Orateur
Description
Weak gravitational lensing maps contain rich, non-Gaussian information that standard two-point statistics miss. Full-field inference tackles this by forward-modeling the entire map—evolving initial conditions through structure formation and observational effects—and using the maps themselves to learn about cosmology, without compressing to summary statistics.
In this talk, I will present a scalable, JAX-based, fully differentiable forward model for weak lensing at LSST scale. Built on the latest JAXPM, it produces spherical shear and convergence (κ) maps and runs efficiently on distributed GPUs; I will describe how we executed the pipeline on the Jean Zay supercomputer. I’ll focus on the practical design choices that make end-to-end differentiation and multi-GPU scaling feasible, and outline the next steps toward a complete inference pipeline on realistic survey data, including survey masks, noise, and photometric-redshift effects.
