Effective monitoring of CO2 plume is critical to environmental safety throughout the life-cycle of a geologic CO2 sequestration project. Although full physics-based techniques such as history matching with numerical simulations can be used for predicting the evolution of underground CO2 saturation, the computational cost of the high-fidelity simulations can be prohibitive. Recent development in data-driven models can provide a viable alternative for rapid prediction of the CO2 plume based on readily available pressure and temperature measurements. In this study, we present a novel deep learning-based workflow that can efficiently visualize CO2 plume in near real-time with considering the uncertainties of CO2 plume images. ‘Deep learning’ refers to a data-driven input-output model development approach involving artificial neural networks with many hidden layers. Our deep learning workflow utilizes field measurements, such as downhole pressure, temperature, and flowrates as input to visualize the subsurface CO2 plume images as a propagating CO2 saturation front in terms of ‘onset time’. The ‘onset time’ is the calendar time when the CO2 saturation at a given location exceeds a specified threshold value. Rather than storing CO2 saturation at multiple time steps, the onset time compresses the data into a single image of CO2 front propagation. To start with, we generate a comprehensive training dataset using flow simulation with diverse geologic model realizations and fluid models. The training data consists of injection rate/pressure at the injection well and measurements in monitoring wells (e.g., distributed pressure and temperature data) and the corresponding CO2 plume propagation onset time maps. To build a machine learning model for CO2 plume evolution, the simple and straight-forward way would be to train a deep learning-based regression model where the input consists of the field measurements, and the output is CO2 saturation distribution at different times. However, the high output dimension of spatial resolution and temporal steps make the training inefficient and impractical. We address this challenge in two ways: first, we output a single onset time map rather than multiple saturation maps at different times; second, we apply a variational autoencoder-decoder (VAE) network that uses lower dimensional latent variables for compressing high dimensional output images, namely the CO2 onset time maps are used as input and output of the VAE network. The use of onset time and image compression using VAE considerably simplifies the deep learning architecture and also makes the training more efficient. In our approach, a feed forward neural network model is trained to predict latent variables of the VAE network. Subsequently the latent variables are fed to the trained decoder network to generate the 3D onset time image, visualizing the evolving CO2 plume in near real time. Since the VAE is used for dimensionality reduction, the trained neural network can provide multiple CO2 plume image predictions considering the uncertainties.The power and efficacy of our approach are demonstrated using both synthetic and field applications. We first demonstrate the deep learning-based CO2 plume imaging workflow using a synthetic example. Next, the visualization workflow is applied to a CO2-enhanced oil recovery and associated geological storage project in a carbonate reef reservoir in the Northern Niagaran Pinnacle Reef Trend in Michigan, USA. The monitoring data set consists of distributed temperature sensing (DTS) data and time-lapse pressure measurements at several locations along the monitoring well. The CO2 plume images obtained from the proposed data-driven approach are compared with the approach based on flow simulation and history matching of 3D geologic models, where similar results are obtained. Additionally, an efficient workflow for optimizing data acquisition and measurement type is demonstrated using our deep learning-based framework.The novelty of this work is the development and application of a deep learning-based framework to interpret field measurements as CO2 plume images. The flexibility of the data-driven workflow allows us to incorporate diverse data types, and the efficiency of the method makes the approach suitable for field-scale applications
