DISCO-DJ - Differentiable Simulations for Cosmology Done with Jax
Erwin Schrödinger International Institute for Mathematics and Physics (ESI) via YouTube
-
39
-
- Write review
Gain a Splash of New Skills - Coursera+ Annual Just ₹7,999
Free courses from frontend to fullstack and AI
Overview
Coursera Flash Sale
40% Off Coursera Plus for 3 Months!
Grab it
Explore the DISCO-DJ framework for modeling cosmic large-scale structure in this 20-minute conference talk from the Workshop on "Putting the Cosmic Large-scale Structure on the Map: Theory Meets Numerics" at the Erwin Schrödinger International Institute. Discover how this comprehensive framework developed at the University of Vienna integrates a linear Einstein-Boltzmann code, perturbation theory models, and a fast particle-mesh N-body module, all built with the Jax library for GPU acceleration and automatic differentiation. Learn about the framework's seamless integration with inference and machine learning frameworks, and examine theory-informed time integrators like the BullFrog method that achieve per-cent-level accuracy in present-day power spectrum predictions using just 6 time steps with 512³ particles in seconds. Understand the numerical techniques employed to control discreteness effects and achieve high accuracy, including custom non-uniform FFT implementation for force evaluation. Examine both forward- and reverse-mode differentiation capabilities with memory requirements independent of time steps, achieved through adjoint formulation in reverse mode, providing a self-consistent, high-performance, and fully differentiable pipeline for universe large-scale structure modeling.
Syllabus
Florian List - DISCO-DJ: DIfferentiable Simulations for COsmology - Done with Jax
Taught by
Erwin Schrödinger International Institute for Mathematics and Physics (ESI)