Fast full waveform inversion with random shot decimation

Abstract

One of the main challenge of full waveform inversion (FWI) is its computing cost because each iteration is equivalent to one reverse time migration. To increase the performance of FWI, few shots can be randomly selected at each iteration. The shot location can be changed after every iteration, or kept constant for few iterations before being changed. This random decimation method helps decreasing the memory footprint of FWI because much fewer shots are needed at any time during the inversion. In addition, FWI can be sped‐up effectively while preserving the ability to recover the model. Using a synthetic 2‐D dataset, the memory footprint is divided by a factor twenty while the computation costs is decreased by a factor four.