Surrogate model for fracture propagation in heterogeneous reservoirs based on generative neural networks

Accurate and efficient simulation of fracture propagation patterns is crucial for optimizing hydraulic fracturing design. However, the impact of rock mechanics heterogeneity and interfaces on fracture propagation is highly complex. Traditional numerical simulation methods often suffer from slow convergence, high computational cost, and the need for manual adjustments of physical parameters under heterogeneous conditions, making fracture propagation simulation a challenging task. To address these issues, this study proposes a fracture propagation prediction method based on a generative neural network named Fracture Propagation GAN (FPGAN). By employing the FPGAN model as a surrogate for fracture propagation simulation, the efficiency of simulating fracture propagation under heterogeneous conditions is significantly enhanced while maintaining the accuracy of the original images. Fracture propagation time-series images under various mechanical parameters were generated using the Finite Discrete Element Method (FDEM) to construct both basic and complex datasets of fracture propagation images. The FPGAN model was trained on these datasets to enable rapid prediction of fracture morphologies under heterogeneous conditions. Experimental results demonstrate that the FPGAN model can predict hydraulic fracture propagation images for any given combination of mechanical parameters within 1 min, achieving computational efficiency improvements of several orders of magnitude compared to traditional numerical methods. The proposed FPGAN model provides a robust foundation for analyzing the influence of different mineral compositions on fracture generation and exhibits significant potential in hydraulic fracture propagation simulation.