Generating high-fidelity discrete fracture networks from low-dimensional latent spaces using generative adversarial network

Characterization of discrete fracture networks (DFNs) in the shallow crust is essential for understanding subsurface flow and transport processes and guiding reservoir exploitation such as water/oil/gas/geothermal/mineral recovery and nuclear waste/CO 2 storage. However, characterizing the geometry of subsurface DFNs is extremely difficult, due to the inherent complexity of DFNs and the generally spatially sparse, low-resolution geological/geophysical data. Traditional DFN parameterization methods may result in a high-dimensional parameter space, making DFN inversion ill-posed and computationally expensive. In this study, we develop a deep learning-based low-dimensional parameterization method to effectively generate complex DFNs from low-dimensional latent spaces, thus significantly alleviating the ill-posedness and computational burden associated with DFN characterization in a data scarce environment. The Wasserstein generative adversarial network with gradient penalty (WGAN-GP) is used to generate random DFNs from latent spaces. Through both qualitative and quantitative comparisons of fracture characteristics between the generated and training DFNs, we demonstrate the extraordinary capability of the method in generating high-fidelity DFNs from extremely low-dimensional latent spaces. The generated DFNs faithfully honor fracture prior knowledge imposed in training samples, including fracture statistics regarding location, length and orientation as well as fracture existence and connectivity identified from geological/hydrogeological surveys. We also demonstrate the ability of the method in generating DFNs that resemble realistic fracture networks mapped from limestone and glacier outcrops. A synthetic DFN characterization case study illustrates the effectiveness of the proposed method in inversion tasks, showing such an effective low-dimensional and conditional parameterization method is particularly useful to facilitate subsurface DFN characterization. • We develop a deep-learning based low-dimensional DFN parameterization method to assist fracture network characterization. • The method can generate high-fidelity DFNs conditioned to pre-defined fracture statistics and connectivity. • Geological-mapped fracture networks from realistic outcrops are successfully generated from low-dimensional latent spaces.