SBI4GALEV — Talks

The final day of the conference focused on the practical application of Simulation-Based Inference (SBI) and deep learning to bridge the gap between complex simulations and observational data. Presentations spanned a wide range of scales, from the internal properties of stars and the circumgalactic medium to the large-scale cosmic web and galaxy merger histories. A recurring theme was the pursuit of computational efficiency, with several talks demonstrating how neural posterior estimators and differentiable simulations can replace slow traditional sampling methods. The day concluded with a look at the broader infrastructure of AI, discussing self-hosted models for education and ambient AI for scientific synthesis.

• Connecting Environment and Assembly to Measure the Formation Histories of Dark Matter Halos With GNN-powered SBI — Christian Kragh Jespersen
• Introducing L-Galaxies AD: an auto-differentiable galaxy evolution simulation — Andrew Green
• Exploring the Impact of Dust on Galaxy Luminosity Functions with Semi-Analytic Models — Sophie Newman
• Quantifying Environmental Information in Galaxy Evolution with Interpretable Machine Learning — Shun-ya S. Uchida
• Fast Stellar Age Inference for Galactic Archaeology with SBI — Shichen Su
• Simulation Based Inference for Understanding the Role of AGN in Galaxy Evolution — Thomas Bebbington
• Deep learning approaches for galaxy merger identification and classification: bridging simulations and observations — Subhrata Dey
• Inferring CGM Cloud Properties Using Simulation-Based Inference — Tanmay Singh
• Disentangling Feedback and Variance in 1,024 Milky Way-Mass DREAMS Simulations — Jonah Rose
• Image-Based Inference: Bridging Photometric Surveys and Spectroscopic Constraints — James Kostas Ray
• Self-Hosted AI for Education — Sotiria Fotopoulou
• Ambient AI for Scientific Meetings — Chris Lovell

Day 3 focused on the practical application of simulation-based inference (SBI) and machine learning to bridge the gap between complex astrophysical simulations and observational data. The sessions spanned a wide range of scales, from the global cosmological constraints of weak lensing and galaxy populations to the detailed physical conditions of the interstellar medium. A recurring theme was the optimization of these models, specifically through techniques to reduce computational costs, handle model misspecification, and leverage data-driven embeddings.

• Galaxy Evolution Meets Cosmological Inference — Hiranya Peiris
• Minimum Distance Summaries for Robust Neural Posterior Estimation — Sherman Khoo
• Field-level cosmology with <100 simulations using multifidelity SBI: a proof-of-concept for KiDS-legacy weak lensing — Alex Saoulis
• Connecting Stars and Gas: a data driven approach to modeling gas emission based on stellar continuum using DESI — Eduardo Albuquerque Hartmann
• Mapping Nebular Structure and ISM Conditions with SBI — Caterina Bracci

The second day of SBI4GALEV focused on scaling simulation-based inference to meet the demands of high-dimensional astronomical data and computationally expensive forward models. Presentations spanned a wide range of applications, from galaxy clustering and SED fitting to gravitational wave analysis for LISA, with a strong emphasis on reducing the simulation bottleneck. A significant portion of the day was dedicated to the emergence of diffusion models as flexible alternatives to normalizing flows for posterior estimation and data inpainting. The day concluded with a call for community standardization through a proposed SBI data challenge and the introduction of the Synference framework.

• Addressing the Simulation Bottleneck in Simulation-Based Inference — Maximilian von Wietersheim
• OASIS: A Robust Framework for Population Inference in Astronomy — Arya Farahi
• GalSBI: Forward Modelling Galaxy Clustering and Population — Silvan Fischbacher
• Continuous Representation of Hydrodynamic Simulations for Robust SBI — Ming-Shau Liu
• Implementing 3D Intrinsic Alignment Models in Projected Simulations — Leonor Simões
• Diffusion Models for Simulation-Based Inference — Giuseppe Viterbo
• Comparing Bayesian Inference Methods in Astronomy: HMC, Nested Sampling, and SBI — Johannes Buchner
• Test-time inference for Prior Change via Density Ratio Estimation — Yichen Zang
• Diffusion based SED fitting — Grégoire Aufort
• SBI for LISA — Philippa Cole
• Simulation-Based Inference (SBI) Data Challenge for Galaxy Evolution — Luca
• Synference Tutorial — Thomas Harvey

The first day of SBI4GALEV focused on the integration of Simulation-Based Inference (SBI) and differentiable programming to accelerate galaxy evolution research. Presentations spanned the development of high-performance JAX-based frameworks and the use of neural emulators to bypass the computational costs of radiative transfer and stellar population synthesis. A strong emphasis was placed on moving beyond traditional SED fitting toward more complex, spatially resolved, and physically nuanced models of galaxy structure and kinematics.

• Introduction to Likelihood-Free Inference — Tom
• Astronomic: A Plugin-Based Platform for Astronomical Data Analysis and Active Learning — Grant Stevens
• Pop-Cosmos: A Generative Model for the Galaxy Population — Stephen Thorp
• SE3D: A Novel Modeling Technique for Chromatic Resolved Galaxy Observations — Stephen Ramnichal
• UNiverse Machine 2 and Tengri: Differentiable Frameworks for Galaxy Formation and Inference — Suchetha Cooray
• Diffusion-Based Inference of HI Kinematics in Ultra-Diffuse Galaxies — Lawrence Faria
• Decoding the SED: Disentangling galaxy physics from the rest-frame optical — Patricia Iglesias-Navarro
• Differentiable SED Simulation with Ceridwen — Amanda Stoffers
• Cheap, physics-informed training data for SBI — Richard Stiskalek
• Degeneracy Distillery: Learning Information Geometry for Simulation-Based Inference — Lucas Makinen