Day 2 Wednesday 24 June 2026

The application of Simulation-Based Inference (SBI) in galaxy evolution and cosmology provides a powerful implicit-likelihood framework to connect fundamental physics with complex observational data from telescopes such as Euclid, the Rubin Observatory, DESI, and JWST. However, the field faces a critical computational bottleneck: the massive simulation budget required to forward-model multi-scale phenomena, from the detailed sub-grid baryonic physics of galaxy formation to large-scale dark matter structure. This discussion session will synthesise recent advancements designed to dismantle these limitations. Drawing on successes from recent massive survey pipelines and proof-of-concept studies (e.g., emulation of the galaxy-halo connection or rapid spectro-photometric SED fitting), we will explore how novel sequential and active learning approaches iteratively target high-dimensional parameter spaces, reducing simulation budgets by orders of magnitude. We will also discuss strategies for maximising the utility of existing simulation suites, such as utilising multi-fidelity transfer learning, or reducing the need for data compression by deploying highly expressive neural density estimators. Furthermore, we will examine the critical role of neural emulators in bypassing expensive hydrodynamical runs altogether, alongside the practical trade-offs of migrating legacy CPU-based astrophysical codes to highly parallelised GPU architectures and differentiable programming frameworks. Ultimately, this session aims to collate the diverse computational tools currently at the community’s disposal, establishing community best practices and identifying which combinations of these approaches will most effectively drive the next generation of cutting-edge SBI analyses.

We introduce OASIS a scalable simulation-based inference framework designed for complex scientific simulators like those in the DREAMS cosmology project, a suite of thousands of state-of-the-art hydrodynamic and cosmological simulations aimed at understanding dark matter and galaxy formation by systematically varying astrophysical and cosmological parameters. OASIS explicitly embeds the noise model into the simulator and performs inference directly at the level of observed-data distributions. The method constructs a pseudo-posterior by utilizing maximum mean discrepancy (MMD) loss between the empirical distributions of the observed data and forward-simulated observations. We demonstrate application of this method in several parametric and non-parametric settings. Our SBI method builds inference around a lightweight projection of the data via a surrogate regression. Candidate parameters are evaluated by simulating small batches and assigning weights based on residual discrepancies with the projected data. The resulting self-normalized pseudo-posterior is computationally efficient, highly parallelizable, and requires only the fitted regression summaries, making it well suited to workflows that must repeatedly interrogate large simulation suites. We demonstrate the practical value of this approach in cosmological parameter calibration tasks using DREAMS data, where traditional likelihood-based techniques are impractical due to the simulator complexity and size. As the DREAMS papers have shown, statistical characterization of dark matter speed distributions and halo structural properties across large ensembles is crucial for interpreting astrophysical measurements and experimental constraints. Our lightweight SBI framework enables such calibration with significantly reduced computational overhead while offering insight into when low-dimensional projections suffice versus when richer summaries are needed to capture the underlying astrophysical variability.

Accurate modelling of galaxy populations is critical for unlocking the full potential of cosmological large-scale structure surveys, from quantifying selection effects to image-based calibration and photometric redshift estimation. In this talk, I will present GalSBI, an open-source framework for constructing highly realistic synthetic galaxy catalogs, constrained end-to-end with simulation-based inference. GalSBI combines parametric descriptions of luminosity functions, morphologies, and spectral energy distributions with SHAM-OT — a novel subhalo abundance matching scheme based on optimal transport — to jointly model the photometric and spatial properties of observed galaxy populations. Model parameters are inferred via an iterative Approximate Bayesian Computation scheme by comparing HSC and DES imaging data to forward-modelled image simulations incorporating instrumental, observational, and source-extraction effects, including PSF and background variations and source blending. The resulting framework simultaneously reproduces observed magnitudes, colors, sizes, redshift distributions, and clustering statistics with high fidelity. By linking intrinsic galaxy populations to realistic image-level observables within a unified inference framework, GalSBI provides a scalable foundation for precision cosmology in current and next-generation surveys.

