FracGen: Natural fracture networks reconstruction and upscaling using generative adversarial networks

Accurate modeling of natural fracture networks is crucial for understanding subsurface structures and their properties, yet traditional methods often struggle with complex geometries and scale transitions. In this study, we present a novel non-parametric machine learning model, FracGen, utilizing the Single Image Generative Adversarial Network (SinGAN) to reconstruct and upscale natural fracture networks. After training, using one of three outcrops from different geological sites, the FracGen model can replicate key statistical properties of natural fractures without explicit parameterization, which is validated further through comparisons with real fracture networks. Furthermore, we address the challenge of fracture network upscaling, ensuring that large-scale simulations retain the critical characteristics observed at small scales. Quantitative analysis, including cosine similarity measurements and probability distribution fitting, validates the model's accuracy (e.g., high cosine similarity values indicate strong correspondence between generated and real fracture networks). The proposed methodology advances our ability to model complex fracture networks and paves the way for more effective resource exploitation, better risk assessment, and improved design of engineering projects. • A novel non-parametric approach is developed to reconstruct natural fracture networks using the generative model. • The proposed methodology effectively addresses the challenge of upscaling natural fracture networks. • The FracGen model is validated through visual and quantitative comparison with real fracture networks.