Deep Learning Assisted History Matching and Forecasting: Applied to the Illinois Basin – Decatur Project (IBDP)

Abstract

To successfully deploy large-scale carbon capture and storage (CCS), it is necessary to have a thorough understanding of subsurface storage formations, given the significant uncertainty in geological properties, such as permeability and porosity. Although many conventional history matching (HM) approaches that integrate numerical flow simulations are available, they often require numerous time-consuming flow simulations (each simulation may take a few hours to a few days to complete), they thus may not be practical for large-scale geological scenarios. To address this challenge, we have developed a fast deep learning-based proxy model using Fourier Neural Operator (FNO) to replace the flow simulations in an ensemble-based HM framework for calibrating uncertain geological parameters. Specifically, we applied the Ensemble Smoother with Multiple Data Assimilation (ES-MDA) algorithm in our HM task for the data obtained from the Illinois Basin - Decatur Project (IBDP). Our results reveal that the trained FNO model can provide reliable and fast predictions of reservoir pressure when compared to conventional flow simulations. Additionally, excellent HM performance was achieved with this HM workflow, as the history-matched realizations closely match the observational data. Our results indicate that a well-trained deep learning-based proxy model can effectively replace full-physics numerical flow simulations, significantly improving the computational efficiency of the HM process.

Keywords: Carbon capture and storage, history matching, IBDP, CO2 injection