CO2 Saturation Prediction from Historical Time-Lapse Seismic Data Using Physics-Constrained VideoMAE

Geological carbon storage is a key strategy for mitigating global CO₂ emissions, and reliable monitoring of subsurface CO₂ migration is critical for storage safety. Time-lapse seismic provides valuable insights into CO₂ plume evolution. However, accurately predicting high-resolution CO₂ saturation from seismic data remains a major challenge. In this study, we propose a novel physics-constrained deep learning framework that treats time-lapse seismic data as video sequences and leverages the Video Masked Autoencoder (VideoMAE) architecture to capture spatial and temporal dependencies. The approach consists of two stages: self-supervised pretraining on seismic data and supervised fine-tuning for CO₂ saturation prediction. During pretraining, masked reconstruction enables the model to extract rich spatiotemporal feature representations from seismic videos. In fine-tuning, the pretrained model is adapted to predict future CO₂ saturation from historical time-lapse seismic data without requiring seismic data from the target year. A physical constraint based on Fick’s law of diffusion is incorporated into the loss function to regularize the temporal evolution of CO₂ saturation during fine-tuning. Results on the Kimberlina synthetic multiphysics dataset demonstrate that the physics-constrained VideoMAE framework consistently outperforms baseline models in both prediction accuracy and spatial consistency. These findings highlight the effectiveness of combining video-based self-supervised learning with physical constraints for time-lapse seismic monitoring and provide a promising physics-informed approach for CO₂ storage surveillance.