Full-wave methods incorporate all wave phenomena into the image reconstructions by solving the Helmholtz equation with no fundamental approximation. These phenomena include refraction, absorption, diffraction, and multiple-scattering of propagating waves.  Although full-wave reconstruction has promising features, it has been thought to be impractical (read it impossible) due to its high computational burden.  This talk will be on our effort to make full-wave imaging attainable with two approaches:  fast algorithms and supercomputing.
 
The Distorted-Born Iterative Method (DBIM) follows a gradient-based approach for the nonlinear optimization, where mathematically-exact functional derivative is found with the distorted-Born approximation.  But this requires solving forward-scattering problems, i.e., inversion of large N-by-N dense matrices, for each object illumination.  We employ the Multilevel Fast Multipole Algorithm (MLFMA) for solving these forward problems with O(N) computational complexity.  Furthermore, we use NCSA’s Blue Waters supercomputing facility with CPU+GPU node architecture for massively-parallel reconstructions.  For efficient implementation of MLFMA, we seek low-level GPU optimizations and effective heterogeneous computing.
 
This talk provides an overview of DBIM and MLFMA, efficient parallelization strategies on large supercomputers, some GPU optimizations, and performance results. The results show good scaling up to 4,096 GPU nodes which provides the largest full-wave image reconstructions to date in near-real time. Several real-life scenarios will be provided where the proposed methodology is especially useful and outperforms conventional approaches in terms of image quality.