2025/08/14 by Carrion, Karim
#Cosmology and Nongalactic Astrophysics (astro-ph.CO) #FOS: Physical sciences
paper · doi:10.48550/arxiv.2508.10876
The present thesis aims to tackle two critical aspects of present and future cosmological analysis of Large-Scale Structure (LSS): accurate modelling of the nonlinear matter power spectrum beyond ΛCDM, and efficient computational techniques for Bayesian parameter estimation. Both are crucial for testing alternative cosmologies and avoiding spurious results. We focus on the Dark Scattering (DS) model, describing pure momentum transfer between dark matter -- dark energy through the parameter A\rm ds. To capture DS effects, we adopt the halo model reaction framework within \ttReACT, compute the nonlinear DS spectrum, and validate it against N-body simulations. We further include baryonic feedback and massive neutrinos, finding degeneracies between DS and baryonic effects but not with neutrinos. We then constrain DS using cosmic shear from KiDS-1000, accelerated by neural emulators from \ttCosmoPower, which speed up predictions by O(104). Our DS emulator, trained on halo model reaction outputs, preserves percent-level accuracy and incorporates baryonic feedback. Analysing KiDS shear statistics, we obtain \vert A\rm ds\vert \lesssim 20 b/GeV at 68 % C.L. Combining KiDS with Planck CMB and BAO data, we find A\rm ds=10.6+4.5-7.3 b/GeV at 68 % C.L., suggesting the DS model as a promising resolution to the S8 tension. Finally, we present weak lensing forecasts for Stage IV surveys using an automatically differentiable pipeline with \ttjax-cosmo and gradient-based samplers in \ttNumPyro, reducing computational cost from months on CPUs to days on GPUs. Model evidence is evaluated with \ttharmonic under multiple scale cuts. To put things into perspective, the modelling strategies and machine learning accelerations developed here provide powerful tools for the next generation of LSS cosmology.