Predicting Ground Surface Deformation Induced by Pressurized Fractures Using Conditional Generative Adversarial Networks

Abstract

ABSTRACT Hydraulic fracturing has been widely applied in subsurface to facilitate hydrocarbon flow or as a preconditioning method to promote normal caving behavior in mining operations. The success of the method prominently relies on the geometry of the induced fractures in the host rock. An array of high-resolution tiltmeters is commonly used to measure the displacement gradient (tilt angle) of ground surface during hydraulic fracturing operation, and to infer the geometry of the induced fractures. In this work, we present a machine-learned model that is capable of predicting the surface tilt resulting from a pressurized fracture, using Conditional Generative Adversarial Networks (cGAN). Fracture apertures along horizontal and vertical planes are given as input while surface tilt vectors are provided as output for the forward model. A 3D finite element model with discrete fractures is utilized as the Full Order Model (FOM) to create training data. The fractures are pressurized uniformly to induce a displacement at the ground surface. Images of the fracture aperture in XY and XZ planes, and surface tilt vector components (tilt X and tilt Y) are used for the training process. Three different loss functions are used, and the predicted tilts are compared to the real (FOM) tilts. The results show that the trained cGAN with Wasserstein loss and gradient penalty (W-model) can predict the ground surface tilt due to the pressurized fractures at variable dip angles. INTRODUCTION Hydraulic fracturing has been frequently applied as a preconditioning method to promote normal caving behavior in mining operations. The success of the method prominently relies on the geometry of the induced fractures in the stiff overburden layer. An array of high-resolution tiltmeters is commonly used to measure the displacement gradient (tilt angle) of ground surface during hydraulic fracturing operation, and to infer the geometry of the induced fractures or pressure plumes using simplified models (Lecampion et al., 2005; Salimzadeh et al., 2022).