Day 4 Friday 26 June 2026

Modelling the connection between galaxies and their host dark matter halos is a fundamental task in galaxy evolution. The modern use of Graph Neural Networks has allowed a deeper exploration of these relationships on both a galaxy-by-galaxy and population level, demonstrating that galaxies, halos, and their relationships are greatly influenced by both their detailed spatial environments and temporal assembly history. In this talk, I demonstrate a full equivalence in the impact of halo environments and assembly histories on a broad range of baryonic galaxy properties. This result holds when simulating galaxy properties using both full magnetohydrodynamic codes or semi-analytic models, demonstrating that the equivalence is driven by the effect of dark matter on galaxies. To achieve the equivalence, the environment cannot be expressed as spherically averaged density, but must be encoded as a graph, linking halos on sufficiently large spatial scales. We furthermore measure the linking lengths that give optimal predictions for each galaxy property, and show that these are directly related to the typical extent of all halo progenitors. The equivalence offers a tantalizing opportunities for Geometric Deep Learning Simulation-Based inference of the assembly histories of host halos in the real Universe. I will finish by showing some early results demonstrating our GNN-SBI’s capability to directly infer key events in the formation of dark matter halos.

Semi-analytic models (SAMs), such as L-Galaxies, play a vital role in modelling galaxy populations across cosmological volumes. These models evolve baryonic physics on pre-computed dark matter halo merger trees and achieve much lower computational cost than full hydrodynamical simulations. This efficiency supports robust parameter inference via methods such as Markov Chain Monte Carlo (MCMC), enabling extensive model calibration and comparison with observations. However, as SAM physics become more complex and parameter spaces expand, run times increase, making traditional inference increasingly challenging. We present a new optimised, parallel C++ implementation of L-Galaxies (Yates et al., 2024) that we have made fully differentiable. We ensure differentiability by using operator overloading for forward- and reverse-mode automatic differentiation (AD). By structuring the SAM as a computational graph, we enable process-level parallelism and enable efficient reverse-mode AD. By exposing exact gradients via AD, our framework enables gradient-based inference methods such as Hamiltonian Monte Carlo (HMC), which deliver higher sampling efficiency and faster convergence in complex, high-dimensional parameter spaces. Our differentiable, parallel SAM will be used to generate extensive galaxy catalogues efficiently, and to directly constrain model parameters using Bayesian inference against super-high-resolution hydrodynamical simulations — demonstrating a scalable gradient-aware approach to simulation-based inference for galaxy evolution.

We generate synthetic observations from the SC-SAM semi-analytic model to directly connect galaxy formation physics with observable galaxy properties. Using Synthesizer, we develop a flexible, physically motivated forward-modelling pipeline that incorporates a suite of galaxy dust prescriptions spanning a range of grain-size distributions, dust masses, and compositions. We demonstrate that this flexible dust modelling produces colours and luminosity functions that differ from those obtained using fixed dust models, with the largest discrepancies occurring in the UV. By embedding this framework within a SBI pipeline, observational data could be used to constrain dust physics. Future work will extend the framework to include self-consistent modelling of AGN emission and dust attenuation.

Understanding galaxy evolution requires disentangling the nonlinear connections between galaxies, dark matter, and their environments across cosmic time. We present an interpretable machine-learning framework that quantifies two complementary definitions of environment. First, we define the spatial environment as the distribution of neighbouring galaxies at z = 0. Using hydrodynamical simulations, we train a neural network to predict galaxy properties from host and neighbouring subhalo properties. Explainable AI analysis then quantifies environmental parameters driving secondary dependencies beyond halo mass, as a function of galaxy type and mass, — providing physical insight into the environmental effect of galaxy formation. Second, we define the primordial environment as the initial dark matter field that causally encodes present-day galaxy properties through structure formation. We train a convolutional neural network that maps the initial density and velocity fields at z = 20 directly to z = 0 galaxy properties. Explainable AI analysis reveals which spatial regions and patterns in the initial conditions carry predictive information for baryonic observables, extending recent CNN-based halo-mass prediction studies to galaxy properties. This forward-modelling approach — from initial conditions to galaxy observables — is relevant to field-level inference with SBI. Our results offer quantitative, interpretable insights into the nonlinear encoding of initial conditions in galaxy observables.

Understanding the assembly history of the Milky Way requires precise stellar ages across large populations. These ages help to disentangle different merger events and constrain galactic chronology. Large-scale spectroscopic surveys are central to this effort: DESI and Gaia XP have already delivered stellar parameters (e.g., Teff and [Fe/H]) for millions of stars, and upcoming surveys like WEAVE and 4MOST will expand this dataset further. These parameters break degeneracies around the main-sequence turn-off and subgiant branch, enabling precise age-determination via isochrone fitting in these regions. However, traditional Bayesian methods require running iterative sampling algorithms independently for each star, creating a computational bottleneck when scaling to millions of targets. Simulation-based inference offers a solution to this problem through amortised inference. In this work, we train a neural posterior estimator on MIST isochrones that can be deployed across different surveys with varying noise levels, bypassing the per-star computational cost of conventional methods. We present preliminary validation results to demonstrate the feasibility of population-scale age inference across current and future spectroscopic surveys.

