COSMO-SBI: a Simulation-Based Inference Framework for Multi-Probe Cosmology

Objetivo

Key open questions in cosmology can be addressed by studying the evolution of the Universe’s structure as a function of scale and time using various observational probes. Traditional methods rely on two-point statistics (e.g. the power spectrum) of cosmological observables and theory-based modeling. In recent years, I have pioneered alternative methods: I developed and applied to weak lensing data a simulation-based inference (SBI) framework to estimate parameter posteriors using simulations. This framework opens the door to studying the data with newer probes for which theoretical models do not exist, enabling the use of higher-order statistics alongside two-point statistics. It also offers a more flexible and accurate incorporation of observational systematics and simplifies the combination of summary statistics from multiple datasets.

With ERC support, I will extend the weak-lensing SBI framework to forward-model CMB secondary anisotropies (CMB lensing, thermal Sunyaev–Zel’dovich) and apply it to Euclid, SPT, and Planck. This enables the first joint analysis of these probes using two- and higher-order statistics. COSMO-SBI will deliver the most precise measurement or the matter fluctuation amplitude σ8 to date (sub-percent in ΛCDM), improve the σ8–Ωm Figure-of-Merit by ~4× over standard approaches, and achieve ~5× tighter w0–wa constraints than conventional two-point pipelines applied to the same LSS data. When combined with external data, the joint fit is expected to yield leading constraints on dynamical dark energy. COSMO-SBI will also provide a sound-horizon independent ~1% measurement of H0, offering a valuable independent cross-check against other methods, and new insight into small-scale baryonic physics across halo mass and redshift.

COSMO-SBI will demonstrate the power of SBI for joint probes, paving the way for more ambitious SBI analyses and motivating the development of SBI frameworks for multiple observables and more beyond-ΛCDM models.