Conference

Galaxies are the primary tracers of the large-scale structure of the Universe and are traditionally used through summary statistics such as correlation functions and power spectra to constrain cosmological models. However, galaxy populations themselves contain rich information through their spatial distribution, environments, and internal properties, potentially extending beyond the information captured by standard summary statistics. In this talk, I will present a series of machine learning approaches designed to extract cosmological information directly from galaxy catalogs. I will discuss work using graph neural networks (GNNs) to infer cosmological parameters from simulated galaxy populations, including studies demonstrating robustness to observational effects and domain shifts across semi-analytic and hydrodynamical galaxy formation models. I will also present recent work exploring the cosmological information content of individual galaxies using symbolic regression, as well as ongoing efforts based on probabilistic generative models. Together, these results suggest that galaxy populations may encode cosmological information in ways that complement traditional large-scale structure analyses, opening new avenues for field-level and simulation-based inference in upcoming surveys

We perform for the first time full simulation-based inference on the Lyman-$\alpha$ forest 1D power spectrum. In particular, we consider the prediction of the Lyman-$\alpha$ forest $P_{\rm 1D}(k)$ at $2.0

The future of large-scale structure is at high redshift. In particular, the surveys of the next decade that will be at the vanguard of precision cosmology will target star-forming galaxies at z>2. Field-level inference efforts will, therefore, inevitably turn toward high-redshift samples in the near future, the clustering of which is not yet well-understood. The computational intractability of hydrodynamical simulations means that EFT-bias models at the field level will be especially compelling choices of field-level inference forward models for these samples. I will present my recent work measuring the EFT-bias parameters of simulated high-redshift galaxies at the field-level for the first time. I will also discuss how these results can be leveraged toward building simulation-based priors, which aid current field-level inference efforts.

Current and future high-redshift spectroscopic surveys such as DESI, DESI-II, and Spec-S5 raise the question of how to fully extract the information contained in these datasets. Field-level inference is opening a new frontier in cosmology by enabling analyses that capture the full information content of cosmological observations, rather than relying on two- or three-point summary statistics. High-redshift surveys probe unprecedented cosmological volumes, access quasi-linear modes sensitive to fundamental physics, benefit from precise calibration against small-scale hydrodynamic simulations, and exhibit diminishing shot noise toward higher redshift. However, accurately modeling these observations — particularly at the field level — remains a major challenge. I will present recent progress in modeling the Lyman-alpha forest at the field level and compare the information content of joint power spectrum and compressed bispectrum analyses with field-level inference approaches, combining methods from effective field theory (EFT) and machine learning (ML). This framework is then extended to additional high-redshift tracers, including Lyman-break galaxies and Lyman-alpha emitters, whose cross-correlations unlock new opportunities for extracting cosmological information and controlling systematic uncertainties. Finally, I will present ongoing work that combines perturbative mock generation with generative ML techniques to bridge large and small scales, enabling simulations that span DESI-like cosmological volumes while retaining the small-scale structure captured by hydrodynamic simulations. Together, these developments provide a scalable path toward fully exploiting next-generation surveys such as DESI-II and Spec-S5, establishing high-redshift structure as a key arena for precision, field-level cosmology.

Polarized Galactic emission is the foremost challenge for searches for a background of primordial gravitational waves imprinted in the polarization of the CMB. We argue that current methods struggle to address this challenge, either by being overly susceptible to model misspecification, or by failing to properly propagate the uncertainty due to residual Galactic emission after foreground cleaning. To address these issues, we explore a novel analysis framework for parameter inference with large-scale CMB polarization data. Our method combines simulation-based inference with the needlet internal linear combination (NILC) algorithm to compress the data into a summary statistic that is robust to model misspecification and small enough for neural posterior estimation with normalizing flows. We show that the semi-blind nature of the NILC-based compression significantly increases robustness to mismodeling of the anisotropic and non-Gaussian properties of the foreground fields. Using an idealized ground-based setup inspired by the Simons Observatory Small Aperture Telescopes, we demonstrate improved statistical constraining power for the tensor-to-scalar ratio r and improved robustness to complex foregrounds compared to other techniques in the literature. Trained on a semi-analytical foreground model, the method yields unbiased results across a range of PySM simulations, including the high-complexity d12 model, for which we obtain r=(1.09±0.27)⋅1e−2 for input r=0.01 and sky fraction fsky=0.21. Our results highlight the importance of designing data compression schemes for SBI that prioritize robustness to model misspecification over statistical optimality, and demonstrate the feasibility and advantages of a complete maps-to-parameters simulation-based analysis of large-scale CMB polarization for current ground-based observatories.