Euclid will observe close to one billion galaxies across a wide range of cosmic time, providing opportunity to investigate the processes behind galaxy evolution and the influence Active Galactic Nuclei (AGN) have on their host systems. Maximising the scientific insights from such large datasets requires inference frameworks that can quickly and efficiently connect observations to theoretical models. AGN are known to play an important role in suppressing galaxy and stellar formation, providing feedback to their local surroundings. In order to understand their role and impact in galaxy evolution, detailed modelling is important to capture all the physics involved. To reflect the data gathered by Euclid, this modelling needs to take place within cosmological scale hydrodynamical simulations. However, at these scales numerical resolution limitations prevent AGN from being resolved directly, requiring their feedback to be implemented through sub-grid models. The parameterisation and calibration of these models strongly affect simulated galaxy properties, leading to modelling uncertainties that limit their comparison with survey data. By combining cosmological simulations and the software package Synthesizer to generate realistic synthetic observables, I plan to use simulation based inference and Euclid data to directly compare cosmological simulation predictions to observational data, with a particular focus on the AGN sub-grid physics implemented. I will outline a general framework for applying SBI techniques to large extragalactic surveys such as Euclid, highlighting their potential to advance our understanding of the role AGN play in galaxy evolution and cosmological simulations.

Galaxy mergers play a central role in the hierarchical assembly of galaxies, driving bursts of star formation, triggering AGN activity, and reshaping galaxy structure. However, robustly identifying mergers and distinguishing their evolutionary stages (pre-merger and post-merger) from imaging data alone remains challenging, particularly in the era of wide-field surveys with millions of galaxies. We present a supervised deep learning framework trained on mock IllustrisTNG galaxy images forward-modeled to match Hyper Suprime-Cam (HSC) observations (Margalef-Bentabol et al. 2024). Using convolutional neural networks (CNNs), we develop a three-class classifier (non-merger, pre-merger, post-merger). To enhance robustness and mitigate architecture-dependent biases, we train multiple CNN backbones and combine them within an ensemble framework. This ensemble strategy leverages complementary feature sensitivities across models, significantly improving classification accuracy and reliability. We evaluate performance on both synthetic and real HSC data, assessing domain transfer and generalizability, key challenges for simulation-based inference in galaxy evolution. The resulting framework produces science-ready merger catalogs for the North Ecliptic Pole (NEP) field, enabling statistically robust studies of merger-driven star formation, black hole growth, and structural transformation.

Understanding the physical state of the circumgalactic medium (CGM) is central to galaxy evolution, but we can only primarily probe it indirectly through absorption lines in background quasar spectra. The standard approach of fitting Voigt profiles and running ionization models works, but it is fundamentally degenerate: very different gas configurations can produce nearly identical absorption features, blended components are hard to disentangle, and propagating uncertainties through the full pipeline is difficult. In this work, we develop a SBI framework that learns to infer multiphase CGM gas properties directly from spectra, trained entirely on synthetic data from the SALSA catalog. SALSA provides millions of sightlines through IllustrisTNG, SIMBA, and EAGLE, each with a realistic synthetic spectrum and full knowledge of every gas cell the sightline intersects. We group these gas cells into physically coherent clouds by clustering in velocity, density, and temperature space, giving us supervised labels for each spectrum: the number of clouds, their individual properties (density, temperature, metallicity, column density, velocity, path length), and halo-scale environmental quantities. Our model uses neural posterior estimation with a 1D convolutional or Transformer encoder feeding into conditional normalizing flows, with physics-informed priors on thermal broadening, density–temperature relations and doublet ratios. We plan to validate first on SALSA sightlines and then apply the framework to real HST/COS data from CGM surveys, benchmarking against traditional Voigt-profile and photoionization analyses. This extends SBI into the structured, multiphase CGM regime.

We introduce a simulation-based inference framework built upon the DREAMS Project, a new suite of 1,024 cosmological hydrodynamical zoom-in simulations of Milky Way-mass halos [2512.00148]. Designed to systematically disentangle theoretical uncertainties in sub-grid physics from intrinsic halo-to-halo variance, this suite varies key astrophysical parameters governing supernova wind energy, wind speed, and AGN feedback efficiency within the IllustrisTNG model. We introduce a novel observational weighting scheme, applying pseudo-posterior constraints based on the empirical stellar mass-halo mass relation [2602.03613]. This inference pipeline reveals broad degeneracies in the fiducial feedback parameters, demonstrating that standard single-model tuning misses complex parameter interdependencies. Using these constraints, we quantify that a Gaia-Sausage-Enceladus-like merger history is exceptionally rare and drives specific structural shifts, though immense halo-to-halo scatter persists. For satellites, our emulator demonstrates that while sub-grid variations regulate stellar mass, intrinsic variance overwhelmingly dominates population statistics [2512.02095]. We also identify a persistent tension where the inferred parameter space fails to reproduce the extended half-light radii of massive satellites observed in the SAGA survey. To determine if this size-mass tension is an artifact of the TNG sub-grid model or a broader theoretical challenge, we will also present preliminary results from ongoing extensions to the DREAMS project framework. This includes three new suites of high-resolution dwarf galaxy simulations run with FIRE3, RAMSES, and ChaNGa. Furthermore, we will introduce new mass-varied and resolution-varied simulation suites utilizing the IllustrisTNG and FIRE3 models. These suites extend the halo mass range to include masses from 10^9 to 10^14, allowing us to recreate uniform box results from a suite of zoom-in simulations.

Modern galaxy surveys are increasingly photometry dominated with wide field imaging campaigns observing orders of magnitude more objects than spectroscopic surveys can feasibly target. While spectroscopy provides physically interpretable constraints on stellar populations, star formation histories, metallicities, and dust, its limited coverage creates a bottleneck for population level inference. This imbalance demands new methodologies that transfer spectroscopic information into the vastly larger imaging domain. This work develops a likelihood free, image based inference framework that directly maps observed multiband galaxy images to posterior distributions over spectroscopically derived physical properties. Only recently have hydrodynamical simulations begun to produce galaxy morphologies that resemble those observed in modern surveys. Rather than relying on simulated images as an intermediary, we instead learn the mapping directly from real data. The approach operates on pixel level imaging information, allowing morphology and structural features to act as primary constraints on physical properties.