Behnam Jafarpour

University of Southern California

Field data-driven CO₂ storage monitoring and prediction with spatial-temporal neural networks

Successful implementation of geologic CO₂ storage projects hinges on reliable and effective monitoring. Traditionally, visualizing and monitoring the CO₂ plume involve integrating field measurements into complex simulation models, which are not suitable for real-time field operation. Furthermore, these models require input parameters such as rock flow properties that are intrinsically heterogeneous and challenging to characterize, introducing considerable uncertainty. We present a novel deep learning framework to efficiently reconstruct and predict the migration of the CO₂ plume using only field measurements, eliminating the need for uncertain geological inputs. The approach utilizes a robust spatiotemporal convolutional neural network that is trained on simulated data across diverse geological scenarios, allowing it to capture the dynamic evolution of the plume without being constrained by specific geological scenario. Once trained offline, the model integrates global and local field measurements from multiple sources to reconstruct the CO₂ plume and forecast its spatiotemporal evolution in real-world applications.