A physics-guided neural network-based inversion framework for monitoring of CO2 storage via angle-dependent seismic images

Abstract

Time-lapse (4D) seismic is a critical technique for reliable monitoring of geological CO₂ storage. Its effectiveness depends on converting time-lapse seismic observations into quantitative estimates of CO2 saturation. We propose a neural-network-based framework trained with a physics-guided objective that embeds seismic and rock-physics relationships. The network adopts a U-Net-style encoder–decoder architecture with residual convolutional blocks and a Transformer bottleneck. The network predicts the non-wetting-phase saturation field from observed angle-dependent seismic reflection data. The utilized loss function compares physics-based predicted angle-dependent 4D seismic images with observed ones and also includes smoothness priors to enforce lateral continuity while suppressing spatial blurring in the estimated saturation fields. We evaluate the proposed framework on synthetically ‘observed’ time-lapse multi-angle seismic images generated from reference saturation fields produced by an 800-year flow simulation. The DNN is trained in a self-supervised manner on the first 720 years, with the remaining 80 years held out for further evaluation. The framework reproduces the large-scale evolution of two migrating plumes over the full 800-year sequence, consistently recovering plume locations, migration paths, and extents. These results indicate the framework is practical for monitoring tasks emphasizing plume localization and extent, enabling fast, stable inference of saturation changes from time-lapse multi-angle images.