Accurate modeling of baryonic physics remains a major challenge for precision cosmology due to our incomplete understanding of complex subgrid processes, like star formation and feedback from supernovae and active galactic nuclei below ~10 Mpc scales. This uncertainty leads to different hydrodynamical simulation suites to implement fundamentally different prescriptions for these unresolved physics. Current simulation-based inference approaches rely therefore on discrete sets of simulators, each encoding specific physical assumptions, making it difficult to robustly quantify theoretical uncertainties and learn about the underlying physics from observations. We introduce a machine learning framework that learns continuous representations of baryonic feedback across multiple simulation suites, to enable interpolation between different physical implementations while providing robust uncertainty quantification. Our approach addresses the key challenge of marginalizing over theoretical uncertainties represented by various simulators while simultaneously constraining the underlying baryonic physics from observations. We frame this as learning a shared continuous latent representation of the physics implemented across different simulators, allowing us to both marginalize over and constrain a continuous baryonic parameter space. Using the CAMELS simulation suite, we demonstrate our method on several baryonic fields including stellar mass, gas density, temperature, and pressure fields. This framework provides a path toward more robust cosmological inference by properly accounting for theoretical uncertainties in baryonic modeling while extracting maximum information about the underlying physical processes from current and future surveys.

Intrinsic alignments (IA) are one of the dominant astrophysical systematics for weak lensing surveys. However, higher-order IA models such as tidal alignment and tidal torquing (TATT) are formulated in three dimensions, whereas many simulations and analyses operate on projected two-dimensional quantities. This mismatch raises both practical and conceptual questions about how such models should be applied when only projected information is available. In this work, I present a field-level implementation of the TATT model applied directly to large-volume N-body simulations, with particular attention to projection effects and ensuring consistency between the three-dimensional model and its projected statistical observables. I quantify the impact of applying the alignment model before versus after projection, and examine how this choice affects the recovered power spectrum. By comparing the full three-dimensional implementation to projected reconstructions, I show that simplified pipelines can introduce biases in both the amplitude and scale dependence of intrinsic alignment signals. I highlight practical numerical and projection-related challenges that arise when modeling IA at the field level, which are important considerations for forward-modelling and likelihood-free inference methods applied to weak lensing and galaxy evolution.

I would like to present a hierarchical simulation-based inference (SBI) framework via compositional score modeling to reconstruct the Milky Way potential using stellar streams observations. Traditional SBI approaches often struggle with scalable aggregation of multiple observations, especially when inference targets include both global physical parameters (e.g., host-galaxy potential, ) and local nuisance parameters (e.g., progenitor or orbital parameters). Compositional score modeling offers a principled mechanism to approximate the score (gradient of the log posterior) of a high-dimensional target posterior by combining learned individual posterior scores associated with each observation, enabling scalable and sample-efficient inference without explicit likelihood evaluations.

How black holes accrete to super-massive monsters in centres of nuclear galaxies and in this luminous process impact galaxies is a key open question in galaxy evolution. The puzzle has remained in many pieces: Selection effects confuse, and most studies focus on two or three parameters. Yet, these indicate a multi-variate, probabilistic link between AGN accretion, obscuration, and the host galaxy's mass, star-formation, and morphology. A new window to cleanly select statistical samples to investigate these inter-dependencies opened with eROSITA. We aim to develop a scalable analysis of the X-ray to IR SED for millions of sources, to decode which host galaxies preferentially undergo a quasar phase and experience its energy release. Starting from the known evolving galaxy population, we trigger AGN with parametric probabilistic distributions and generate X-ray survey samples. To generate galaxy-AGN fluxes with realistic errors, we use the new GRAHSP SED fitting engine, validated to retrieve host galaxy mass even in bright AGN. To feasibly compute survey likelihoods we introduce a new machine-learning-based survey likelihood, supported by an explicit generative model and advances in deep neural network models. The likelihood allows building extensible models. In our application, we explore the mass and star-formation-dependence of AGN triggering, jointly with the geometry of the AGN accretion engine.

Simulation-Based Inference (SBI) is a cornerstone for parameter inference in scientific disciplines where simulators involve intractable likelihoods. While amortized generative models enable rapid posterior estimation, they are typically constrained by the specific priors used during training, which limits their flexibility when prior knowledge evolves. To overcome this dependency, PriorGuide has been proposed as an inference-time guidance method that attempts to adapt to new priors via reverse transition kernel approximations, making it suffer from systematic bias. Moreover, we find that PriorGuide can exhibit severe numerical instability (e.g., NaNs) in out-of-distribution (OOD) scenarios. To address these challenges, we propose a ratio-estimation-based test-time guidance framework. By leveraging an auxiliary neural ratio estimator (NRE) to learn a score guidance term, our method aims to decouple the inference process from the training prior. Preliminary results suggest that our approach improves robustness to OOD shifts and produces samples that align well with target posteriors. This framework shows promise as a tool for prior shift and sensitivity analysis, potentially allowing researchers to explore various prior updates with greater flexibility.

