that, given the assumption of stationarity in each detector’s data stream and independence between multiple detectors, an upper limit can be established on the rate of false detections. Even under weaker assumptions, a theoretical limit on false detections can be found. Our initial findings suggest that coincident detection claims should be reassessed as additional post-event data become available.
Alex Gagliano, Massachusetts Institute of Technology
Mixture-of-Expert Variational Autoencoders for Multi-Modal Embedding of Supernova Data
Time-domain astrophysics requires the collection and analysis of multi-modal data. Specialized models are often constructed to extract information from a single modality, but this approach ignores the wealth of cross-modality information that may be relevant for the downstream tasks to which the model is applied. In this work, we propose a multi-modal, mixture-of-expert variational autoencoder to learn a joint embedding for supernova light curves and spectra. Our method, which is inspired by the perceiver architecture, natively accommodates variable-length inputs and the irregular temporal sampling inherent to supernova light curves. We train our model on radiative transfer simulations of SNe Ia and validate its use for cross-modality generation and the recovery of physical parameters. Our model achieves superior performance in cross-modality generation to nearest-neighbor searches in a contrastively-trained latent space, showing its promise for learning informative latent representations of multi-modal astronomical datasets.
Arkaprabha Ganguli, Argonne National Laboratory
Enhancing interpretability in generative modeling: statistically disentangled latent spaces guided by generative factors in scientific datasets
Scientific discovery often requires identifying relationships among noisy, biased, and uncertain measurements—tasks at which purely data-driven models, despite their strong predictive power, frequently lack interpretability. In extragalactic astronomy and cosmology, for example, we wish to link observed galaxy images and spectra to underlying physical drivers such as the dark-matter environment and evolutionary history; however, many dynamical parameters remain poorly constrained. Typical AI pipelines excel at classification or redshift estimation yet provide limited insight into these mechanisms, in part because their latent spaces entangle multiple generative factors and rarely exploit the data’s multi-fidelity nature.
We present an encoder–decoder generative framework that learns disentangled representations by injecting domain-specific auxiliary information into the latent space. Well-understood generative factors are allocated to separate, interpretable dimensions, while uncertain or unknown factors remain entangled. This design improves semantic structure, enhances explainability, and increases robustness to adversarial or out-of-distribution inputs.
We demonstrate the method on two testbeds: Synthetic galaxy images with known structural parameters, and cosmic microwave background lensing maps with associated halo properties from simulations. The model maintains high reconstruction fidelity, and the disentangled dimensions correspond to physically plausible perturbations—providing a transparent link between learned features and underlying cosmic processes.
Aritra Ghosh, University of Washington
Harnessing ML & Large Surveys to Probe Galaxy Evolution: from HSC Size-Environment Correlations to Rubin Unsupervised Discovery
The combination of machine learning and large astronomical surveys is revolutionizing our ability to study galaxy evolution. I will present two key results that exemplify this synergy.
First, we used a novel Bayesian deep neural network to measure the sizes of 3 million galaxies from the Hyper Suprime-Cam survey with high precision and with robust uncertainties. Investigating the long-debated correlation between galaxy size and environmental density, we find with >5σ confidence that galaxies in denser environments are up to 25% larger than their counterparts of similar mass and morphology in less dense regions. This surprising result provides new insights into the connection between galaxy structure, halo assembly, and merger history—challenging current models to more comprehensively account for environmental effects.
Second, I will introduce a new ML-driven image-based unsupervised discovery framework for Rubin-LSST data. Capable of processing hundreds of millions of images, it combines multiple unsupervised deep learning techniques with image similarity search and interactive unsupervised exploration within a single tool. It can detect a wide range of irregular galaxies, gravitational interactions, galaxy mergers, tidal remnants, and other low surface brightness features. This framework, combined with Rubin’s wide-fast-deep survey, provides us with a powerful new tool to investigate dark matter dynamics and the hierarchical growth of structure in the universe.
