CSE Seminar Announcement

Fast and Parallel Algorithms for Inverse Multiple-Scattering Solutions and Applications on Tomographic Imaging

Wednesday, September 27, 2017
1030 National Center for Supercomputing Applications (NCSA)
12:00pm - 1:00pm
Lunch Provided

Mert Hidayetoglu
CSE Fellow
Electrical and Computer Engineering

Abstract

Conventional inverse-scattering methods usually assume a single-scattering mechanism for linear reconstruction of wave properties of the object of interest. This assumption provides well-approximate images of low-contrast objects, such as in diffraction tomography or in projection tomography; however, this mechanism breaks down when the multiple-scattering phenomena come into the picture images due to the inaccurate physical model. To solve more demanding problems involving higher-order wave scattering phenomena, we employ full-wave formulations which make no fundamental approximation.

Implemented into one’s favorite gradient-based nonlinear solver (e.g., steepest-descent, conjugate-gradient, or Newton-type solver), the distorted-wave Born approximation provides a semi-analytical way to obtain exact functional derivatives which are required in each iteration. However, compounding multiple object illuminations and the corresponding measurements into the iterative optimization process requires many (hundreds to thousands, typically) forward-scattering solutions in each iteration. Considering each forward solution involves large (millions-by-millions) dense matrices to be inverted, the computational cost of the object reconstruction is immense. As a result, the multiple-scattering formulations were thought to be impractical until recently.

This talk will be on the fast and parallel algorithms for reducing the computational complexity of solutions and for utilizing large supercomputers, respectively. Specifically, we use the multilevel fast multipole algorithm to multiply N-by-N dense matrices with O(N) computational complexity (without making any fundamental approximation). For parallelization, we propose a hierarchical scheme which provides good granularity for spreading the load of iterative inversions among large number of computing nodes. For further speedup, we employ GPUs. The talk will demonstrate efficient solutions of nonlinear tomographic reconstruction problems in near-real time (which are the largest and fastest to date). To do so, we employ up to 4,096 GPUs of NCSA’s Blue Waters Supercomputer. Additionally, several approaches will be demonstrated to avoid local-minima stagnation as well as incorporation of multiple-scattering phenomena for imaging details which are invisible otherwise.

Biography

Mert Hidayetoğlu went to Bilkent University, Turkey, where he received his B.S. and M.S. degrees in electrical engineering in 2013 and 2015, respectively. Starting from late 2010, until his arrival at ECE Illinois, he was with the Computational Electromagnetics Research Center, where, he worked on radar cross section calculations of chaff clouds, scattering analysis involving real-life jet aircrafts, and radiation analysis of complicated antenna structures mounted on large platforms, all with in-house, numerical, full-wave, fast, and parallel solvers. His research led accurate solutions of extremely-large scattering and radiation problems involving dense NxN linear systems with more than one billion unknowns, which was a world record by then. He achieved this on a relatively modest parallel computer.

Mert has been with the electromagnetics group at ECE Illinois for the last two years under supervision of Prof. Weng Cho Chew. His research involves fast and parallel inverse-scattering algorithms for multiple-scattering tomographic imaging. With Prof. Wen-Mei Hwu at the Coordinated Science Laboratory, who co-advises Mert through his research, he implements the fast imaging algorithms on large-scale computing systems like NCSA Blue Waters for solving challenging computational problems in near-real time. Mert’s areas of interest include electromagnetics, integral equations, inverse scattering, fast algorithms, and parallel & high-performance computing. He is the first author of more than a dozen conference papers published at respectable venues.