Realtime forecasting of CO2 flow using variational autoencoder with ensemble-based data assimilation

Abstract:

This work presents a novel framework for accomplish machine learning-driven CO₂ modeling by combining a variational autoencoder (VAE) with ensemble-based data assimilation (EnDA) for data assimilation, resulting in real-time history matching of CO₂ operations and forecasting CO₂ and pressure plume development. VAE as a generative model can be used to extract a latent representation of the data used in training (i.e., encoder) and perform data generation and reconstruction (i.e., decoder). To demonstrate this method, the permeability/porosity fields are generated from a widely used geostatistical model. For fast testing and optimal implementation of VAE, CO₂ saturation and pressure distribution are simulated using an open-source percolation model (pyPERC) and single phase flow model (MODFLOW). An open-source multiphase flow model, PFLOTRAN is also used to validate the approach of pyPERC and MODFLOW simulations. Then, VAE is trained with porosity, permeability, CO₂ saturation and pressure distributions to generate permeability, CO₂ saturation, and pressure distribution maps. The optimal encoder-decoder NN structure and its hyperparameters are determined to extract the features of these data and develop more robust generative models. 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 such as PCA and SVD. For data assimilation, the latent space z obtained from non-linear dimension reduction through the encoder in VAE will be continuously updated conditioned on the available data instead of the original high-dimensional state parameter space for effective sampling with better convergence on the low dimensional subspace. Overall this work demonstrates that dynamic models developed with ML method(s) and data assimilation approach can be effectively used for fast forecasting of CO₂ saturation and pressure plume development. SNL is managed and operated by NTESS under DOE NNSA contract DE-NA0003525.