Constructing large-scale high-fidelity fracture networks based on generative AI

Abstract

Accurate fracture network modeling is crucial for many fields, such as geothermal resources exploitation, underground infrastructure construction and nuclear waste disposal. However, only small-scale fracture network images can generally be obtained in field work, which is insufficient for engineering-scale analysis. Therefore, based on generative artificial intelligence (AI), we introduced a novel algorithm, termed Upscaling-GAN, to generate large-scale high-fidelity fracture networks by directly learning from small-scale fracture network images collected in the field. The algorithm employs a two-stage generation process: (1) Generative Adversarial Network (GAN) model is trained to generate small-scale fracture networks; (2) based on the patch-by-patch generation paradigm, a large-scale fracture network is generated from these small-scale fracture networks, which are obtained by applying the trained model. The achieved AI-generated fracture networks have significant advantages: 1) compared to fractal-geometry-based methods or stochastic discrete fracture network model, the present method can exactly characterize the geometrical characteristics and topological structures of natural fracture systems; 2) compared to previous GAN-based method, our algorithm can rigorously capture the spatial variability of fracture apertures by directly learning from raw fracture network images without laborious preprocessing procedures such as binarization and skeletonization; 3) when the size of the target image increases, the GPU memory consumption remains nearly unchanged. And the effectiveness of the Upscaling-GAN model in fracture network modeling is rigorously validated through qualitative and quantitative evaluations.