Abstract
Reduced-order models are much in demand to make field management optimization and uncertainty quantification computationally tractable. In this talk, I will outline some recent graph-based approaches and compare two specific approaches.
The interwell family of methods, exemplified by GPSNet (Ren et al., 2019), represent the reservoir as a graph of 1D interwell connections. In the alternative CGNet approach, the graph topology mimics the intercell connections of a traditional 3D grid. Both model types can be trained to match well responses obtained from underlying fine-scale simulations. However, our experience shows that CGNet is easier to setup/train and generalizes better than GPSNet.
The interwell family of methods, exemplified by GPSNet (Ren et al., 2019), represent the reservoir as a graph of 1D interwell connections. In the alternative CGNet approach, the graph topology mimics the intercell connections of a traditional 3D grid. Both model types can be trained to match well responses obtained from underlying fine-scale simulations. However, our experience shows that CGNet is easier to setup/train and generalizes better than GPSNet.