Jimena González, University of Wisconsin–Madison
Does Machine Learning Work? A Comparative Analysis of Strong Gravitational Lens Searches in the Dark Energy Survey
We present a systematic comparison of three independent machine learning (ML)–based searches for strong gravitational lenses applied to the Dark Energy Survey (Gonzalez et al. 2025; Rojas et al. 2022; Jacobs et al. 2019a,b). Each search uses a distinct ML architecture and training strategy, enabling evaluation of their relative performance and complementarity. Using a visually inspected sample of 1,651 lens candidates, we assess how each model scores these systems and quantify their agreement with expert classifications. We find successive ML efforts show marked improvement, with AUROC values of 0.67, 0.77, and 0.90 for the Jacobs, Rojas, and Gonzalez models, respectively. Among high-confidence candidates, 88% receive ML scores above 0.8 from the Gonzalez model, compared to 69% and 51% from Rojas and Jacobs. Only 7% of high-confidence lenses are missed (score < 0.8) by all three models. We also explore ensemble techniques to combine individual predictions, including averaging, median, linear regression, decision trees, random forests, and an Independent Bayesian method. All ensemble methods—except averaging—outperform individual models, with the top three achieving up to 2x improvement in precision. Our findings highlight the potential of ensemble strategies and provide practical insights for optimizing ML-searches in upcoming astronomical surveys.
Keiya Hirashima, RIKEN Center for Interdisciplinary Theoretical and Mathematical Sciences
Star-by-star Galaxy Simulations Accelerated by Surrogate Modeling for Supernova Feedback
In recent decades, galaxy simulations have revealed the multiscale interdependence of gas physics—such as star formation, stellar feedback, and inflows/outflows—through improved models and resolution. However, due to limitations in resolution and scalability, simplified sub-grid models remain prevalent. Even zoom-in simulations of Milky-Way-sized galaxies typically cap mass resolution at ~1,000 solar masses (e.g., Applebaum et al. 2021), roughly the scale of molecular clouds. To overcome this, we are developing the N-body/SPH code ASURA-FDPS-ML, optimized for exascale systems (e.g., Fugaku), capable of handling ~100 billion particles and resolving individual stars and their feedback. Achieving efficient parallelization at this scale is challenging. Although hierarchical time-stepping methods have been introduced to reduce costs, communication overhead limits scalability beyond ~1,000 cores—especially for short-timescale, localized events like supernovae.
To address this, we developed a deep-learning-based surrogate model that replicates supernova feedback in real-time (Hirashima et al., 2025). During simulation, it predicts physical quantities 100,000 years into the future per explosion in 3D, reducing computational cost to ~1% of direct resolution. In this presentation, we demonstrate simulation results using this approach, focusing on star formation and outflows driven by resolved stellar feedback, and discuss the fidelity and computational benefits of our deep-learning integration.
Matthew Ho, Columbia University
Learning the Universe: Building a Scalable, Verifiable Emulation Pipeline for Astronomical Survey Science
Learning the Universe is developing a large-scale, ML-accelerated pipeline for simulation-based inference in cosmology and astrophysics. By combining high-resolution physical models with fast emulators, we generate realistic training sets at the scale required for field-level inference from galaxy survey data. This enables us to constrain models of galaxy formation and cosmology from observations with unprecedented scale and precision. In designing this pipeline, we have also developed validation methodologies to assess emulator accuracy, identify sources of systematic error, and support blinded survey analysis. I will present results from its application to the SDSS BOSS CMASS spectroscopic galaxy sample and discuss how this approach is scaling to upcoming cosmological surveys.
Lindsay House, SkAI Institute
Harnessing AI and Citizen Science to Improve Our Understanding of Dark Energy
This talk will discuss research at the intersection of AI, cosmology, citizen science, and public engagement. We harness a large-scale participatory science effort combined with machine learning to enhance the Hobby-Eberly Telescope Dark Energy Experiment (HETDEX). Classifications from the citizen scientists inform our models and improve the interpretation of our AI results. We have utilized these techniques together to remove false positives, allowing us to use lower signal-to-noise data and sources with low goodness of fit. With seven million classifications through Dark Energy Explorers, we can confidently determine if a source is not real at over a 94% confidence level when classified by at least ten individuals; this confidence level increases for higher signal-to-noise sources. To date, we have labels from 257,000 Lyman-alpha emitting (LAEs) galaxies, all of which have been classified by a Dark Energy Explorer. We can accommodate a tenfold increase by innovative machine learning methods combined with the visually vetted samples from Dark Energy Explorers. Using this technique, we generate a real or false positive classification for the full candidate sample of 1.2 million LAEs. We currently have around 20,000 volunteers from 159 countries worldwide, providing a free educational opportunity that requires no prior technical knowledge.
