A fast least-squares reverse time migration method using cycle-consistent generative adversarial network

Abstract

With high imaging accuracy, high signal-to-noise ratio, and good amplitude balance, least-squares reverse time migration (LSRTM) is an imaging algorithm suitable for deep high-precision oil and gas exploration. However, the computational costs limit its large-scale industrial application. The difference between traditional reverse time migration (RTM) and LSRTM is whether to eliminate the effect of the Hessian operator or not while solving Hessian matrix explicitly or eliminating the effect of the Hessian matrix implicitly has a very high requirement on computation or storage capacity. We simulate the inverse Hessian by training a cycle-consistent generative adversarial network (cycleGAN) to construct a mapping relationship between the RTM results and the true reflectivity models. The trained network is directly applied to the RTM imaging results, which improves the imaging quality while significantly reducing the calculation time. We select three velocity models and two velocity models respectively to generate the training and validation data sets, where the validation data is not involved in the training process. The prediction results on the validation data sets show that the trained network significantly improves the imaging quality with almost no additional in computational effort. Finally, we apply the network trained with only synthetics to the field data. The predicted results confirm the effectiveness and good generalization of the proposed method.