Christopher Scott Sherman, Hewei Tang, Pengcheng Fu, Joseph P. Morris, Lawrence Livermore National Laboratory
Estimating reservoir pressure distribution in geological carbon storage (GCS) projects plays a key role in forecasting induced seismicity and detecting potential leakage. Due to the high drilling cost, GCS projects usually have spatially sparse measurements from wells and limited prior knowledge for initial geologic descriptions of saline aquifers. These characteristics lead to significant uncertainties in predicting a spatial distribution of reservoir pressure and evolution of the pressure distribution. To address this challenge, we propose to use low-cost Interferometric Synthetic-Aperture Radar (InSAR) data as the monitoring data to infer reservoir pressure build up. We present a deep learning accelerated workflow to assimilate surface displacement map interpreted from InSAR and to forecast dynamic reservoir pressure. The target application is fast decision making for carbon storage reservoir management.
The workflow was developed based on an Ensemble Smoother Multiple Data Assimilation (ES-MDA) framework to naturally handle uncertainty quantification. We consider an inverse estimation of non-Gaussian porosity and permeability fields with heterogeneous porosity/permeability distributions in each facies. This high-dimensional inverse problem requires a significant number of forward model runs to obtain converged results. We applied two strategies to speed up the workflow. First, we reduce the dimension of the parameter space through a principal component analysis (PCA)-based parameterization strategy. The traditional PCA method cannot be applied to multimodel distributions (introduced by different facies). We apply a convolutional neural network (CNN) as a post process filter for the PCA realizations. The method efficiently reduces the parameter space dimensionality for more than one hundredfold. Second, we develop two residual U-Net based surrogate models for surface displacement and reservoir pressure predictions respectively. The surrogate model significantly reduces a single forward model evaluation time from 2 core hours to 0.06 core seconds. The application of the two deep learning models helps reduce the computational cost of the data assimilation and forecasting workflow by more than 300 times.