An Efficient Hybrid Deep Learning Framework for CO2 Plume Monitoring in Geological Sequestration

Abstract Geological CO2 sequestration (GCS) represents a critical technology for mitigating CO2 emissions, but faces significant challenges in environmental monitoring and risk assessment. While physics-based reservoir simulations provide accurate predictions, their computational cost severely limit practical applications. Deep learning-based surrogate models offer 100 to 1000 times speedup but meet training difficulties for large-scale systems. Moreover, practical GCS monitoring relies on seismic technologies limited to detecting CO2 saturation changes, necessitating rapid predictive methods aligned with real-world detection capabilities. This study introduces a novel deep learning workflow for GCS to predict seismicity-detectable CO2 plume fronts. Following established practices for seismic monitoring data interpretation, continuous saturation plume fields are converted to binary maps using a 10% CO2 saturation threshold, to mimic the practical detection limit of current seismic technologies. The workflow employs a hybrid CNN-Transformer architecture. Convolutional neural networks (CNNs) excel at spatial feature extraction, while transformer networks have strength in sequence modeling. In order to capture complex spatiotemporal CO2 plume evolution patterns, the framework comprises: a geological encoder using 3-dimensional (3D) CNNs to process high-dimensional geological properties (porosity and permeability), a well-control encoder coupled with Transformer blocks to handle dynamic well control sequences, and ultimately a decoder to reconstruct seismicity-detectable saturation distribution. Based on the large-scale 3D SPE11-C reservoir model, the hybrid CNN-Transformer model reaches an average binary cross-entropy (BCE) loss at 0.099 ± 0.0788 and Dice score at0.799 ± 0.181, and achieves nearly 10,000 times (on GPU) and 3,500 times (on CPU) acceleration over reservoir simulation in the prediction. Therefore, the hybrid CNN-Transformer model successfully captures essential spatiotemporal dynamics of CO2 plume evolution, and is expected to provide a better balance between computational speed and physics fidelity.