Patricia Iglesias-Navarro, Institute of Astrophysics of the Canary Islands
Simulation-based inference of galaxy properties from JWST pixels
We present a novel simulation-based inference framework addressing the challenging inverse problem of extracting physical parameters from high-dimensional, multiwavelength astronomical data. Our approach tackles inherent degeneracies in spectral energy distribution fitting through amortized Bayesian inference, enabling rapid posterior sampling across millions of individual galaxy pixels with quantified uncertainties.
The framework leverages neural posterior estimation trained on synthetic datasets from stellar population models, incorporating parametric and non-parametric star formation histories with realistic observational noise. We implement physics-informed priors and robust uncertainty quantification methods for sparse, noisy astronomical observations. Our neural architecture captures complex parameter correlations while maintaining computational efficiency for large-scale surveys.
Validation demonstrates excellent calibration with R² = 0.99 for stellar mass estimation and well-calibrated posteriors in high-noise regimes. Benchmarking against traditional MCMC methods achieves four orders of magnitude speedup while maintaining comparable accuracy. Applied to the JWST Advanced Deep Extragalactic Survey, our framework processes pixel-by-pixel photometry for ~2 million individual pixels from 1,083 galaxies in ~1CPU-day, demonstrating exceptional scalability.
Our methodology exemplifies modern ML’s transformative potential in astronomy, enabling previously computationally prohibitive population-level analyses. The approach generalizes beyond galaxy fitting, offering a solution for inverse problems where traditional methods become intractable, with broad applications for automated pipelines in next-generation surveys.
Sreevani Jarugula, Fermilab
Cosmology constraints from Strong Gravitational Lensing using Neural Ratio Estimation
Strong gravitational lenses are powerful cosmic tools for probing the nature of dark energy and dark matter. Modern cosmic surveys will uncover millions of lenses and traditional methods such as monte carlo algorithms will be computationally expensive for inference. New tools like Simulation-Based Inference (SBI) using machine learning, present an opportunity for faster inference where the likelihood is intractable. In this work, we apply Neural Ratio Estimation (NRE) to infer the posterior distributions of cosmological parameters — the dark energy equation-of-state parameter w and the dark matter density Omega_m. The NRE model is trained on simulated strong lens images generated from cosmologies sampled with uniform prior. During inference, we perform population-level analysis by combining multiple lenses to construct a joint likelihood. Our results demonstrate that while Omega_m can be effectively constrained even from single strong lens image, recovering w proves more challenging due to degeneracies between the two parameters. However, when either w or Omega_m is fixed, the other can be robustly estimated even from single images. Population-level inference improves overall constraints, though the NRE model’s performance varies across the prior space with some regions showing bias and overconfidence. The best-fit reduced chi-square values occur at w=1.7 and Omega_m=1.1, while the worst cases reach values as high as 7.5 and 16.9, respectively, highlighting the variation in model reliability. This study showcases both the promise and hurdles of NRE in extracting cosmological insights from strong lensing. Our findings motivate further investigation into how parameter degeneracies impact NRE model reliability and the need for more robust SBI approaches for static strong lens studies.
Tianao Li, Northwestern University
Probabilistic Imaging of Galaxies for Weak Gravitational Lensing
Astrophysical images contain rich information about the universe. We investigate uncertainty quantification for inverse problems in astronomical imaging and the propagation of uncertainty from image reconstruction to a downstream astrophysical task. In weak gravitational lensing, the subtle distortions in galaxy shape, also known as the cosmic shear, carry information about the dark matter and can be used to reconstruct the structure of the cosmic web. We formulate the deblurring and denoising of galaxy images from telescopes as the initial inverse imaging task and treat cosmic shear estimation as a downstream task. We show how various inverse imaging algorithms using unrolled optimization, variational inference, and score-based diffusion priors lead to different trade-offs among image quality, uncertainty calibration, and computational cost. We also explore the propagation of uncertainty from image to downstream astrophysics, where the goal is to constrain the astrophysical parameter predictions with meaningful and accurate error bars.
