Enhancing Computational Efficiency of Numerical Simulation for Subsurface Fluid-Induced Deformation Using Deep Learning Reduced Order Models

The development of subsurface exploitation projects, including CO2 storage processes, requires a large number of numerical simulations where fluid and transport in porous media are coupled, at a certain stage, with the solution of the Biot problem, for instance, to evaluate the potential of faults destabilization and associated induced seismicity. These simulations are computationally expensive, therefore we consider the possibility of using a surrogate for full-physics simulations. We follow a data-driven approach based on neural networks, denoted as deep learning reduced order modeling (DL-ROM), to build a reduced model where the reduced space is identified using an autoencoder trained on full-physics numerical solutions. In this work, we consider physical uncertain parameters, such as Young's moduli, permeabilities, and fault transmissibility, as well as process controls, such as the injection rate of CO2, and we train the DL-ROM using numerical simulations of reservoir operations. We apply the surrogate modeling to two synthetic problems developed in the context of underground CO2 storage, where the injection sites are characterized by a sloping fault that could be destabilized during operations. High-fidelity simulations are performed using a one-way coupled strategy: the flow in porous media is simulated with a finite volume-based commercial software for both cases, while the solid problem is solved either with commercial finite element-based software or through a multi-point stress finite volume approximation available in the open-source library PorePy. We investigate the capability of the reduced model to accurately reproduce new scenarios by comparing the outcome with a standard numerical solution. We can observe how the DL-ROM can be used in place of full-physics modeling to rapidly compute the stress state along a fault and the characteristic quantities used to estimate whether subsurface operations can destabilize or not the fault itself. After training, the DL-ROM is computationally inexpensive, enabling multi-query analysis for statistical investigations on fault stability. Even though this is not the first example ofa surrogate model for fault stability problems, to the best of the authors' knowledge, this is at least one of the first examples of a data-driven DL-ROM used in this context.