Donald Joseph Remson, David Morgan, U.S. DOE's National Energy Technology Laboratory (NETL); Kolawole Bello, Derek Vikara, NETL Support Contractor
With advancements in data analytics, supervised deep learning (DL) approaches are being used for modeling complex multiphase flow in subsurface reservoir applications. DL can be used to generate surrogate models which can be fast and accurate at estimating complex spatio-temporal response variables. Their prediction speed and accuracy offer utility over traditional reservoirs simulators offering applications towards uncertainty quantification of reservoir properties, optimization of operations, and use for near real-time decision support in subsurface operations. Despite the benefit of DL-based surrogate models, challenges remain in how they can be efficiently utilized in subsurface applications. One of the challenges in utilizing DL-based surrogate models for the capability to predict subsurface processes is that they require lots of data to build the models which can also lead to higher training times and prediction turnarounds. To maintain the efficacy of DL-based models in subsurface applications, it is crucial to appropriately preserve input and output mapping relationships when employing techniques that reduce the dimensionality of the data needed – ultimately enabling sustained accuracy and speed of the model. In this study, we developed deep learning models for CO₂ geologic storage that are capable of accurate prediction of spatio-temporal outputs of CO₂ saturation, pressure, and brine production in 3D space over a storage operation’s injection (30 years) and post-injection (100 years) timeframes. Training datasets used for this study were generated using CMG-GEM and are specific to a carbonate reef depositional environment in the Scurry Area Canyon Reef Operators (SACROC) oilfield located in the Permian Basin. The training datasets encompass 81 realizations assuming a saline storage reservoir system. Training sets vary porosity and permeability of the storage reservoir and CO₂ injection rates to capture operational variability aspects of a CO₂ storage project. The model framework involves ensembling multi-layer autoencoder networks that provide dimensionality reduction of geologic inputs with fully connected long short-term memory (LSTM) neural networks that generate time-series prediction. This approach offers a means to maximize training time efficiency, reduce computational memory burden, and minimize prediction turnaround. The use of autoencoders also offers flexibility for adapting the models to account for alternative realizations of the geology of the SACROC CO₂ storage site. Despite the substantially reduced dimensionality fed to the network, the modeling framework enables minimal compromise to performance accuracy (>99 percent compared to held out testing data) in generated models while providing for extremely fast predictive turnarounds (< 1 second). This research is part of NETL’s SMART-CS initiative Task 5 aimed at developing a virtual learning environment that enables different stakeholders’ ability to explore and test CO₂ storage reservoir behavior.