Shunyuan Mao, Rice University
Multi-resolution neural representation for self-supervised image reconstruction in radio interferometry
Image reconstruction in radio interferometry presents a classic ill-posed inverse problem, where the goal is to recover a continuous image from sparse Fourier domain samples. Conventional methods like CLEAN fail on extended astronomical sources due to a restrictive point-source prior, while Regularized Maximum Likelihood (RML) approaches are hindered by an explosion of parameters that scales with image resolution, leading to convergence difficulties.
To overcome these limitations, we introduce a self-supervised neural regularized maximum likelihood framework that leverages a coordinate-based neural network as a continuous model of the image. Our model represents the sky brightness distribution with a compact network that maps 2D coordinates to intensity values. By utilizing multi-resolution feature grids, the network can capture both large-scale structures and fine details, promising orders of magnitude faster convergence and more accurate reconstruction than pixel-grounded RML.
We validate our approach on complex, extended-source datasets from the DSHARP and exo-ALMA surveys. The proposed method demonstrates superior reconstruction quality compared to both CLEAN and traditional RML, offering a promising new direction for interferometric imaging using machine learning graphics representations.
Matiwos Mebratu, Stanford University
Hybrid Prior Wavelet based Conditional Flow Matching Model (HyWave-CFM)
Cosmological inference from CMB observations typically relies on Gaussian approximations for foreground components due to computational limitations in generating sufficient non-Gaussian realizations. We address this constraint by developing a wavelet-based conditional flow matching (CFM) model for simulating high-resolution (512×512 pixels, 1 arcmin) CMB secondary maps. Our approach jointly models non-Gaussian lensing convergence ($\kappa$), cosmic infrared background (CIB), and thermal Sunyaev-Zel’dovich (tSZ) fields, capturing their cross-correlations and statistical properties. By recursively simulating conditional distributions of wavelet coefficients with scale-specific prior distributions, we produce samples that reproduce power spectra and Minkowski functionals within a few percent of validation data. The models enable direct inference due to their tractable log-likelihood but also can be used to generate non-Gaussian foreground samples for existing inference pipelines. As a next step, we are investigating the latter potential by combining these simulations with the Marginal Unbiased Score Expansion (MUSE), a fast approximate Bayesian method for high-dimensional field-level inference. This prospective integration aims to facilitate cosmological parameter estimation that accounts for the complex statistical nature of foreground contamination in CMB observations, potentially enhancing the results from current and upcoming CMB experiments.
Peter Melchior, Princeton University
Optimizing Instrument Utilization and Survey Designs with Structured Learning
Wide-field surveys or follow-up campaigns require making decisions with many choices (dozens or even billions) and often unknown outcomes. The complexity of such tasks overwhelms humans and conventional optimization algorithms. I will show how to side-step the discontinuous nature of decision-making and optimize surveys for maximal scientific yield with Graph Neural Networks. I will also demonstrate how such decisions can be updated on-the-fly by connecting the decision architecture with suitable data encoders. These developments provide the building blocks for (semi)autonomous control of robotic telescopes.
Tri Nguyen, Northwestern University
Generating Halo Merger Trees with Graph Generative Models
Merger trees track the hierarchical assembly of dark matter halos across cosmic time and are essentials input for semi-analytic models of galaxy formation. However, traditional methods for constructing merger trees rely on ad-hoc assumptions and struggle to incorporate environmental information. We present FLORAH, a novel generative model for merger trees that represents them as directed acyclic graphs, employing recurrent neural networks to capture the temporal evolution of merger events combined with normalizing flows to model the distributions of halo properties. Trained on cosmological N-body simulations, our model can accurately reproduce key merger statistics across a wide mass and redshift range. We further validate the scientific utility of the model by applying the Santa Cruz Semi-Analytic Model to the synthetically generated merger trees and successfully reproducing galaxy-halo scaling relations. FLORAH offers a computationally efficient alternative to simulation-based merger trees while maintaining statistical fidelity, and can be generalized to a broad class of tree-generative problems beyond galaxy formation.
