Deep learning-enhanced multiscale permeability prediction in fractured-vuggy carbonate reservoirs

Predicting equivalent permeability in fractured-vuggy carbonate reservoirs is critical for optimizing hydrocarbon recovery and aquifer management yet remains challenging due to the multiscale heterogeneity of matrix, fracture, and vug (MFV) systems. Traditional numerical methods, such as continuous medium models (CMM) and discrete fracture-vug network models (DFVNM), inherent trade-offs between computational efficiency and accuracy. This study proposes a deep learning (DL)-based approach to predict permeability by integrating coupled Stokes-Brinkman-Darcy flow physics with data-driven techniques. A deep neural network (DNN) was trained on 200 samples generated via Latin Hypercube Sampling (LHS), incorporating four key parameters: fracture-vug number, fracture-cavity density, fluid viscosity, and fracture aperture. Numerical simulations in COMSOL provided permeability data, which were normalized and split into training/testing sets. The DNN achieved high predictive accuracy (R2 = 0.92) through 3-fold cross-validation. The results demonstrate that permeability is governed by the nonlinear synergy of these parameters: structural parameters (aperture, number, density) enhance permeability by optimizing flow paths, with their effectiveness significantly amplified under low-viscosity conditions, while increasing fluid viscosity emerges as the dominant limiting factor, suppressing improvements and ultimately leading to saturation in permeability enhancement. This DL framework effectively captures these complex multiscale flow interactions, offering a computationally efficient and accurate alternative to traditional methods, and provides practical insights for reservoir characterization where permeability maximization depends on developing well-connected fracture networks under favorable fluid viscosity conditions.