A massively parallel frequency - domain full - waveform inversion algorithm for imaging acoustic me -

We present a massively parallel algorithm for distributed-memory platform to perform 2D acoustic frequency-domain Full-Waveform Inversion (FWI) of global offset seismic data. Our code is written in Fortran 90 and uses Message Passing Interface (MPI) for parallelism. The linearized inverse problem is solved by a classical gradient method which consists in finding a model perturbation which minimizes the least-squares cost function. A key feature of our code is the direct solver used for frequency-domain full-waveform modeling. Frequency-domain full-waveform modeling requires to solve a huge sparse system of linear equations with complex coefficients whose multiple Right-Hand Side (RHS) terms are the seismic sources and solutions of the system are the pressure wavefield. We use the MUlti frontal Massively Parallel direct Solver (MUMPS) for distributedmemory computer to solve this system (LU factorization). MUMPS includes a functionality to perform parallel multiple-shot resolutions once the LU factors have been distributed over the processors. Once the resolution phase is completed, the multiple-shot solutions are left in core and distributed over the processors allowing straight forward parallelization of subsequent tasks in the FWI program such as gradient computation. We first validated our algorithm with a realistic synthetic test consisting of a dip section of the Overthrust model. We first inverted 7 frequencies successively following a multiresolution strategy. Second, two groups of 4 and 3 frequencies respectively were inverted successively. A fastest and more robust convergence was observed using the first multiresolution strategy. This multiresolution strategy was applied to a real wide-angle 2D dense Ocean Bottom Seismometers (OBS) data set (100 OBS gathers) recorded with a 1-km spacing in the eastern-Nankai trough (Japan) to image the deep crustal structure of the subduction system (105 km X 25 km). A speedup of 14 was obtained using 40 processors of a PC cluster and an Infiniband interconnect. The same parallel strategies was implemented in a 3D full-waveform inversion algorithm.