A Predictive DeepONet Model of CO2 Plume Propagation in Geological Formations Based on Sparse Well Data

Numerical simulations for predicting stored CO2 behaviour, while informative, can be computationally costly, making rapid real-time monitoring of the plume propagation challenging. Machine learning offers a data-driven approach to monitoring carbon storage operations. Recent studies have demonstrated the potential of machine learning to predict CO2 plume propagation using sparse well data, though they require substantial, large datasets to be accurate. On the other hand, deep operator networks (DeepONets) efficiently map inputs to outputs with smaller datasets, reducing error and overfitting, and combining DeepONets with self-supervised learning has proven effective for geologic carbon storage. This research investigates DeepONets’ performance as a carbon storage proxy, comparing results using full or sparse porosity/permeability data. Given sparse well data, the most fundamental question is whether DeepONets may be used as a proxy for monitoring the CO2 plume propagation in the subsurface.