Reservoir Model Optimization under Uncertainty

Computerized reservoir simulation models are widely used in the industry to forecast the behavior of hydrocarbon reservoirs and connected surface facilities over long production periods. These simulation models are increasingly complex and costly to build and often use millions of individual cells in their discretization of the reservoir volume. Simulation processing time and memory requirements increase constantly and even the utilization of ever faster computers cannot stem the growth of simulation turnaround time. On the other hand, decision makers in reservoir management need to quickly assess the risks associated with a certain model and production strategy and need to come up with high/low scenarios for Net Present Value (NPV) and the likelihood of these scenarios. To achieve reduced turnaround time in this difficult environment, reservoir engineers and applied mathematicians explore ways to employ optimization techniques that use surrogate models (i.e. a response surface) to perform these tasks. The costly simulation model is used to seed the design space and to assist with local refinement of the surrogate model. In this report we summarize our findings from the IMA Math Modeling Workshop, in which we examined different algorithms to build surrogate models for a simple oil reservoir. Besides estimating NPV for certain high/low scenarios, we used surrogate models to find optimal producer well locations and to perform simple history matching of a 2D permeability field.