A Deep Learning Model for CO₂ Storage in a Depleted Gas Reservoir Using Sparse Well Data

This study uses Deep Operator Networks (DeepONets) combined with Barlow Twins (BT) to monitor geological carbon storage, focusing on predicting CO₂ plume propagation based on sparse well data. Machine learning offers a promising alternative for predicting CO₂ plume behavior. Recent works have successfully showcased the usage of machine learning for predicting the CO₂ plume propagation given sparse well data. This study extends previous work by employing BT-DeepONets with sparse well data alongside 3D/4D seismic inputs to model CO₂ plume propagation in a 3D analog of a depleted gas field.

Our work centers on the Harald field in the North Sea, where various wells provide static logs for training and validating the BT-DeepONet model. The methodology includes simulating scenarios based on these logs coupled with a rock physics model to assess changes in acoustic impedance resulting from gas production and CO₂ injection. This innovative approach employs a combination of static and dynamic well logs, injection history, and 3D/4D seismic data in the training process.

Including seismic data significantly enhances the accuracy of CO₂ plume predictions. This approach demonstrates the potential of BT-DeepONets as a proxy for monitoring geological CO₂ storage, leveraging commonly available data in CCS projects.