Jack O’Brien, University of Illinois Urbana-Champaign
Physics Informed Latent Space in Foundation Models of Transients
Time-series foundation models for astronomical observations of transients have promising applications in rapid data-driven classification schemes and light curve prediction for follow-up. Especially important for Rubin, which will observe thousands of transients per night, the ability to select targets and plan follow-up with foundation models will be important for more detailed spectroscopic data collection required for physical parameter inference. With the ultimate goal of developing a physical understanding of astrophysical transients, inferring the underlying physics from observations requires both accelerated methods for physical simulation as well as a foundation model that is capable of understanding transients in a physics informed manner. We aim to provide both, with detailed emulated SEDs coupled with a data-driven photometric time-series encoder mapping to a shared latent embedding, our model will produce posterior probabilities for intrinsic physical models of transients from early observations to all future points in the object’s evolution.
Hyosun Park, Yonsei University
Transformer-based Reduction of PSF Effect and Correlated Noise for Precision Dark Matter Mapping
Current and upcoming weak-lensing (WL) surveys, such as Euclid, LSST, and Roman, will enable us to map the three-dimensional dark matter distribution with unprecedented statistical power. However, the success of these missions depends on the ability to control systematics, including PSF effects and noise correlations. To address the challenge, we propose a Transformer-based approach to perform deconvolution and denoising to enhance galaxy images. Our model is trained on paired datasets consisting of high-quality images and their degraded counterparts. We first apply the method to HST-quality images and achieve exceptional restoration of photometric, structural, and morphological information, approaching the quality of the JWST observations for the same targets. We then extend the model to restore ground-based images contaminated by correlated pixel noise using spatially varying PSFs as additional inputs. Finally, we employ the trained model to enhance Subaru images of the massive galaxy cluster MACS J0717.5+3745 and demonstrate a significant improvement of the WL S/N value. We also present detailed comparisons with HST-based WL analyses.
Helen Qu, Flatiron Institute
An Astrophysical Case Study in Robustness
Robustness of machine learning models is a particularly challenging yet often overlooked problem in scientific applications. In particular, models trained on a labeled source domain (e.g., low-redshift, bright supernova lightcurves) often generalize poorly when deployed on an out-of-distribution (OOD) target domain (e.g., high-redshift, noisy supernova lightcurves). When unlabeled target data is available, self-supervised pretraining (e.g., masked autoencoding or contrastive learning) is a promising method to mitigate this performance drop. To better leverage pretraining for distribution shifts, we propose Connect Later: after pretraining, fine-tune with targeted augmentations designed with knowledge of the distribution shift. Pretraining learns good representations within the source and target domains, while targeted augmentations connect the domains better during fine-tuning. Connect Later improves average OOD error over standard fine-tuning and supervised learning with targeted augmentations on astronomical time-series classification and redshift estimation tasks, as well as other real-world applications.
Aayush Saxena, University of Oxford
Using Deep Learning to classify high-redshift galaxy spectra from JWST: uncovering exciting galaxy populations through AI
The public nature of large JWST observing programmes has resulted in the availability of over 10,000 high-quality galaxy spectra in the archive, many of which have reliable spectroscopic redshifts. With early JWST observations revealing a small number of unexpectedly exciting galaxy populations such as the puzzling Little Red Dots (LRDs), extreme emission line galaxies (EELGs), and nitrogen-enhanced galaxies, to name a few, the logical next step is to assemble large populations of these galaxies to enable a deeper understanding of their origin. In this talk, I will introduce LearnSpec, a new deep learning model using a variational autoencoder (VAE) architecture, which accurately encodes high-redshift JWST spectra to a multi-dimensional latent space, enabling the discovery of exciting classes of objects by means of clustering. When run over 4,000 galaxies at z>4 with reliable spectroscopic redshifts, LearnSpec has robustly identified statistical samples of LRDs, EELGs, nitrogen-rich galaxies, quiescent galaxies, and Lyman-alpha emitters. With small reconstruction errors, LearnSpec offers the possibility of automatically classifying large samples of galaxies from all current and future observing missions across telescopes, making it a powerful tool for galaxy evolution studies.
