Using the parallel MUMPS solver for frequency domain full waveform inversion

Scott Keating, Kristopher A. Innanen

Frequency domain finite difference (FDFD) wave propagation is often used in full waveform inversion (FWI) research. The major cost in FDFD is the factorization of a large, sparse matrix. This means that solving the problem for an additional source term is inexpensive, but also raises challenging problems for parallelization. Parallelization over source terms offers negligible benefits as no re-factorization of the matrix is required. Parallelization over different frequencies is possible, but limited by the number of frequencies considered, which may be small. This work focuses on using the MUMPS solver, which allows for massive parallelization of the factorization itself.