Lossy Wavefield Compression for Full-Waveform Inversion

Abstract:

We present lossy compression techniques, tailored to the inexact computation of sensitivity kernels, that significantly reduce the memory requirements of adjoint-based minimization schemes.

Adjoint methods are a powerful tool to solve tomography problems in full-waveform inversion (FWI). Yet they face the challenge of massive memory requirements caused by the opposite directions of forward and adjoint simulations and the necessity to access both wavefields simultaneously during the computation of the sensitivity kernel. Thus, storage, I/O operations, and memory bandwidth become key topics in FWI.

In this talk, we present strategies for the temporal and spatial compression of the forward wavefield. This comprises re-interpolation with coarse time steps and an adaptive polynomial degree of the spectral element shape functions. In addition, we predict the projection errors on a hierarchy of grids and re-quantize the residuals with an adaptive floating-point accuracy to improve the approximation. Furthermore, we use the first arrivals of adjoint waves to identify "shadow zones" that do not contribute to the sensitivity kernel at all. Updating and storing the wavefield within these shadow zones is skipped, which reduces memory requirements and computational costs at the same time.

Compared to check-pointing, our approach has only a negligible computational overhead, utilizing the fact that a sufficiently accurate sensitivity kernel does not require a fully resolved forward wavefield. Furthermore, we use adaptive compression thresholds during the FWI iterations to ensure convergence.

Numerical experiments on the reservoir scale and for the Western Mediterranean prove the high potential of this approach with an effective compression factor of 500-1000. Furthermore, it is computationally cheap and easy to integrate in both, finite-differences and finite-element wave propagation codes.