Hybrid optimization methods for Full Waveform Inversion

Thursday, 18 December 2014
Debanjan Datta and Mrinal K Sen, University of Texas at Austin, Austin, TX, United States
FWI is slowly becoming the mainstream method to estimate velocity models of the subsurface from seismic data. Typically it makes use of a gradient descent approach in which a model update is computed by back propagating the residual seismograms and cross correlating with the forward propagating wavefields at each grid point in the subsurface model. FWI is a local optimization technique, which requires the starting model to be very close to the true model. Because the objective function is multimodal with many local minima, the requirement of good starting model becomes essential. A starting model is generated using travel time tomography. We propose two hybrid FWI algorithms one of which generates a very good starting model for a conventional FWI and the other, which works with a population of models uses gradient information from multiple starting locations in guiding the search.

The first approach uses a sparse parameterization of model space using non-oscillatory splines, whose coeffiencts are estimated using an optimization algorithm like very fast simulated annealing (VFSA) by minimizing the misfit between the observed and synthetic data. The estimated velocity model is then used as a starting model for gradient-based FWI. This is done in the shot domain by converting the end-on marine geometry to a split spread geometry using the principle of reciprocity. The second approach is to uses an alternate global optimization algorithm called particle swarm optimization (PSO) where PSO update rules are applied. However, we employ a new gradient guided PSO that exploits the gradient information as well. This approach avoids the local minima and converges faster than a conventional PSO.

We demonstrate our methods with application to 2D marine data sets from offshore India. Each line comprises over 1000 shots; our hybrid methods produce geologically meaningful velocity models fairly rapidly on a GPU cluster. We show that starting with the hybrid model gives a much better velocity model than starting with a simple smooth model.