We compared FWI using three different kinds of boundary conditions: absorbing, random and continuation of velocity.
The random and continuation-of-velocity boundary conditions eliminate almost all I/O cost associated with wavefield storage and transfer at the expense of two extra wavefield propagations per iteration. Random boundary inversion results are almost as good as those from the absorbing boundary condition, which is current industry practice. The continuation-of-velocity boundary condition, on the other hand, works quite well in well-constrained areas, and less so in poorly constrained areas This is particularly attractive on unconventional architecture where computational cost is much less than memory access cost ( i.e. GPUs ).