Abstract:Monitoring CO 2 plumes throughout the operation of geologic CO 2 sequestration projects is essential to environmental safety. The evolution of underground CO 2 saturation can be predicted using high-fidelity numerical simulations. However, high-fidelity simulations can be prohibitively expensive to compute. As a result of recent developments in data-driven models, rapid predictions of the CO 2 plume can now be made using readily available pressure and temperature measurements. This study presents a novel deep learning-based workflow for efficiently visualizing CO 2 plumes in near real-time while considering their uncertainties. In our deep learning workflow, we visualize the CO 2 plume images in the reservoir as a propagating saturation front, represented by 'onset time', using field measurements, including downhole pressure and temperature. At a given location, the 'onset time' is the calendar time when the CO 2 saturation exceeds a certain threshold. Therefore, a single image of the CO 2 front propagation is captured using the 'onset time' rather than storing multiple CO 2 saturation images at different time steps. The use of 'onset time' significantly reduces memory and computational cost of deep learning-based framework, enabling large-scale field applications. We use a variational autoencoder-decoder (VAE) network to compress high dimensional 'onset time' images into low dimensional latent variables while considering uncertainties of the predicted images. The use of VAE and onset time, simplifies the overall neural network architecture and significantly enhances the training efficiency. To estimate the latent variables of the VAE network, we train a feed forward neural network model that incorporates available monitoring data such as downhole pressure and temperature measurements. The estimated latent variables are then fed into a trained decoder network to generate 3D onset time images, visualizing the propagation of CO 2 plume in near real time. The proposed workflow is applied to both synthetic and field cases, where the field application is a large-scale geological carbon storage project in a carbonate reef reservoir in the Northern Niagaran Pinnacle Reef Trend in Michigan, USA. The field measurements include distributed temperature sensing (DTS) and distributed pressure responses from down-hole gauges along the monitoring well. The predicted CO 2 plume images provided by the proposed workflow are shown to be consistent with the simulation results of traditional history matched model using genetic algorithm. Our deep learning-based framework can predict the CO2 front propagation in terms of onset time in seconds, making it well-suited for real time decision-making and operational optimization.

Advancing spatiotemporal forecasts of CO 2 plume migration using deep learning networks with transfer learning and interpretation analysis

Spatio-temporal neural networks for monitoring and prediction of CO 2 plume migration from measurable field data

An efficient deep learning-based workflow for CO 2 plume imaging considering model uncertainties with distributed pressure and temperature measurements

A Robust Deep Learning Workflow to Predict Multiphase Flow Behavior during Geological CO2 Sequestration Injection and Post-Injection Periods

Deep learning for characterizing CO 2 migration in time-lapse seismic images

A Robust Autoregressive Long-Term Spatiotemporal Forecasting Framework for Surrogate-Based Turbulent Combustion Modeling Via Deep Learning

A deep learning-accelerated data assimilation and forecasting workflow for commercial-scale geologic carbon storage

LSTM-driven Forecast of CO2 Injection in Porous Media

Towards efficient and accurate CO 2 sequestration optimization: Integrating hierarchical spatio-temporal information into deep-learning-based surrogate models

CO2 Leakage Rate Forecasting Using Optimized Deep Learning

Time-lapse seismic inversion for CO 2 saturation with SeisCO2Net: An application to Frio-II site

Learning CO$_2$ plume migration in faulted reservoirs with Graph Neural Networks

Deep learning applied to CO 2 power plant emissions quantification using simulated satellite images

Spatiotemporal Predictions of Toxic Urban Plumes Using Deep Learning

Accurate and Rapid Forecasts for Geologic Carbon Storage via Learning-Based Inversion-Free Prediction

Uncertainty Quantification Based on Deep-Learning Approach Integrating Time-Lapse Seismic Data for Geological Carbon Storage

Uncertainty Quantification of CO2 Plume Migration in Highly Channelized Aquifers Using Probabilistic Convolutional Neural Networks

Connect the Dots: In Situ 4D Seismic Monitoring of CO2 Storage with Spatio-temporal CNNs

Multiphase flow prediction with deep neural networks

Deep Learning-Based Geological Parameterization for History Matching CO2 Plume Migration in Complex Aquifers

Real time hydrogen plume spatiotemporal evolution forecasting by using deep probabilistic spatial-temporal neural network