Abstract
There are numerous complex characteristics that impact the long-term decline behavior of wells in tight oil and gas reservoirs. Numerical simulation with fine spatial discretization is required to capture important characteristics such as: the presence of natural fractures (stimulated and unstimulated); changes in total fluid mobility and compressibility; heterogeneous matrix and fracture properties (including permeability loss due to compaction); early flow transients; and nonuniform stimulation during hydraulic fracturing. This discretization requirement, when combined with the need to include multiple wells to capture interference effects, can result in model sizes in the tens of millions of cells over a few 640-acre sections. While these model sizes can sometimes be addressed with current-generation simulators, the excessive run time limits the ability to simulate multiple realizations for history matching and sensitivity analysis. In practice, lower fidelity simulations are often substituted.
We present our efforts to help eliminate this tradeoff by building a fully-implicit black-oil simulator that combines recent advances in simulation algorithms with the high performance of GPUs. All major computational tasks are executed on GPU, including property evaluation, Jacobian construction and assembly, and linear solution with CPR-AMG. This approach allows models with many millions of cells to be simulated within minutes on a single workstation with multiple GPUs. For example, on a tiled SPE 10 model with 55 million cells and 250 wells, we simulate 2000 days of production in 20 minutes using eight GPUs. We summarize our approach to address the challenges in building a fine-grained, scalable simulator. We discuss two challenges in particular: the need to expose massive parallelism while retaining the robustness of linear solvers; and managing the complexity of the many features required by engineers for practical application.
We discuss how we apply this fully-accelerated technology to increase fidelity and throughput on tight reservoir workflows, improving our understanding of the complex nature of production decline and possible long-term well interference. Furthermore, we illustrate workflows for simulations based on detailed reservoir/fracture descriptions.
