Due to the gradient calculation requiring cross-correlation of the forward wavefields and back-propagated residual wavefields at each time step, the great storage amount becomes an obstacle of practical application of full-waveform inversion, especially in three-dimensional elastic case in time domain. In this paper we extend the efficient boundary storage to the time domain three-dimensional elastic full-waveform inversion on multi-GPU. Based on the efficient boundary storage strategy, the storage amount can be reduced dramatically. As a result, we can save the partial forward wavefields directly on the GPU memory and reconstruct the full forward wavefields synchronized with back-propagated residual wavefields along the reverse time direction. This strategy avoids frequent CPU-to-GPU or GPU-to-CPU memory copy (extremely time-consuming) at the cost of the recomputation (little time-consuming) of the forward wavefields. Our forward simulation tests show that the GPU’s supercomputing effect can be fully exploited with this strategy. In addition, we perform a three-parameter simultaneous inversion of P-, S-wave velocities and density. The favorable inversion results verify that our algorithm is feasible and efficient.
Number of Pages
Looking for more?
Some of the OnePetro partner societies have developed subject- specific wikis that may help.