Rapid High-Fidelity Forecasting for Geological Carbon Storage Using Neural Operator and Transfer Learning

Abstract

Abstract Carbon sequestration is a promising technique to minimize the emission of CO2 to the atmosphere. However, the computational time required for CO2 forecasting using commercial numerical simulators can be prohibitive for complex problems. In this work, we propose the use of transfer learning to rapidly forecast the CO2 pressure plume and saturation distribution under uncertain geological and operational conditions, specifically for variations in injector locations and injector rates. We first train a Fourier Neural Operator (FNO)-based machine learning (ML) model on a limited set of simple scenarios. Then, we use transfer learning to fine-tune the FNO model on a larger set of complex scenarios. Most importantly, the CMG forecasting time for one scenario requires approximately 40 to 50 minutes, which was drastically reduced to 12 seconds by using Fourier Neural Operator and then reduced further to 8 seconds by implementing transfer learning on the Fourier neural operator. The mean relative errors 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 operator to learn complex task with less amount of data and low training time. Our results show that transfer learning can significantly reduce the computational time required for CO2 forecasting. The data generation and model training times were reduced by 50% and 75%, respectively, by using transfer learning on the Fourier neural operator. Additionally, the total number of trainable parameters was reduced by 99.9%. 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 CO2 forecasting and to help mitigate the risks associated with CO2 leakage.