Paper Summary
We present a newly developed 3D inversion code for Towed streamer EM data utilizing the Gauss-Newton method. The code is designed and implemented on a standard cluster of nodes to handle large-scale inversions of the order 25 million unknowns. The corresponding Jacobian matrix of about 2 Tb is split over the subscribed computational nodes in order to reduce the memory consumption on each node as well as to reduce the computational time. This enables the use of the full Gauss-Newton method on this amount of data. The forward modeling part is based on a parallelized integral equation formulation of the electric field. The forward and inversion grids cover the entire survey area including bathymetry and down to 5 km below the seasurface. The performance is demonstrated with an unconstrained anisotropic 3D inversion on Towed Steamer EM data over a survey area covering a total of 5206 sq. km, acquired in the Barents Sea in 2014. The survey consisted of 38parallel acquisition lines with a spacing of 1.25 km and an average length of 115 km. The selected number of data observations for this case was 505820 with a total of 6 frequencies in the range from 0.2 to 2.2 Hz, 7226 shotpoints and 16 receiver offsets. The inversion grid consisted of 13 million cells and extended from the water column down to 5 km below the sea surface. The inversion ran 14 Gauss-Newton iterations and converged to a misfit of 3.7%with a geologically feasible anisotropic resistivity model.