Day 1 Tuesday 23 June 2026

Modern galaxy surveys and simulation-based inference pipelines increasingly require tools that can connect large datasets, machine learning models, domain expertise, and heterogeneous contextual information in an interactive and reproducible way. AstronomicAL was originally developed to address this need through active learning: allowing users to visualise catalogues, inspect informative or uncertain sources, update labels, and iteratively improve classification models with expert input. In this talk I will present the next stage of AstronomicAL: a plugin-based platform for building scientific pipelines around data, models, simulations, and expert decisions. The motivation is that many research workflows do not fail because a model cannot be trained, but because the work around the model is fragmented. In practice, these surrounding steps are often scattered across scripts, archive queries, manual inspection, and informal records. AstronomicAL aims to bring these pieces into a shared workspace, where model outputs, contextual evidence, annotations, and follow-up decisions become connected parts of the same reproducible workflow. This changes AstronomicAL from an active learning dashboard into an interactive workbench for scientific analysis. In astronomy, this could mean moving from an unexpected model prediction or unusual SBI posterior to a connected evidence-gathering process that helps researchers decide whether to trust a model, retrain it, refine the simulation design, or pursue observational follow-up. More generally, the same architecture can support workflows in which researchers compare competing models through their failure modes, steer expensive simulations, reconcile conflicting data sources, or construct benchmark samples. By making these steps composable through plugins, AstronomicAL aims to turn fragmented research code into reusable, auditable workflows guided by domain expertise.

Projects such as the Vera C. Rubin Observatory and Euclid offer a huge opportunity to study galaxies and their evolution across wide areas of the sky. However, this is only possible if we are able to efficiently and accurately model our observations of the galaxies, and thus infer their redshifts and physical properties from limited photometric data (typically 4-10 bands). The pop-cosmos framework is a stellar population synthesis (SPS)-based generative model for the galaxy population, built around a machine-learned prior (a score-based diffusion model) over 16 galaxy-level parameters, and trained on 26-band photometry taken from the Cosmic Evolution Survey (COSMOS). In this tutorial, we will demonstrate several aspects of the public pop-cosmos codebase (https://github.com/Cosmo-Pop/pop-cosmos): (i) using the population model as a forward simulator of galaxy photometry and spectral energy distributions (SEDs); (ii) performing galaxy-level Bayesian inference with the trained diffusion model as a prior; and (iii) amortizing galaxy-level inference with flow-based models built using our public library (https://github.com/Cosmo-Pop/flowfusion).

Galaxy growth has traditionally been explored through two complementary perspectives: the build-up of stellar mass inferred from galaxy integrated SED modelling, and the growth in size captured by high-resolution imaging. Together, these methods have driven substantial progress in our understanding of galaxy evolution. However, SED fitting typically relies on simplified assumptions, such as a foreground dust screen and a non-varying stellar population, limiting the physical realism of inferred properties. In parallel, increasingly rich datasets on high-redshift galaxies (from UV to mm) gathered from deep multi-wavelength observations suffer from differing spatial resolutions, resulting in a mixture of highly resolved, marginally resolved and unresolved data. To address both limitations, we developed a novel modelling technique aimed at maximally exploiting all of this rich yet varied information without sacrificing unresolved or poorly resolved bands, and without sacrificing resolution where available. Our modelling approach is fundamentally 3D in nature, and extracts information on stellar populations and their radial variations as well as star-dust geometries, without imposing an attenuation law a priori. We jointly model observational constraints on wavelength dependent fluxes, sizes, Sérsic profile shapes, and axial ratios using a computationally efficient machine learning emulator trained on an extensive toy model library of dust radiative transfer calculations. Applications to TNG50 galaxies and the COSMOS/GOODS-South fields will be discussed in detail.

Empirical galaxy formation models offer a powerful complement to more physics-first simulations: they are flexible, observation-driven, and computationally tractable enough to enable rigorous statistical inference. We present UniverseMachine (UM) DR2 , a major update to the pre-JWST UniverseMachine framework (Behroozi et al. 2019). UM DR2 now matches a comprehensive compilation of observational constraints spanning z ∼ 0–20, including the latest JWST observations at z > 10 that challenged earlier models. Now the model simultaneously reproduces key observable statistics, like stellar mass functions, UV luminosity functions, bursty star formation rate distributions, cosmic star formation rate density, and quiescent fractions across the full observed redshift range. The model encodes the galaxy-halo connection through a minimal, physically motivated parametrization of star formation rates as a function of halo mass, mass accretion rate, and redshift, preserving cosmological consistency while remaining sufficiently flexible to capture the uncertain physics. We leveraged simulation-based inference to perform posterior estimation over the full model parameter space, enabling rigorous uncertainty quantification and principled model comparison that would be intractable with traditional sampling approaches. I will present the current calibrated model and discuss about the physical drivers of galaxy growth and quenching across cosmic time. We provide public catalogs that will enable forward-modeling and SBI applications with JWST, Roman, Rubin LSST, and PFS.

