Multi-Pattern Generalization in CO₂‑EOR: Physically Consistent Surrogate for Saturation-Field Evolution.

Area of Science:

• Petroleum Engineering
• Artificial Intelligence
• Geoscience

Background:

• Accurate forecasting of CO2 saturation-field evolution is crucial for optimizing CO2-Enhanced Oil Recovery (EOR) and Carbon Capture, Utilization, and Storage (CCUS) strategies.
• Conventional simulation methods are computationally expensive for rapid scenario analysis, and existing data-driven models struggle with diverse well-pattern topologies.

Purpose of the Study:

• To develop a novel, efficient, and physically constrained surrogate model for predicting CO2 saturation-field evolution.
• To address the limitations of conventional simulators and data-driven models in handling complex well patterns and dynamic production data.

Main Methods:

• Proposed a multimodal, physics-constrained, generative spatiotemporal network (MPG-STNet).
• Integrated an enhanced Fourier Neural Operator for physics-aware inference on static geological data.
• Utilized a Graph Attention Network for well-pattern topology encoding and a temporal Transformer for dynamic production sequence analysis.
• Employed a conditional generative adversarial network for high-fidelity CO2 saturation field generation, incorporating material balance constraints in the discriminator.

Main Results:

• MPG-STNet achieved a mean squared error of 0.0062 and a structural similarity index measure of 0.9788 on a test set of 4,500 simulated cases.
• The model demonstrated strong generalization capabilities on unseen well-pattern configurations.
• The surrogate model provided fast and physically plausible predictions, outperforming conventional methods in efficiency and accuracy for complex scenarios.

Conclusions:

• The MPG-STNet offers a significant advancement in forecasting CO2 saturation-field dynamics for CO2-EOR/CCUS applications.
• The model's ability to integrate spatial, temporal, and physical information enables robust predictions across diverse operational settings.
• This physics-constrained generative approach enhances decision-making for injection-production strategies and operational support.