Day 3 Thursday 25 June 2026

Next-generation cosmological surveys — LSST, Euclid, Roman — will be limited not by statistical power but by systematic uncertainties rooted in our understanding of galaxy populations. Photometric redshift calibration, intrinsic alignments, and baryonic effects on the matter power spectrum are all manifestations of the same underlying problem: cosmological inference requires an accurate model of the galaxy population and its evolution. I will present pop-cosmos, a simulation-based inference framework that learns the joint distribution of galaxy properties — redshift, stellar mass, star formation history, dust, metallicity, and black hole activity — from deep multiwavelength photometric data. The resulting generative model simultaneously delivers calibrated redshift distributions for weak lensing tomography, enables physically-motivated sample selection that mitigates intrinsic alignment contamination and unlocks multi-tracer clustering analyses, and provides improved Type Ia supernova standardisation through host galaxy property inference. It also reveals the astrophysical processes — including black-hole-driven feedback and the cessation of star formation — that shape the very populations we use as cosmological tracers. I will show applications spanning the Kilo-Degree Survey, DESI, the Zwicky Transient Facility, and COSMOS-Web, illustrating how a unified model of galaxy evolution connects to the next generation of cosmological constraints.

Simulation-based inference (SBI) enables amortized Bayesian inference by first training a neural posterior estimator (NPE) on prior-simulator pairs, typically through low-dimensional summary statistics, which can then be cheaply reused for fast inference by querying it on new test observations. Because NPE is estimated under the training data distribution, it is susceptible to misspecification when observations deviate from the training distribution. Many robust SBI approaches address this by modifying NPE training or introducing error models, coupling robustness to the inference network and compromising amortization and modularity. We introduce minimum-distance summaries, a plug-in robust NPE method that adapts queried test-time summaries independently of the pretrained NPE. Leveraging the maximum mean discrepancy (MMD) as a distance between observed data and a summary-conditional predictive distribution, the adapted summary inherits strong robustness properties from the MMD. We demonstrate that the algorithm can be implemented efficiently with random Fourier feature approximations, yielding a lightweight, model-free test-time adaptation procedure. We provide theoretical guarantees for the robustness of our algorithm and empirically evaluate it on a range of synthetic and real-world tasks, demonstrating substantial robustness gains with minimal additional overhead.

The weak lensing shear field encodes substantially more cosmological information than standard summary statistics such as the power spectrum. Field-level inference can fully exploit this information, but physical realism at the field-level requires very high fidelity simulations. This poses a major challenge for SBI: extracting maximal information with deep-learning-based neural compression requires many training simulations, but achieving physical realism at the field level makes each simulation extremely costly. We demonstrate that multifidelity SBI can alleviate this tension by dramatically reducing the number of high-fidelity simulations needed for accurate cosmological inference. We pre-train neural compression models on KiDS-legacy-like shear mocks using fast log-normal GLASS simulations and fine-tune them on a small set of high-fidelity $N$-body simulations. We show that as few as 60-100 high-fidelity simulations are sufficient to obtain informative and well-calibrated cosmological posteriors, enabling an order-of-magnitude reduction in simulation cost for accurate field-level inference in a realistic setting.

Recent and upcoming large-scale spectroscopic surveys such as DESI, WEAVE, and 4MOST are providing an unprecedented view of galaxies across a wide redshift range. This vast amount of data poses a challenge for analysis with traditional tools. At the same time, it coincides with a revolution in machine learning, particularly deep learning and foundation models, which are uniquely capable of absorbing this wealth of information. In this context, we have leveraged data from DESI DR1 to develop a foundation model that encodes fundamental information from the stellar populations and gas emission in galaxies. We demonstrate that the resulting aligned embedding space can be used to robustly and efficiently obtain properties such as color, emission-line equivalent widths, and even fundamental galaxy properties such as mass and age. More importantly, it can be used to reconstruct emission lines using continuum input alone, providing a fast, reliable, and purely empirical alternative to codes such as Cloudy, thereby establishing a powerful connection between stellar and gas physics. Finally, we show how this approach is being integrated into an SBI pipeline to infer star formation histories and galaxy properties, enabling deeper insights into galaxy evolution.

Inferring the physical conditions of the ionised interstellar medium (ISM) from nebular emission lines is central to understanding star formation and chemical enrichment in galaxies. Yet, it remains a challenging inverse problem, especially as modern spectroscopic surveys deliver data at unprecedented scale and complexity. Hence, I present two frameworks that leverage machine learning and physically motivated forward simulations to analyse spectroscopic observations of the ionised ISM in local galaxies. First, I show how we use synthetic spectra generated with photoionisation models under realistic observational conditions to train a neural network to identify different nebular regions and separate them from the surrounding diffuse ionised gas in integral field data cubes. By training directly on simulated data, the model learns robust spectral features beyond simple line-ratio thresholds. Consequently, the model defines physically meaningful nebular boundaries that account for line-of-sight superpositions. Additionally, we apply interpretability methods to identify which spectral features drive the network’s classification, allowing us to both verify that the model relies on physically meaningful emission lines and explore potential new diagnostics. Building on a similar forward-modelling method, I then present a simulation-based inference framework to constrain ISM physical conditions, delivering full posterior distributions for metallicity, ionisation parameter, and relative abundances without requiring an explicit likelihood. The retrieved posterior distributions are consistent with results from traditional Bayesian approaches, such as Nested Sampling, while the increased inference speed enables efficient analysis of modern spectroscopic surveys. Together, these machine learning techniques offer a promising route to extract detailed ISM physics from modern IFS campaigns probing the star formation and feedback scales in the local Universe.