Three-dimensional discrete fracture network identification based on deep learning and reversible jump Markov chain Monte Carlo algorithm

Characterizing fracture networks is critical for groundwater development, geothermal exploitation, hydrocarbon recovery, and geological CO 2 sequestration, yet their complex and uncertain spatial distribution poses a persistent challenge. This study proposes an intelligent inversion framework that integrates a 3D-UNet surrogate model, reversible jump Markov Chain Monte Carlo (rjMCMC), and multi-source data fusion for three-dimensional discrete fracture network (DFN) characterization at the field scale. Within this framework, a 3D-UNet model trained on large datasets of fracture configurations, hydraulic head, and electrical potential provides an efficient initial inversion of fracture parameters. Fracture geometries are then extracted with the RANSAC algorithm and iteratively refined via rjMCMC, where the surrogate 3D-UNet replaces conventional forward modeling. This innovation reduces computational costs by an order of magnitude, enabling efficient large-scale inversion. Furthermore, the fusion of electrical potential with hydraulic head data enhances inversion accuracy by about 10 %. Validation demonstrates that the framework reliably reconstructs the spatial distribution of fracture networks, capturing both low-density zones and dominant hydraulic pathways in highly heterogeneous domains. By combining computational efficiency with improved accuracy, this approach offers a practical and scalable solution for field-scale fracture network characterization in a wide range of hydrogeological and engineering applications.