Integrating cross-section and seismic data for 3D reservoir structure reconstruction using a hard-soft-constrained CycleGAN

A significant challenge to improve the understanding of subsurface hydrocarbon accumulation is to characterize realistic three-dimensional (3D) reservoir models from sparse observations. These data are usually divided into hard data (e.g. well logs) and soft data (e.g. seismic volumes). The integration of hard and soft data is essential for characterizing complex spatial heterogeneity of subsurface structures. This study proposes HS-CycleGAN, a reliable reconstruction method for 3D reservoir models by integrating hard and soft data into a cyclic generative adversarial network. HS-CycleGAN includes a dual-generator and a dual-discriminator to fully utilize the implicit features of hard and soft data. An enhanced U-Net is exploited in the generator to improve the extraction and reconstruction of spatial heterogeneity. A Markovian discriminator is also integrated to strengthen the extraction of high-frequency features in seismic data. And a reverse validation strategy is employed to enhance the agreement of the realizations with the seismic data. In addition, a joint loss function combining the reconstructed loss and the original GAN loss is used to optimize the training process of HS-CycleGAN. A real-world reservoir modeling experiment shows that HS-CycleGAN can capture the spatial heterogeneous features of both hard and soft data, and reproduce high-resolution 3D reservoir models under the constraints of seismic data.