Geological carbon storage (GCS) is the most popular technique for sequestering CO2. Usually, GCS is modeled using commercial numerical simulators to make CO2 forecasts for mapping CO2 subsurface movement. However, simulators require high computational resources for complex problems. In this paper wavelet and Fourier neural operator (WNO and FNO) based – machine learning models were employed to rapidly forecast the reservoir pressure and CO2 saturation distribution, under fixed injection locations. Two geological models, SACROC and Sleipner, were used to generate CO2 injection datasets. The efficacy of the WNO model was evaluated through CO2 forecasts on the SACROC dataset, while the trained FNO model on the SACROC dataset was "transferred" to make predictions on the Sleipner dataset. The WNO-based ML method was accurate and efficient, such that the overall mean relative errors for pressure and saturation predictions on the test set were 2.21% and 0.84% respectively. More importantly, the WNO-ML algorithm reduced the prediction time by 90%. The overall mean relative errors for pressure and saturation prediction using transfer learning with the FNO algorithm were 2.48% and 1.79% respectively. Additionally, the application of transfer learning reduced the machine learning model training time and data storage requirement by 61% and 45% respectively. Through reduction of the computational time and data storage requirements for CO2 forecasting, transfer learning makes it possible to conduct more detailed and accurate forecasts, which can help to improve the safety and efficiency of CO2 storage projects.

You can access this article if you purchase or spend a download.