Seismic impedance inversion based on geophysical-guided cycle-consistent generative adversarial networks

Abstract

Deep learning algorithms have shown great potential in geophysical areas such as seismic interpretation and seismic inversion. However, when applied to seismic inversion, high dependence on label data amount and lack of geophysical constraints severely influences the feasibility and accuracy of results predicted by deep learning-based methods. To address these two problems, we design an improved cycle-consistency generative adversarial network (Cycle-GAN) to mitigate the dependence on labeled data amount. Only a small amount of label data (less than 1%) can render high accuracy of predict results with this method, eliminating the need for an initial model. Besides, novel geophysical constraints are added to the network. Firstly, the deterministic inversion results and labeled data are substituted into hybrid-geophysical data as a training set. Then the residuals between field data and synthetic seismic data are used as part of the loss function to make the predicted results meet the geophysical laws. The geophysical inconsistency factor is introduced as an index to quantitatively evaluate the consistency between synthetic seismic data and field data. The Marmousi model test results prove that this method significantly improves the prediction accuracy and the matching degree of synthetic seismic data compared with other methods such as Cycle-GAN and Convolutional Neural Network (CNN). We apply this method to field data, and the prediction results show that the predicted impedance of this proposed method is highly correlated with the real impedance.