Field-Level Inference with Differentiable GridSPT Forward Model

Explore field-level inference techniques for cosmological analysis using a differentiable GridSPT forward model in this 21-minute conference talk. Learn how to move beyond traditional summary statistics like 2-point correlation functions and power spectra to directly compare observed galaxy density fields with theoretical predictions. Discover the development of differentiable GridSPT, a forward modeling method based on standard perturbation theory that leverages accelerated Fourier transforms and automatic differentiation with GPUs for enhanced computational efficiency. Examine the application of this approach to reconstruct initial density fields and extract cosmological parameters from non-linear halo density fields in N-body simulations, understanding how field-level methods can access information that remains hidden in conventional statistical approaches, particularly in the non-linear regime of cosmic structure formation.