Wavefield compression for adjoint methods in full-waveform inversion

Abstract

Adjoint methods are a key ingredient of gradient-based full-waveform inversion schemes. While being conceptually elegant, they face the challenge of massive memory requirements caused by the opposite time directions of forward and adjoint simulations and the necessity to access both wavefields simultaneously for the computation of the sensitivity kernel. To overcome this bottleneck, we present lossy compression techniques that significantly reduce the memory requirements with only a small computational overhead. Our approach is tailored to adjoint methods and utilizes the fact that the computation of a sufficiently accurate sensitivity kernel does not require the fully-resolved forward wavefield. The collection of methods comprises re-interpolation with a coarse temporal grid as well as adaptively chosen polynomial degree and floating-point precision to represent spatial snapshots of the forward wavefield on hierarchical grids. Furthermore, the first arrivals of adjoint waves are used to identify "shadow zones" that do not contribute to the sensitivity kernel. Numerical experiments show the high potential of this approach achieving an effective compression factor of three orders of magnitude with only a minor reduction in the rate of convergence. Moreover, it is computationally cheap and straightforward to integrate in finite-element wave propagation codes with possible extensions to finite-difference methods.

Efficient full waveform inversion using the excitation representation of the source wavefield

“…The boundary saving scheme is a solution (Berkhout 1988;Clapp 2008), but at the cost of an additional wavefield extrapolation. Also, techniques based on wavefield compression either temporally or spatially or both are viable alternate solutions maintaining a balance between computational overhead and time (Unat et al 2009;Dalmau et al 2014;Boehm et al 2016). Although promising, these kind of techniques are also not a panacea in complex models, involving trade-offs in the degree of compression and the amount of distortion while decompressing (Mittal & Vetter 2016).…”

Compression Challenges in Large Scale Partial Differential Equation Solvers

“…Adjoint, or dual, equations are important in PDE-constrained optimization problems, e.g., optimal control in electrophysiology [71] or inverse problems [72,73], and goal-oriented error estimation [74]. Consider the abstract variational problem find x ∈ X such that c(x; ϕ) = 0 ∀ϕ ∈ Φ,…”

Stochastic Seismic Waveform Inversion Using Generative Adversarial Networks as a Geological Prior

“…Efficient parameterizations (Akcelik et al, 2002;Boehm et al, 2016) that allow a dimensionality reduced representation of the high-dimensional parameter space of possible Figure 2a shows the acoustic recordings as individual continuous wave forms of the measured acoustic signals. Figure 2b shows the same dataset of acoustic measurements visualized as a collection of discrete pixels of an image.…”