A Deep-learning-based Surrogate Model for Data Assimilation in Geologic Carbon Storage

Abstract:

In geologic carbon storage (CO₂ injecting into deep saline aquifers), predicting CO₂ plume migration and pressure field distribution is essential for multiple reservoir management tools, such as history matching, well control optimization, etc. The computational costs of these tasks are extremely high if a high-fidelity reservoir model is employed. To address this problem, we propose to leverage physical understandings of porous medium flow and transport behavior with deep-learning (DL) techniques to develop a surrogate model targeting complex geologic carbon storage scenarios.

The DL-based surrogate model generates accurate predictions of well pressure, CO₂ plume extent, and overpressure distribution for any given combination of random geo-model realization (porosity and permeability fields), injection rate, and injection duration. The model was developed by combining a deep convolutional encoder-decoder network and a recurrent neural network. After training with a wide range of reservoir properties and under various injection conditions, the model was separately tested on an interpolated dataset and an extrapolated dataset beyond the range of training data.

Three types of (synthetic) observation data are considered by the model: (1) pressure and CO₂ saturation from monitoring wells, (2) CO₂ plume extents inferred from seismic survey, and (3) reservoir overpressure distribution based on Interferometric Synthetic Aperture Radar (InSAR) images. The incorporation of various observation data reduces the uncertainty in predictions. The capability of handling reservoir dynamic state changes efficiently will be an important technology breakthrough in reservoir management.