Deep Learning-Based Geomechanical Upscaling Technique for Reservoir Models Considering Lithological Heterogeneities and Discontinuities

ABSTRACT: Understanding the geomechanical response of reservoirs with lithological heterogeneity and natural fracture networks is crucial for assessing their stability and mechanical behavior in subsurface. Complex interactions between fractures, weak beddings, and host materials introduce significant uncertainties, especially under deformation and failures. Traditional numerical models often simplify fracture networks to improve computational efficiency, yet this oversimplification limits their predictive accuracy. To address this challenge, we propose a deep-learning-based upscaling technique to efficiently predict the geomechanical response of heterogeneous rock masses containing weak beddings and discrete fracture networks (DFN). The proposed method leverages convolutional neural networks (CNNs) to learn stress-strain behavior directly from fracture geometry and lithological heterogeneity, enabling rapid and accurate predictions. This approach provides a computationally efficient framework for analyzing complex fractured heterogeneous reservoirs, facilitates the upscaling of coupled flow and geomechanical processes from the microscopic to macroscopic scale. The findings advance geomechanical upscaling methodologies by incorporating both lithological heterogeneity and discontinuites, providing a valuable tool for reservoir management and subsurface engineering applications.