Ved Shah, Northwestern University
ORACLE: A Real-Time, Hierarchical, Deep-Learning Photometric Classifier for the LSST
ORACLE is the first hierarchical deep-learning model for real-time, context-aware classification of transient and variable astrophysical phenomena. ORACLE is a recurrent neural network with Gated Recurrent Units (GRUs), and has been trained using a custom hierarchical cross-entropy loss function to provide high-confidence classifications along an observationally-driven taxonomy with as little as a single photometric observation. Contextual information for each object, including host galaxy photometric redshift, offset, ellipticity and brightness, is concatenated to the light curve embedding and used to make a final prediction. Training on ∼0.5M events from the Extended LSST Astronomical Time-Series Classification Challenge, we achieve a top-level (Transient vs Variable) macro-averaged precision of 0.96 using only 1 day of photometric observations after the first detection in addition to contextual information, for each event; this increases to >0.99 once 64 days of the light curve has been obtained, and 0.83 at 1024 days after first detection for 19-way classification (including supernova sub-types, active galactic nuclei, variable stars, microlensing events, and kilonovae). We also compare ORACLE with other state-of-the-art classifiers and report comparable performance for the 19-way classification task, in addition to delivering accurate top-level classifications much earlier.
Liren Shan, Toyota Technological Institute at Chicago
Volume Optimality in Conformal Prediction with Structured Prediction Sets
Conformal Prediction is a widely studied technique to construct prediction sets of future observations. Most conformal prediction methods focus on achieving the necessary coverage guarantees, but do not provide formal guarantees on the size (volume) of the prediction sets. We first prove the impossibility of volume optimality where any distribution-free method can only find a trivial solution. We then introduce a new notion of volume optimality by restricting the prediction sets to belong to a set family (of finite VC-dimension), specifically a union of k-intervals. Our main contribution is an efficient distribution-free algorithm based on dynamic programming (DP) to find a union of k-intervals that is guaranteed for any distribution to have near-optimal volume among all unions of k-intervals satisfying the desired coverage property. By adopting the framework of distributional conformal prediction (Chernozhukov et al., 2021), the new DP based conformity score can also be applied to achieve approximate conditional coverage and conditional restricted volume optimality, as long as a reasonable estimator of the conditional CDF is available. While the theoretical results already establish volume-optimality guarantees, they are complemented by experiments that demonstrate that our method can significantly outperform existing methods in many settings.
Nolan Smyth, University of Montreal
Towards the automated discovery of differential equations in (astro)-physics
Symbolic Regression (SR) seeks to automate the discovery of analytic expressions that fit data, enabling interpretability – an essential feature in physics and astrophysics. Recent advances in deep learning have revitalized interest in SR, yet most efforts have focused on finding empirical laws rather than searching for differential equations, whose solutions inherently encode deeper physical principles.
While empirical laws are valuable, true interpretability often requires working in the space of differential equations, which provide a more abstract yet fundamental description of physical systems. This shift is particularly important in astrophysics, where a wealth of complex observational data and the need for physically meaningful models intersect. Differential equations can offer simple governing laws even when their solutions are highly intricate or lack explicit closed-form expressions.
I will present an extension of the Φ-SO (Physical Symbolic Optimization) framework, which employs deep reinforcement learning to discover differential equations whose solutions fit data and (possibly) satisfy additional physical constraints – such as symmetries, conservation laws, or asymptotic behaviors. This is achieved by extending the framework’s operator set to include differential terms. I will demonstrate the power of this approach across a range of astrophysical applications, from high-energy astrophysics to galactic dynamics and cosmology, illustrating how symbolic learning can bridge the gap between data-driven discovery and more fundamental physical insight.
Shubhendu Trivedi, Fermilab
Conformal Hierarchical Simulation-based Inference with Local Validity
Trustworthy and interpretable uncertainty quantification remains a core challenge in artificial intelligence. Simulation-based inference (SBI) offers a flexible toolkit for estimating latent parameters with uncertainty, often using expressive neural density estimators to model complex, high-dimensional posteriors. However, these estimators frequently yield miscalibrated credible regions, relying on heuristic coverage diagnostics. We introduce the first SBI framework with finite-sample, local coverage guarantees–valid in the neighborhood of each observation. Our method augments any hierarchical SBI engine with a conformal Bayesian post-processing step applied to the posterior predictive density. A kernel-weighted conformity score adapts the conformal quantile to the local data geometry, enabling prediction sets that are (i) marginally calibrated, (ii) locally valid, and (iii) hierarchical–handling both global and observation-specific parameters in a single pass. Across synthetic examples and real-world benchmarks from neuroscience and physics, our method achieves guaranteed coverage where existing SBI methods tend to under- or over-cover. It also maintains competitive credible set sizes with minimal computational overhead. Methods like this one, which respect statistical constructs and epistemic intent, will be crucial for trustworthy uncertainties and decision-making in fast or automated inference contexts.
