Integrating Neural Operators and Transfer Learning for Efficient Carbon Storage Forecasting

Abstract Efficient forecasting involves generating accurate and reliable predictions while minimizing the use of resources, such as time, training data, and computational power. This study investigates the potential of Fourier Neural Operators (FNOs) combined with transfer learning (FNO+TL) for forecasting the spatiotemporal evolution of pressure and CO2 saturation in a large, heterogeneous geological carbon-storage aquifer. The enhanced FNO models, trained on the SACROC geological model, achieved high accuracy, with mean absolute errors (MAEs) of 0.046 MPa for pressure and 0.011 for saturation in large aquifer models. Transfer learning (TL) further enabled the adaptation of a trained FNO model to diverse operational conditions with minimal new training data. The adapted models achieved pressure prediction MAEs of 0.27 MPa, 0.51 MPa, and 0.59 MPa, and saturation prediction MAEs of 0.05, 0.056, and 0.14, under varying injection rates, schedules, and injector locations, respectively. Additionally, the FNO models generalized successfully to a geologically distinct site at the Illinois Basin - Decatur Project (IBDP), which had significantly different carbon storage operations. The models achieved prediction MAEs of 4.5 psia for pressure and 0.11 for saturation, with reductions of 87% and 80% in training data requirements, respectively. Furthermore, the FNO+TL models significantly reduced computational costs. The study concludes that integrating transfer learning with Fourier Neural Operators provides an efficient and scalable approach for CO2 storage forecasting.