Fast waveform inversion without source encoding

Randomized source encoding has recently been proposed as a way to dramatically reduce the costs of full waveform inversion. The main idea is to replace all sequential sources by a small number of simultaneous sources. This introduces random crosstalk in the model updates and special stochastic optimization strategies are required to deal with this. Two problems arise with this approach: i) source encoding can only be applied to fixed-spread acquisition setups, and ii) stochastic optimization methods tend to converge very slowly, relying on averaging to get rid of the cross-talk. Although the slow convergence is partly offset by the low iteration cost, we show that conventional optimization strategies are bound to outperform stochastic methods in the long run. In this paper we argue that we don¬øt need randomized source encoding to reap the benefits of stochastic optimization and we review an optimization strategy that combines the benefits of both conventional and stochastic optimization. The method uses a gradually increasing batch of sources. Thus, iterations are very cheap initially and this allows the method to make fast progress in the beginning. As the batch size grows, the method behaves like conventional optimization, allowing for fast convergence. Numerical examples suggest that the stochastic and hybrid method perform equally well with and without source encoding and that the hybrid method outperforms both conventional and stochastic optimization. The method does not rely on source encoding techniques and can thus be applied to non fixed-spread data.

https://slim.gatech.edu/Publications/Public/Journals/GeophysicalProspect...