Laura Trouille, Adler Planetarium, Zooniverse
Zooniverse and SkAI: Human-AI Collaboration for Scalable Scientific Discovery
Zooniverse—the world’s largest platform for people-powered research—combines human insight at scale with artificial intelligence (AI) to accelerate discovery, improve data efficiency, and foster inclusive public engagement. With over 2.8 million volunteers and over 70 active research projects, Zooniverse enables hybrid human-AI workflows that are especially valuable in anomaly detection, ambiguity resolution, and training data generation.
This talk will highlight the opportunities and impact of Zooniverse in the era of big data and AI, including how researchers are integrating AI across disciplines—and particularly in astronomy—to complement machine-only approaches. I’ll give an overview of the half dozen planned SkAI-Zooniverse project collaborations, including results from this summer’s Rubin Observatory Data Preview 1 (DP1) kickoff projects.
I’ll also introduce the human-in-the-loop latent space exploration tool being co-developed with the SkAI community to support more efficient data exploration and AI-human collaboration. Finally, I’ll share an update from our Kavli-funded work-in-progress AI Ethics workshops, focused on developing ethical and transparent approaches to integrating AI into participatory science workflows.
Ana Sofia Uzsoy, Harvard University
Bayesian Component Separation for DESI LAE Automated Spectroscopic Redshifts and Photometric Targeting
Component separation is a critical step in disentangling multiple signals and extracting useful information from spectra. We present MADGICS (Marginalized Analytic Dataspace Gaussian Inference for Component Separation), a data-driven Bayesian component separation technique that can separate a spectrum into any number of Gaussian-distributed components. We then discuss the application of this technique for automatically determining spectroscopic redshifts for Lyman Alpha Emitter (LAE) galaxies observed with the Dark Energy Spectroscopic Instrument (DESI) while marginalizing over sky residuals to separate sky from target emission lines. We create a covariance matrix from real spectra of visually inspected DESI LAE targets to provide physically motivated priors, and determine redshifts by jointly inferring sky, LAE, and residual components for each individual spectrum. We demonstrate this method on 910 spectroscopically observed DESI LAE candidate spectra and determine their redshifts with >90% accuracy. Using the chi-squared value from our pipeline as a proxy for detection confidence, we then explore potential survey design choices and implications for targeting LAEs with medium-band photometry. This novel method allows for scalability and accuracy in component separation and automatic redshift determination while maximizing interpretability, and is also broadly generalizable to any other spectral features of interest.
Georgios Valogiannis, The University of Chicago
Saturating Cosmological Information with AI: Field-Level Inference and Beyond
Next-generation cosmological surveys such as LSST and DESI will produce petabyte-scale datasets containing the positions, shapes, and redshifts of millions to billions of galaxies. While traditional analyses rely on compressed summary statistics like the power spectrum or correlation functions, these methods leave significant information untapped—especially in the non-Gaussian regime. This project aims to explicitly quantify the total constraining power encoded in cosmological fields, using field-level inference (FLI) as a gold standard. By modeling the likelihood of observed fields directly—rather than through intermediate statistics—we can characterize the full information content available for constraining cosmological parameters.
In parallel, we explore whether modern summary statistics can efficiently approximate this information. Specifically, we compare the performance of neural network–based summaries (e.g., convolutional neural networks) and mathematically interpretable transforms such as the wavelet scattering transform (WST). These tools are evaluated on simulated and observationally realistic data using simulation-based inference pipelines. By systematically assessing how closely these summaries saturate the FLI benchmark, we aim to identify optimal representations for downstream cosmological inference.
