Data-Driven Model Reduction, Scientific Frontiers, and Applications ()
- Texas A&M University
- College Station, TX
- Joe C. Richardson Petroleum Engineering Building (RICH) 910
- Yusuf Falola, Harold Vance Department of Petroleum Engineering
- Neural Operator-Based Rapid Forecast of CO2 Pressure and Saturation Distribution During Geological Carbon Storage
Geological carbon storage is a proven technique for storing CO2 in the subsurface, thereby reducing global warming. Although this process could result in environmental concerns such as underground drinking water contamination and micro-seismicity, prior forecasts of the CO2 injection can minimize these risks. The computational time required for CO2 forecast using commercial numerical simulators can be prohibitively long for complex problems. Therefore, in this work, we leverage Fourier Neural Operators for rapid forecast of CO2 pressure plume and saturation distribution during geological carbon storage. Additionally, we propose the application of transfer learning to rapidly forecast CO2 sequestration under varying geologies. Firstly, we evaluated the efficacy of a Fourier Neural Operator (FNO)-based machine learning (ML) model on the SACROC (USA) Geomodel. Thereafter, we used transfer learning to fine-tune and evaluated the FNO model on the Sleipner (Norway) Geomodel. Most importantly, the simulator forecasting time for one scenario requires approximately 40 to 50 minutes, which was drastically reduced to 3.5 minutes by using Fourier Neural Operator. The mean relative errors (MRE) of the neural operator predictions of pressure and saturation were 1.42% and 7.9%, respectively. These errors get slightly higher when transfer learning is implemented on neural operators to learn complex tasks with less amount of data. Thus, the MRE for pressure and saturation distribution using transfer learning is 2.48% and 8.8% respectively. The data generation and model training times were reduced by 50% and 61%, respectively, by using transfer learning on the Fourier neural operator. Our results demonstrate the potential of transfer learning for rapid forecasting of CO2 pressure plume and saturation distribution. This technique can be used to improve the efficiency of geological carbon storage and to help mitigate its associated risks.