We present a novel framework for generating reduced-order models that combines agglomeration of cells from existing high-fidelity reservoir models and flow-based upscaling. The framework employs a hierarchical grid-coarsening strategy that enables accurate preservation of geological structures from the underlying model. One can also use flow information to distinguish regions of high or low flow, and use this division, or other geological or user-defined quantities, to select and adapt the model resolution differently throughout the reservoir. Altogether, the framework provides a wide variety of coarsening strategies that allow the user to adapt the reduced model to important geology and explore and identify the features that most impact flow patterns and well communication. By preserving these features, while aggressively coarsening others, the user can develop reduced models that closely match an underlying high-fidelity model. Various types of simple flow diagnostics based on time-of-flight and volumetric well communication are used to predict the accuracy of the resulting reduced models.
In this paper, we systematically apply this framework to the Great White Field, but also present results from other real or synthetic models, to demonstrate the asymptotic scaling of accuracy metrics with coarsening levels. Our aim is to identify and illustrate best practices when designing and improving coarsening strategies that can guide future applications of the framework to other reservoir models. We also discuss practical limitations when applying the framework to new simulation models where flow regimes or geologic features may differ.