Enhancing the Performance of Geomechanical Reservoir Simulations through Multiscale Preconditioning Techniques

The simulation of reservoir geomechanical models presents significant challenges. The need to include under, over and side burden, besides the reservoir itself, results in meshes that can reach several hundred million elements. In this work, we explore a multiscale methodology designed to enhance the performance of iterative solvers in reservoir geomechanical simulations, leveraging our previous findings presented in Figueiredo Et. Al., "A parallel viscoplastic multiscale reservoir geomechanics simulator" (Upstream Oil and Gas Technology, 2022). Our approach employs multiscale techniques as a preconditioning strategy to accelerate the convergence of the conjugate gradient method, which is essential given the complexity and size of the problems we address. In this study, we focus on optimizing the parameters of our iterative method to maximize efficiency and convergence speed when solving large scale geomechanical models. We will systematically investigate the impact of various configurations for iteratively solving the coarse scale problem, including single versus double precision, ILU(0) versus ILU(1) preconditioning, and different convergence tolerances. Our goal is to identify the optimal settings that yield the best overall performance for practical geomechanical simulations. Through a series of real field geomechanical test cases, we will demonstrate the impact the configuration of the iterative linear solver for the coarse problem has on the effectiveness of our proposed methodology. These test cases will be executed in parallel on a high-performance computing environment utilizing both distributed and shared memory architectures. We will utilize the linear solver framework outlined in Gasparini et al., "Hybrid parallel iterative sparse linear solver framework for reservoir geomechanical and flow simulation" (Journal of Computational Science, 2021), to ensure robust and efficient computations. Particular emphasis will be placed on the advantages of employing single precision computations, which have shown promising results in our initial tests. This choice aligns with current trends in computational practices, highlighting the potential for significant performance improvements mixed precision techniques have on modern hardware.