Ultra-Diffuse Galaxies (UDGs) are an extreme population of low surface brightness galaxies that challenge our understanding of galaxy formation and evolution. Despite their faint stellar component, many are rich in neutral hydrogen (HI), making them ideal targets for kinematic studies. However, the limited spatial resolution, low rotation speeds, and uncertain inclinations typical of UDGs place these systems in a regime where traditional tilted-ring kinematic models are poorly constrained, complicating efforts to determine their dark matter content and evolutionary pathways. Using new VLA HI observations together with forward-modelled synthetic data cubes, we are developing diffusion models as a simulation-based inference approach to recover galaxy kinematics directly from HI data. Trained on simulated observations, these models provide a framework for inferring kinematic parameters and their uncertainties in regimes where traditional methods are uncertain. In this talk, I will present the VLA HI observations, preliminary kinematic modelling, and the development of diffusion-based inference for HI-rich UDGs.

Simulation-based inference (SBI) enables fast, well-calibrated posterior estimation for high-dimensional inverse problems in astronomy. We present SBIPIX, an amortized neural density estimation framework that performs full Bayesian inference on galaxy spectral energy distributions at the pixel level ($\sim 10^{-4}$ s per pixel)---orders of magnitude faster than traditional MCMC methods while maintaining accurate uncertainty quantification. SBIPIX leverages FSPS (Flexible Stellar Population Synthesis) forward models to generate synthetic training datasets with both parametric and non-parametric star formation histories \citep{Leja2019, Iyer2017}. Validation against MILES-based synthetic galaxies \citep{Vazdekis2010} demonstrates robust recovery of stellar mass, age, dust attenuation, and SFR across $z \approx 0.5$--$10$. The amortized network generalizes efficiently to new observations, enabling application to millions of pixels in JWST imaging. We apply SBIPIX to JADES (JWST Advanced Deep Extragalactic Survey; \citealt{Eisenstein2023}) multiband NIRCam and HST/ACS data, generating spatially-resolved physical property maps for thousands of galaxies. This enables two science applications: (1) identifying star-forming substructures and clumps in galactic disks with unprecedented precision using resolved stellar mass and SFR, and (2) using Synthesizer \citep{Lovell2025, Roper2025} with pristine stellar population synthesis models (Yggdrasil; \citealt{Zackrisson2011}) to quantify Population III detectability limits using SBI-derived posteriors, demonstrating how amortized SBI accelerates Bayesian analysis in extragalactic surveys and enables scientific discoveries at sub-galactic scales.

The depth, resolution, and sheer volume of high-redshift spectra from JWST and many more large scale surveys have outpaced the models we use to interpret them. Simulations and observations agree that real star-formation histories (SFHs) are more complex than parametric models can describe, with phases of star bursts, mini-quenching and rejuvenation. Flexible, non-parametric models address this, but at the cost of a steep rise in parameter dimensionality that strains conventional inference. I present CERIDWEN, a JAX-based, GPU-accelerated SED-fitting framework built around a fully differentiable forward model. Because the simulator is differentiable and vectorized, it sustains a roughly constant evaluation throughput as model complexity grows, making it tractable to fit hundreds of SFH bins, time-varying chemical-enrichment histories, and modular per-age-bin dust, recovering these jointly from a single spectrum. On real JADES galaxies it reproduces the state-of-the-art SED fitting code Prospector at a large reduction in wall-clock cost. Beyond its current gradient-based samplers, this fast differentiable simulator is natural infrastructure for simulation-based inference: a source of cheap simulation budgets, and a problem where exact posteriors can validate amortized methods. I will close with an open question for discussion: how best to bring a physically realistic but expensive, non-differentiable nebular model into an SBI framework.

Learning cosmology from galaxy surveys demands large simulation suites spanning cosmological and astrophysical parameters, yet hydrodynamical galaxy-formation simulations are prohibitively expensive. Semi-analytic models are cheaper and physically grounded, but still require N-body merger trees, and densely sampling parameter space remains costly. We extend cosmological rescaling to operate directly on merger trees in the Ωₘ–σ₈ plane, running the Santa Cruz semi-analytic model on the rescaled trees to generate galaxy populations at new parameters for negligible cost. We further demonstrate that gradients can be propagated through the existing merger trees, opening a route to gradient-based inference. A halo-profile-based correction, set by a single free parameter, suppresses rescaled-mass bias below the per-cent level. For inference of Ωₘ and σ₈ from the stellar mass function or two-point correlation function, as few as 64 base simulations rescaled to ~1000 samples match 750 dedicated N-body runs; rescaling to 3200 improves Ωₘ by ~25%. Rescaling one CAMELS-SAM simulation costs ~0.1 CPUh versus several thousand.