A sparse reduced Hessian approximation for multi-parameter wavefield reconstruction inversion

Multi-Parameter full-waveform inversion is a challenging problem, because the unknown parameters appear in the same wave equation and the magnitude of the parameters can vary many orders of magnitude. This makes accurate estimation of multiple-parameters very difficult. To mitigate the problems, sequential strategies, regularization methods and scalings of gradients and quasi-Newton Hessians have been proposed. All of these require design, fine-tuning and adaptation to different waveform inversion problems. We propose to use a sparse approximation to the Hessian derived from a penalty-formulation of the objective function. Sparseness allows to have the Hessian in memory and compute update directions at very low cost. This results in decent reconstruction of the multiple parameters at very low additional memory and computational expense.

https://slim.gatech.edu/Publications/Public/Conferences/SEG/2014/peters2...