Prediction of flow and transport behavior in complex geologic formations is critical for the efficient development of the underlying water resources. Accurate simulation models can provide reliable forecasts about the response of subsurface flow systems to various development strategies and can be used to guide future planning and to optimize performance. Intrinsic rock flow properties, such as permeability and porosity, play an important role in constraining the spatiotemporal evolution of groundwater flow and transport processes in aquifers (Pirot et al., 2015). These properties tend to exhibit complex architectures and strong multi-scale spatial heterogeneity, making their representation a nontrivial task. The cost and technical difficulties associated with measuring these properties at a high enough resolution and coverage have necessitated the use of indirect and scattered flow response measurements for their characterization, which is often carried out through an inverse modeling or model calibration process (
Reservoir model calibration against dynamic response data is often constrained by a prior conceptual model of geologic scenario that specifies the expected types of spatial variability and features in the solution. However, geologists have significant uncertainty in developing a conceptual model, e.g., due to limited data, process-based modeling assumptions, and subjectivity. Therefore, it is prudent to consider the uncertainty in the geologic scenario when solving the model calibration problem as it will provide an opportunity to utilize the response data in supporting or rejecting the proposed scenarios. We present a new approach for geologic scenario identification based on dynamic response data by combining gradient-based inversion for feature extraction and a convolutional neural network for feature recognition. To compactly represent each scenario while ensuring data sensitivity, the approach relies on extremely low-rank parameterization of individual geologic scenarios based on the Principal Components Analysis (PCA). The PCA basis elements of each scenario are then combined to capture the salient features in any of the scenarios, or their possible combinations. An iterative least-squares formulation is then formulated to detect scenarios that are supported by the observed data. The inversion results in an approximate (smooth) spatial solution that contains the dominant spatial patterns. A pre-trained convolutional neural network (CNN) is then used to identify the relevant geologic scenarios based on the reconstructed spatial solution. Two main advantages of the workflow include: (i) the ability to combine different scenarios if supported by data, instead of evaluating individual scenarios, and (ii) efficient gradient-based implementation that does not require extensive forward simulation runs. In addition, the training of CNN is implemented using only geologic realizations without requiring additional reservoir simulation. The performance of the workflow is evaluated using tomographic inversion and model calibration in fluvial reservoirs.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.