Model Compression by Grid Shaking for Nonlinear Electromagnetic Inversions

CSEM inversion for large 3D models is computationally very intensive, making the use of advanced inversion algorithms like Gauss-Newton particularly challenging. To overcome this problem model compression techniques can be applied. By decreasing the number of parameters they reduce both memory needs and computing time at the cost of a decreased resolution. We propose a new model compression approach where the compression operator is slightly changed at each iteration. Like many compression techniques it uses a coarse grid to reduce the number of parameters, but random lateral and vertical shifts are applied to this grid at the beginning of every iteration. Compared to usual model compression with a static compression operator, the new method presents exactly the same benefits in terms of computational efficiency but it preserves resolution much better. We have illustrated this approach with Gauss-Newton inversion of synthetic 2D CSEM data. It shows that a significantly higher degree of model compression than with static approaches can be used: good inversion results are obtained with 132 times less parameters than the number of nodes in the modeling grid. This approach will help making Gauss-Newton algorithms applicable to 3D inversion of larger 3D CSEM data sets.