This work bridges cosmology, statistics, and machine learning, and informs how best to extract information from high-dimensional astrophysical data in the LSST and DESI era and beyond.
Amanda Wasserman, University of Illinois Urbana-Champaign
Improving Supernova Cosmology with Active Learning Follow-up
The Legacy Survey of Space and Time (LSST) will observe over 10^6 transients per year, increasing the number of yearly discovered transients by a factor of 100. Type Ia supernovae (SNe Ia) within this dataset are essential for accurately estimating cosmological parameters and improving our understanding of the universe. Spectroscopic classification is the only way to classify transients such as SNe Ia unambiguously, but spectroscopic resources are scarce, an expected <1% of transients will be followed up. To maximize the SNe Ia used in the cosmology analysis, we must create a photometrically classified sample of SNe Ia, but this process is imperfect and allows contamination particularly from SNe Ib/c and SNe II.
We will create an ideal, representative transient training set through active learning to obtain a pure sample of SNe Ia. We will spectroscopically observe and classify the hardest to photometrically label objects, adding these objects to our training set ensuring AI algorithms have an easier time classifying similar light curves going forward (as seen in Moller, 2025). We implement the automated system, the Recommendation System for Spectroscopic Follow-up (RESSPECT) to sort through the millions of Rubin SNe and prioritize the most cosmologically useful for followup. Preliminary results using the simulated ELAsTiCC dataset show that SNe Ia photometrically classified using a RESSPECT chosen training set leads to tighter constraints and more accurate predictions on w_0 and w_a.
Jiezhong Wu, Northwestern University
A foundation AI model to infer the physics of transients
The discovery rate of optical transients will explode to 10 million public alerts per night once the Rubin Observatory’s LSST comes online, overwhelming the traditional physics-based inference pipelines. A foundation AI model is of interest because it can deliver millisecond-scale inference for thousands of objects per day, whereas legacy MCMC codes need hours per object. Moreover, Rubin requires light curve extrapolation and prediction from few samples at early times to optimize very limited spectroscopic follow-up resources. We present an on-the-fly autoencoder trained on truncated partial light curves. A tailored GRU/DeepSets encoder, multi-band parametric basis decoder, and population-size insensitive loss jointly learn low-dimensional representations that help predict both peak time and peak flux, which can be readily incorporated into Rubin Alert Brokers for real-time predictions.
Bin Xia, Georgia Tech and Argonne National Laboratory
Towards a Generalizable Multi-Modal Foundation Model for Astrophysical Data
Advancements in machine learning have enabled new approaches to understanding complex astrophysical phenomena, yet current models often remain domain-specific and modality-limited. This work proposes a generalizable foundation model capable of ingesting and jointly reasoning over diverse astrophysical data modalities, including scalar properties (e.g., redshift, mass), time series (e.g., star formation histories), and high-dimensional spectra or images. Built on a transformer-based architecture, the model is trained on both simulation and observation datasets and can flexibly predict missing or complementary physical quantities, such as inferring the spectral energy distribution given a galaxy’s formation history. The ultimate goal is to enable multi-purpose, plug-and-play inference across a wide range of astrophysical tasks. Our approach aims to bridge the gap between simulation and observation while offering a scalable framework for future general-purpose scientific AI models.
Yuanyuan Zhang, NSF NOIRLab
Lessons Learned from a Simulation-Based Inference Approach for Galaxy Cluster Abundance Cosmological Analysis
We study the robustness of a simulation-based inference (SBI) method in the context of cosmological parameter estimation from galaxy cluster abundance in mock optical datasets. For this application, we train an SBI model, based on a mixture density network (MDN), to derive posteriors for cosmological parameters from a data vector that we construct using an analytic model for the galaxy cluster halo mass function and an analytic model for the observed richness (number of observed member galaxies) given halo mass and redshift. We compare the SBI posteriors to posteriors from an equivalent MCMC analysis that uses the same analytic form for the likelihood. We discuss the results from the SBI and MCMC analyses and lessons learned from the comparison.
The SkAI Institute is one of the National Artificial Intelligence Research Institutes funded by the U.S. National Science Foundation and Simons Foundation. Information on National AI Institutes is available at aiinstitutes.org.