I will present the cosmological analysis from the simultaneous Bayesian estimates of gravitational-lensing potential bandpowers and unlensed cosmic microwave background (CMB) EE bandpowers directly using the polarization maps from the South Pole Telescope (SPT) observed in 2019/20. These observations produce the deepest high-angular-resolution CMB polarization maps at 90, 150, and 220 GHz to date, making the standard Quadratic Estimation (QE) method suboptimal for lensing reconstruction. In this analysis, we use the Marginal Unbiased Score Expansion (MUSE) method, which is an optimal map-level Bayesian inference method for CMB lensing potential bandpowers and unlensed CMB EE bandpowers, effectively using all N-point statistics of the CMB polarization maps. The constraints on the Hubble constant (H0) and the amplitude of structure growth (S8) from this work are comparable to those from Planck using full-sky temperature and polarization observations, enabling a powerful test of the LCDM model. With the lensing potential bandpowers reconstructed from the CMB polarization signal, we test the anomaly of excess lensing power from the LCDM prediction, and detect the impact of non-linear structure evolution on CMB lensing. We also explore the extensions of the LCDM models.

The work presented in this talk is part of a program of performing field-level cosmological analyses using galaxy and weak lensing observables in density slabs of $O[100\ \mathrm{Mpc}]$. The inference pipeline requires a model for the mass overdensity field which is accurate to scales of a few Mpc across cosmologies, and also fast in the generation of sample fields. I will describe a scheme in which the nonlinear mass overdensity field is obtained by applying some point transformation to a Gaussian random field. I will present our transformation function that, with a small number of parameters, characterizes the mass distribution in a slab for different redshifts, slab widths, and cosmologies. I will show that the model accurately reproduces the statistical properties of mass overdensity fields from Gower Street simulations, demonstrating its viability for field-level inference. Finally, I will discuss a validation test which uses the point-transformed Gaussian model to infer the cosmology from the output of an $N$-body simulation.

Upcoming photometric surveys such as LSST, Roman, and Euclid will map billions of galaxies, opening the door to field-level cosmological analyses. In this talk, I will present a density-slab framework for performing a field-level analog of the standard 3x2pt analysis, where we model galaxy and weak lensing observables by forward modeling density slabs of O[100 Mpc] at the map level rather than through two-point correlation functions. A central challenge in this program is modeling the field-level distribution of galaxies down to small scales (a few Mpcs), where the galaxy–matter connection is non-linear, non-local, non-Poissonian, and correlated across different tracer populations. I will describe our empirical probabilistic model of galaxy bias that captures all of these effects through a flexible parametric form calibrated against simulations. I will show that the model accurately reproduces the statistical properties of galaxy fields from both the UniverseMachine and IllustrisTNG simulations, demonstrating its robustness across different galaxy formation prescriptions and its viability as a forward model for field-level inference of photometric surveys. I will also briefly discuss a related model for the joint non-Poissonian distribution of multiple galaxy populations within halos in the Halo Occupation Distribution framework.

Stage-IV dark energy wide-field surveys, such as the Vera C. Rubin Observatory Legacy Survey of Space and Time (LSST), will observe an unprecedented number density of galaxies. As a result, the majority of imaged galaxies will visually overlap, a phenomenon known as blending. Blending is expected to be a leading source of systematic error in astronomical measurements. We present the Bayesian Light Source Separator (BLISS), a framework for probabilistic detection, deblending, and measurement. We demonstrate the potential of this method with numerical experiments using synthetic observations where truth is known. We highlight experiments that show (i) robustness to spatially varying backgrounds and point spread functions, (ii) how propagating the probabilistic detections to per-object flux posteriors substantially improves aperture flux residuals, and (iii) how our method retains accurate and well-calibrated posterior approximations for shear estimation under increasingly complex observational systematics. BLISS is a scalable, uncertainty-aware tool for mitigating blending-induced systematics in next-generation cosmological surveys.

