Seismic impedance inversion based on cycle-consistent generative adversarial network

Nowadays, convolutional neural network (CNN) has achieved great attention in varieties of research fields. Benefit from the strong learning ability of CNN, the knowledge contained in labeled data set can be efficiently extracted. However, the validity of CNN is guaranteed by a sufficient number of labeled data. In the field of seismic impedance inversion, the amount of labeled data is always limited. In order to mitigate the dependence of CNN on the amount of labeled seismic data, we propose to solve the seismic impedance inversion problem using 1D cycle-consistent generative adversarial network (Cycle-GAN), which consists of two generative subnets and two discriminative subnets. The generative subnets are used to model seismic forward and inversion process and the discriminative subnets can constrain the distribution consistency between the generated data and the true data. We investigate this algorithm on both synthetic data set and real data set. The experimental results on synthetic data set show that the proposed Cycle-GAN based inversion algorithm has better performance than conventional CNN based algorithm. And the experiments on real data set also show promising inversion results.