Towards improving crosstalk suppression in multiparameter FWI by decorrelating parameter classes
Mariana Lume, Scott Keating, Kristopher A. Innanen
Multiparameter FWI is commonly affected by parameter crosstalk. These effects are described and corrected by the Hessian, which also impact the shape of the objective function iso-surfaces and the convergence of the optimization algorithms. This study focuses on finding an intermediate model space where the parameter classes are decorrelated, i.e., where the Hessian is an identity matrix, to minimize crosstalk and reach an accurate minimum that could be transformed to the density, Vp and Vs model space. Transformation rules between model spaces were applied in a FWI workflow, using transformation matrices (T) constructed to satisfy constraints imposed by the Hessian of the intermediate system. Overall, this FWI method produced relatively good Vs estimations, but did not overcome a reference FWI in the Vp and density results, since more crosstalk was introduced. However, improvements on the structure of the Hessians with respect to those from the reference inversion were brought for some areas of the model grid, which makes the main decorrelation ideas promising to minimize these coupled effects. The drawbacks were related to a localized approach to compute T, which might need to account, in future work, for crosstalk contributions of multiple grid cells.