Motivation and Overview

Oil and gas exploration nowadays typically demands tremendous amount of computation. Consider for example a 3D velocity model with a moderate grid size $\gls{nx}\times\gls{ny}\times\gls{nz} = 1000\times 1000\times 500$ , with $\gls{Stot}=100\times 100$ shots and $\gls{nt}=4000$ time steps for one run of wavefield propagation. Suppose one rtm (McMechan, 1983; Whitmore, 1983; Baysal et al., 1983) requires three such runs, and the nflop per grid point per time step is 25, then the total flop count is nflop= $\gls{nx}\times\gls{ny}\times\gls{nz}\times \gls{nt} \times \gls{Stot} \times 3 \times 25= 1.5 \cdot 10^{18}$ , or 1.5 exa. Factor in a dozen iterations and more complexity per iteration due to elaborate optimization schemes, as encountered typically in fwi (Tarantola, 2005,1984; Lailly, 1984), the total nflop for this moderately sized model presents a considerable computational cost.

One approach to reduce this cost is known as the multisource technique (Romero et al., 2000; Tang, 2009; Dai and Schuster, 2009; Virieux and Operto, 2009; Morton and Ober, 1998; Krebs et al., 2009), whereby every S sources are active (deemed a supershot) in one finite-difference simulation and every S shot gathers are correspondingly blended into one supergather. This blending of shot gathers is also done for the many shot gathers recorded in the field, which are recorded one shot gather at a time. Consequently, the number of effective gathers to cope with is cut down by S-fold. The downside of this technique, however, is crosstalk noise that tarnishes the reconstructed model image. Think of a similar task of determining the acoustic property of a lecture hall, while hosting a cocktail party!