Dark matter direct detection experiments require information about the local dark matter speed distribution to produce constraints on dark matter candidates, or infer their properties in the event of a discovery. I will discuss how the uncertainty in the dark matter speed distribution near the Sun is affected by baryonic feedback, halo-to-halo variance, and halo mass. I will utilize the statistical power of the new DREAMS Cold Dark Matter simulation suite, which is comprised of 1024 zoom-in Milky Way-mass halos with varied initial conditions as well as cosmological and astrophysical parameters. Applying a normalizing flows emulator to these simulations, the uncertainty in the local dark matter speed distribution is dominated by halo-to-halo variance and, to a lesser extent, uncertainty in host halo mass. Uncertainties in supernova and black hole feedback (from the IllustrisTNG model in this case) are negligible in comparison. Using the DREAMS suite, I will present a state-of-the-art prediction for the dark matter speed distribution in the Milky Way. Although the Standard Halo Model is contained within the uncertainty of this prediction, individual galaxies may have distributions that differ from it. Lastly, I will discuss applying the DREAMS results to the XENON1T experiment and demonstrate that the astrophysical uncertainties are comparable to the experimental ones, solidifying previous results in the literature obtained with a smaller sample of simulated Milky Way-mass halos.

We introduce a novel framework for simulation-based inference using the DREAMS Project, a suite of 1,024 cosmological hydrodynamical zoom-in simulations of Milky Way-mass halos. This suite is designed to systematically disentangle theoretical uncertainties in galaxy formation physics from intrinsic halo-to-halo variance by varying key astrophysical parameters governing supernova wind energy, wind speed, and AGN feedback efficiency within the IllustrisTNG model [arXiv:2512.00148]. To overcome the computational bottleneck of evaluating this high-dimensional parameter space, we utilize a hierarchical generative machine learning framework. By incorporating conditional normalizing flows and Variational Diffusion Models, we accurately emulate both central host properties and variable-length satellite populations [arXiv:2409.02980]. We then introduce a novel observational weighting scheme constrained by the empirical stellar mass-halo mass relation [arXiv:2602.03613]. This approach yields pseudo-posterior constraints that reveal broad degeneracies in fiducial feedback parameters, demonstrating that standard single-model tuning misses complex parameter interdependencies. Applying this inference framework allows us to robustly assess the impact of feedback variations versus accretion history. For central galaxies, we demonstrate that specific structural shifts are driven by specific merger histories, such as the Gaia-Sausage-Enceladus event, though immense halo-to-halo scatter persists. For satellites, we show that intrinsic variance overwhelmingly dominates population statistics, while identifying a persistent tension regarding extended half-light radii observed in the SAGA survey [arXiv:2512.02095]. Finally, we outline ongoing extensions of this inference framework to multi-code simulations (FIRE3, RAMSES, ChaNGa) along with new simulation designs that incorporate mass-varied and resolution-varied suites spanning $10^9$ to $10^{14}$ solar masses; establishing a new method to understand parameter and resolution variations across a full range of halo masses.

We investigate the cosmological information content of higher-order statistics of the CMB lensing convergence field for near-term experiments similar to the Simons Observatory. Using a field-level forward-modeling pipeline based on ray-traced $N$-body simulations with realistic SO-like lensing reconstruction, we measure non-Gaussian statistics such as Minkowski functionals, peak/minima counts, and train emulators to model their dependence on $\Omega_m$, $A_s$, and $M_\nu$. We quantify the information gain beyond the lensing power spectrum and identify which statistics are most robust to reconstruction noise, highlighting the potential of non-Gaussian statistics to enhance cosmological constraints from upcoming CMB surveys.