Hongkyu Yoon, Teeratorn Kadeethum, Sandia National Laboratories; Jonghyun Harry Lee, University of Hawaii at Manoa
This work presents a novel framework for deep learning-based CO₂ flow modeling of pressure and saturation distribution over time and data assimilation approach for real-time forecasting. Geologic carbon storage (GCS) is considered one of scalable strategies to mitigate the climate change due to greenhouse gas emission. Scalable GCS operations require fast forward modeling capability with data assimilation with real-time monitoring data for forecasting and operation optimization. First, combination of convolutional neural networks (CNNs), long-short term memory (LSTM), and deep neural networks (DNNs) is employed to construct deep learning (DL)-based surrogate modeling of predicting multiphase CO₂ flow, pressure propagation, and displacement if poroelsticity is considered, over time in heterogeneous subsurface fields. Physics-based loss functions are also used to improve the prediction accuracy. Second, a variational autoencoder (VAE) in the data assimilation (DA) framework is developed for a real-time history matching of CO₂ and pressure plume development. The encoder in VAE works as a nonlinear dimension reduction method that determines a low-dimensional latent space “z” with encoded/compressed information, possibly performing better than traditional linear dimension reduction methods. For data assimilation, the latent space z is updated conditioned on the observed data for effective sampling with better convergence on the low dimensional subspace. The decoder in VAE using the updated latent data is used to generate distribution of state variables (i.e., permeability and porosity), which are used as input to CNN-LSTM-DNN model for fast forecasting of CO₂ saturation and pressure plume development. With computationally efficient CNN-LSTM-DNN forward model the fast data assimilation can be completed in a timely fashion. In addition, a self-supervised learning approach with the Barlow Twins concept is adopted to enhance predictive accuracy as a ML-driven forward model. SNL is managed and operated by NTESS under DOE NNSA contract DE-NA0003525.