Key points are not available for this paper at this time.
Accurate and efficient forecasting of CO2 saturation-field evolution is essential for screening injection–production strategies in CO2-EOR/CCUS, yet conventional simulators are too expensive for rapid multiscenario evaluation and many data-driven models generalize poorly across well-pattern topologies. To bridge these gaps, we propose a multimodal, physics-constrained, generative spatiotemporal network (MPG-STNet). The model employs an enhanced Fourier neural operator to perform efficient, physics-aware inference on static geological fields, a Graph Attention Network to explicitly encode well-pattern topology, and a temporal Transformer encoder to extract features from dynamic three-phase production sequences. These signals jointly condition a conditional generative adversarial network that produces high-fidelity CO2 saturation fields, thereby coupling spatial, temporal, and physical information within a single generative pathway. We further embed physics constraints on the discriminator side─most notably material balance─treating mass conservation as a hard criterion to enhance the physical acceptability of the generated fields. Evaluated on a high-fidelity data set of 4,500 simulated cases covering nine heterogeneous well patterns over a 150-month horizon, MPG-STNet achieves a mean squared error of 0.0062 and structural similarity index measure of 0.9788 on the test set and maintains strong performance on unseen well-pattern configurations. The proposed surrogate provides fast, physically plausible saturation-field predictions for well-pattern generalization and operational decision support.
Zhao et al. (Wed,) studied this question.