Uncertainity Quantification for Large Scale Inverse Scattering.

摘要

Our goal is the design of fast parallel algorithms statistical inference for scalar and wave propagation problems. We have looked at source inversion and inverse medium problem problems. We use a Bayesian approach in which the regularization appears as prior information and the data mismatch appears as a likelihood information, given known noise probability density functions. A key component of all of our algorithms is the approximation of the Hessian operator. Key components of our work are rank-revealing factorizations, fast extraction of the diagonal of the inverse, adaptivity, and integration of all of these components within a particle filter methodology. In addition, our implementations are being designed to scale on manycore and heterogeneous parallel architectures.