Petrophysical characterization of deep saline aquifers for CO2 storage using ensemble smoother and deep convolutional autoencoder

Abstract

Carbon dioxide sequestration in deep saline aquifers requires accurate and precise methods to monitor carbon capture and storage and detect leakage risks. The assessment of the CO2 plume location during injection and storage depends on the accuracy of the spatial distribution model of petrophysical properties, such as porosity and permeability. This work focuses on stochastic methods for petrophysical characterization and presents a method for the prediction of porosity and permeability using borehole observations and surface geophysical data. This study utilizes injection and monitoring measurements at the borehole locations and time-lapse seismic surveys. The proposed method is based on a stochastic approach to inverse problems with data assimilation, namely the ensemble smoother with multi-data assimilation. Ensemble-based methods are generally unfeasible when applied to large geophysical datasets, such as time-lapse seismic surveys. In the proposed approach, a machine learning method, namely the deep convolutional autoencoder, is applied to reduce the dimension of the seismic data. The ensemble smoother is then applied in a lower dimensional data space to predict the aquifer petrophysical properties. This method updated predictions of porosity and permeability every time new data, either seismic surveys or borehole data, are available, to reduce the uncertainty in the CO2 plume prediction. The method is tested and validated on a synthetic geophysical dataset generated for the Johansen formation, a potential large-scale offshore site for CO2 storage.