U-net generative adversarial network for subsurface facies modeling

Subsurface models are central pieces of information in different earth-related disciplines such as groundwater management and hydrocarbon reservoir characterization. These models are normally obtained using geostatistical simulation methods. Recently, methods based on deep learning algorithms have been applied as subsurface model generators. However, there are still challenges on how to include conditioning data and ensure model variability within a set of realizations. We illustrate the potential of Generative Adversarial Networks (GANs) to create unconditional and conditional facies models. Based on a synthetic facies dataset, we first train a Deep Convolution GAN (DCGAN) to produce unconditional facies models. Then, we show how image-to-image translation based on a U-Net GAN framework, including noise-layers, content loss function and diversity loss function, is used to model conditioning geological facies. Results show that GANs are powerful models to capture complex geological facies patterns and to generate facies realizations indistinguishable from the ones comprising the training dataset. The U-Net GAN framework performs well in providing variable models while honoring conditioning data in several scenarios. The results shown herein are expected to spark a new generation of methods for subsurface geological facies with fragmentary measurements.

Automated summary

Subsurface models are crucial in various earth-related disciplines, typically obtained using geostatistical simulation methods. Recently, deep learning algorithms have been applied for generating subsurface models, with challenges remaining in including conditioning data and ensuring model variability. Generative Adversarial Networks (GANs) show potential in creating unconditional and conditional facies models.