Accelerated full-waveform inversion using dynamic mini-batches. Issue 2 (21st February 2020)

• Record Type: Journal Article
• Authors: van Herwaarden, Dirk Philip; Boehm, Christian; Afanasiev, Michael; Thrastarson, Solvi; Krischer, Lion; Trampert, Jeannot; Fichtner, Andreas
• Is Part Of: Geophysical journal international. Volume 221:Issue 2(2020)
• Page Start: 1427
• Page End: 1438
• Publication Date: 2020-02-21
• Subjects: Inverse theory -- Waveform inversion -- Computational seismology -- Seismic tomography; Geophysics -- Periodicals; 550
• DOI: 10.1093/gji/ggaa079
• Languages: English
• ISSNs: 0956-540X

Abstract

SUMMARY: We present an accelerated full-waveform inversion based on dynamic mini-batch optimization, which naturally exploits redundancies in observed data from different sources. The method rests on the selection of quasi-random subsets (mini-batches) of sources, used to approximate the misfit and the gradient of the complete data set. The size of the mini-batch is dynamically controlled by the desired quality of the gradient approximation. Within each mini-batch, redundancy is minimized by selecting sources with the largest angular differences between their respective gradients, and spatial coverage is maximized by selecting candidate events with Mitchell's best-candidate algorithm. Information from sources not included in a specific mini-batch is incorporated into each gradient calculation through a quasi-Newton approximation of the Hessian, and a consistent misfit measure is achieved through the inclusion of a control group of sources. By design, the dynamic mini-batch approach has several main advantages: (1) The use of mini-batches with adaptive size ensures that an optimally small number of sources is used in each iteration, thus potentially leading to significant computational savings; (2) curvature information is accumulated and exploited during the inversion, using a randomized quasi-Newton method; (3) new data can be incorporated without the need to re-invert the complete data set, thereby enabling an evolutionary mode of full-waveform inversion. We illustrate our method using synthetic and real-data inversions for upper-mantle structure beneath the African Plate. In these specific examples, the dynamic mini-batch approach requires around 20 per cent of the computational resources in order to achieve data and model misfits that are comparable to those achieved by a standard full-waveform inversion where all sources are used in each iteration.