Colloq Abstract - Thyagarajan

by Hubertus Intema last modified Jan 21, 2016 by Lori Appel

Jan 29, 2016

11am Mountain

 

Nithyanandan Thyagarajan

Arizona State University

 

E-field Parallel Imaging Correlator: An Ultra-efficient Architecture for Modern Radio Interferometers

Abstract

Next generation radio interferometers driven by cosmology and time-domain applications will require thousands of antenna elements to achieve the survey speed and large scale sensitivity. The computational complexity of traditional correlators (FX and XF) scales as the number of antennas squared, and will quickly approach the petaFLOPS regime in the coming decades, making the correlator a dominant cost of these instruments. I present the E-field Parallel Imaging Correlator (EPIC) software to provide the first detailed demonstration of the Modular Optimal Fast Fourier (MOFF) imaging, a generic FFT-based direct imaging algorithm. This software is an efficient alternative to those implementing conventional correlator-imaging algorithms, particularly for dense antenna array layouts. It takes raw antenna voltages as input, computes the spectrum using a temporal FFT (F-engine) and calibrates these antenna voltage spectra without ever forming cross-correlation products from antenna pairs unlike FX/XF correlators. Instead, for calibration, we correlate antenna voltages with a single output image pixel. This method scales only as the number of antennas, accounts for complex sky models, and produces results equal in quality to traditional visibility-based calibration. Using the finite extents of antenna apertures, the calibrated electric fields and the antennas' collecting areas are projected onto a grid on the ground. Images are obtained by squaring the spatial Fourier transform of the electric fields projected on this grid, thereby completely avoiding the need to estimate O(N^2) cross-correlations between antenna pairs. For densely packed antenna array layouts, the imaging cost scales efficiently as O(N log N). Another significant advantage of our software is that the gridding naturally allows for direct imaging even from irregularly placed antenna array layouts and antennas whose power-patterns are not identical. We demonstrate that the images so obtained with our code are equivalent to those from conventional correlators and imaging techniques. We identify specific science cases that will be more effectively enabled with this software implementation like real-time search of radio transients with future arrays such as HERA, LWA and the SKA.