Large-Scale Structure Cosmology with AI-Accelerated Forward Modeling

  • March 16, 2020, 2:00 pm US/Central
  • Curia II
  • Andrew Hearin, Argonne National Laboratory

Present-day and near-future galaxy surveys create the capability to measure cosmological structure growth using multiple, distinct populations of galaxies as tracers of the underlying density field. Multi-tracer cross-correlations contain rich information that can be used to self-calibrate systematics and significantly enhance cosmological constraining power. However, conventional theoretical models for making these predictions bear the mark of having been developed in an era of more limited datasets and modest simulations, leaving present-day analyses ill-equipped to reap the potential benefits of multi-tracer measurements. In this talk, I discuss a differentiable formulation of the connection between galaxies and dark matter halos that meets the predictive needs of multi-wavelength measurements. This new generation of AI-accelerated modeling makes computationally efficient use of high-resolution survey-scale simulations by leveraging the multi-architecture scalability of contemporary deep learning libraries, and so should prove useful in deriving future cosmological constraints with larger galaxy surveys using high-resolution, extreme-scale simulations.