Autonomous Gaussian Decomposition: A Tool For Analyzing Spectral Big Data in the SKA Era

Type Conference Paper
Names Robert Lindner, 21 Sponge
Conference Name Exascale Radio Astronomy
Volume 46
Pages 40302
Date April 1, 2014
Short Title Autonomous Gaussian Decomposition
Library Catalog
Abstract Statistical comparisons between simulated and observational data are crucial for understanding detailed physical processes in astrophysics. Simulations are increasing in spatial dynamic range, and spectral-line observations in the SKA era will produce millions of high resolution and high-sensitivity spectra. However, interpreting the huge volumes of radio spectra-line data from next-generation radio telescopes and simulations will be an insurmountable challenge for current algorithms. For example, in analyzing HI emission/absorption spectra, the well-known technique of "Gaussian components" can provide a good description of the physical properties of individual ISM clouds, but requires a significant degree of human interaction (and time) in choosing the model's initial parameters-- therefore, it does not scale to large data volumes. We have produced a new algorithm, built on ideas from computer vision and machine learning, for autonomously predicting the number of Gaussian components, as well as their widths and positions, in an arbitrary 1D spectrum. The "unsupervised" nature of the technique allows for truly unbiased comparisons between observations and simulations, and also allows the algorithm to scale up to handle the very large data-volumes from the upcoming SKA and pathfinders, e.g., GASKAP. In my talk, I will present the algorithm, discuss its performance in decomposing synthetic spectra, and present initial results in using it to compare recent, high-sensitivity "21-SPONGE" observations to high-resolution hydrodynamic ISM simulations.
UW-Madison Astronomy Home