We present a simulation-based inference (SBI) framework for galaxy SED fitting based on conditional diffusion models, and, to our knowledge, the first application of diffusion-based SBI to this problem. Instead of learning a direct map from photometry to physical parameters, we model the joint distribution of observables and latent quantities, (p(\mathrm{mags}, \theta)), and cast inference as an inpainting task. In this formulation, any subset of entries can be treated as known and the remaining ones are sampled from the posterior, enabling a single trained model to perform both parameter inference and photometric reconstruction. This approach provides two practical advantages for SED fitting. First, it naturally supports conditioning on externally measured physical quantities (e.g. spectroscopic redshift) by simply marking them as observed variables in the joint state. Second, it handles missing photometric bands at inference time without retraining, imputation heuristics, or Monte Carlo resampling over missing data configurations. The same model can therefore accommodate heterogeneous observational coverage and variable data quality in a unified probabilistic framework. We demonstrate posterior inference for galaxy physical parameters together with inpainting of masked magnitudes, and show how posterior uncertainties broaden as photometric information is removed while remaining informative through learned cross-band and parameter-photometry correlations. The method yields full posterior samples, supports flexible conditioning, and is computationally efficient at inference once trained. More broadly, this work illustrates how diffusion models can serve as general-purpose SBI engines for high-dimensional astrophysical inverse problems with incomplete observations.

One of the biggest challenges facing the next generation of gravitational wave physics is tackling the LISA data analysis problem. There are expected to be thousands of overlapping sources buried in non-stationary, non-Gaussian noise, all of which need to be detected and analysed precisely so that parameter measurement errors do not cascade. Excellent progress has been made with traditional likelihood-based methods so far, but there is scope for tackling this challenge with SBI methods in a complementary way. I will present results from recent work using SBI to analyse a particularly troublesome class of binary black holes that we hope to observe with LISA: extreme mass ratio inspirals. I will reflect on the successes and limitations of this approach, as well as fruitful directions for improvements.

Simulation-based inference (SBI) is rapidly becoming a practical tool for galaxy-evolution problems, but its application raises recurring questions about robustness, model misspecification, scalability, and fair comparison across methods. This is exactly the sort of challenges and solutions this meeting aims to surface through discussion and practical sessions. In this discussion session, Chris Lovell and I will present a proposal for a community data challenge for SBI in galaxy evolution, and we will actively solicit input from participants to define a benchmark that is scientifically meaningful, technically tractable, and future-proof for Stage-IV data regimes. The core motivation is simple: we want to stress-test SBI pipelines beyond “in-model” validation by including deviations that mimic reality, like incomplete or wrong forward models, un-modelled selection effects, and domain shifts between simulations and data. We propose an open conversation around key design choices for the challenge, which includes: - What data products? photometry, spectroscopy, or both—and if spectroscopic, 1D spectra or IFU-like data? - What galaxy population model and volume? how complex should the simulator be, and what “truth” should be hidden vs disclosed? - What metrics define success? posterior calibration/coverage, parameter recovery under misspecification, posterior predictive checks, out-of-distribution detection, and others. - Speed vs robustness: should we explicitly reward amortized/fast inference, or prioritize reliability and robust uncertainty quantification? - Scalability: can the method realistically scale to Stage-IV data volumes and systematics? - Publication model: a coordinated paper series per method vs a single benchmark paper that collect the results from the various groups. - Timeline: how much time should we give to researchers before collecting the results?

Understanding galaxy formation and evolution is moving from a data-limited science to a methodology limited science with the advent of new, large photometric and spectroscopic surveys which will observe billions of galaxies over the next decade. Understanding the physics driving these systems often relies on SED fitting, which is a generally a complex, high-dimensional infernce problem with fundamental degeneacies between key parameters of interest. Fitting such models with traditional Bayesian techniques can take minutes to hours per galaxy per core, which is entirely unfeasible for ongoing and future surveys such as those of Euclid, Roman and LSST. This talk will explore Synference, a new Python framework aiming to accelerate the use of SBI approaches to SED fitting, which has been shown to produce comparable posterior constraints 1000-10,000x faster than traditional techniques. I will demonstrate the capabilities and results obtainable with Synfernce when applied to JWST and Euclid photometry. I will discuss other applications of Synference enabled by the flexibility and speed of SBI, including resolved, pixel-by-pixel SED fitting, model comparison and selection, and the application of Synference for constraints from spectroscopic confirmation.