Workshops Archive

SkyMatch.jl: Fast, Differentiable Modeling for Millimeter-Sky Inference


Workshop: Scale to See: Scalable AI for Scientific Imaging and Spatiotemporal Data

Authors: Vimarsh Sathia, Kedar A. Phadke, and Siyuan Brant Qian (University of Illinois at Urbana-Champaign, NSF-Simons AI Institute for the Sky (SkAI)) and Joaquin Vieira and William S. Moses (University of Illinois at Urbana-Champaign)

Abstract: Generating realistic millimeter/sub-millimeter (mm/sub-mm) sky simulations is essential for de-blending sources and understanding cosmic dust-obscured star formation, but remains computationally expensive. While the previous tools could simulate only a small fraction (<0.25%) of the mm-sky, the next-generation telescopes are now mapping nearly 25% of the sky at high sensitivity, probing the faintest and farthest populations of galaxies. Interpreting this data and de-blending sources effectively requires accelerating forward modeling and inverse problem-solving techniques.

In this paper, we present SkyMatch.jl, a fast and differentiable framework for computing mock mm-sky maps and catalogs. Leveraging the Reactant.jl compiler, SkyMatch.jl automatically optimizes the simulation's tensor computations to run efficiently on multi-core CPUs and GPUs. Overall, SkyMatch.jl achieves speedups of up to 90× on CPUs and 3300× on GPUs, enabling us to realistically simulate data from state-of-the-art mm-wave telescopes and interpret the resulting source populations.


Back to Scale to See: Scalable AI for Scientific Imaging and Spatiotemporal Data Archive Listing

Back to Full Workshop Archive Listing