Fast least-squares reverse time migration based on cycle-consistent generative adversarial network

The high computational costs of the least-squares iterative solution limit the large-scale industrial application of the least-squares reverse time migration (LSRTM)method. The difference between traditional reverse time migration (RTM) and least-squares reverse time migration is whether to solve the inverse Hessian matrix or not. This paper proposes a solution by simulating the inverse of the Hessian matrix using a cycle-consistent adversarial neural network (cycleGAN). The network constructs a mapping relationship between the reverse time migration and high-precision imaging, improving imaging quality while significantly reducing computation costs. The trained network is applied to the reverse time migration results of the Marmousi model and the Sigsbee2A model, and the imaging results obtained from the network prediction demonstrate that this method improves the offset imaging quality better with almost no increase in computational effort. © 2023 University of Petroleum, China. All rights reserved.