Modern cosmology faces a grand challenge: How to extract reliable physical knowledge from incomplete, biased, and heterogeneous observations of a highly complex Universe. Progress no longer depends solely on improved instrumentation and refined theoretical models. It also depends on the ability to design end-to-end inference frameworks that connect raw observational data to cosmological interpretation. This manuscript presents work that spans this full chain of the cosmological field of research: from photons recorded by detectors up to physical knowledge obtained from Bayesian inference through synthetic universes generated on supercomputers. Still, while each component of this chain can be addressed with increasing precision, it also highlights a fundamental limitation: These elements are still largely developed and applied independently. The major challenge that emerges is therefore not only to improve each layer, but also to unify them within a single, coherent, and bias-controlled framework in which uncertainties are propagated consistently from data acquisition to cosmological parameters. Such a unification, relying on realistic digital twins of the Universe, would enable controlled experiments on systematic effects and the validation of inference pipelines under fully consistent conditions. It would open the way to percent-level accuracy in cosmological measurements and provide a decisive test of current inconsistencies in the standard model of cosmology. It would clarify whether these discrepancies originate from residual biases or from new physics beyond the standard cosmological model. This work should thus be understood both as a set of methodological advances and as the foundation for a large project aimed at closing a unified loop between observation, simulation, and inference in cosmology.
defended on 02/10/2026
français
English