Multi-Modal Data-Driven Fracture Response Surrogate: Integrating Development History and Prior Knowledge for Efficient CO2 Storage Assessment

Abstract

Abstract Depleted deep carbonate reservoirs with natural fracture networks are attractive targets for CO2 storage, yet they pose leakage risks from fracture reactivation. Existing approaches use numerical simulation and neural networks in isolation, sacrificing either interpretability or efficiency. To bridge this gap, we propose a knowledge graph enabled multi modal fusion framework. A physics constrained graph neural network is pretrained on over 500 synthetic fracture geometries and then finetuned with field monitoring data (microseismic, fiber optic strain, and pump curves). The resulting surrogate provides rapid, interpretable 3D fracture characterization. Applied to a well in the Sichuan Basin, the surrogate delivers high accuracy predictions within minutes, dramatically accelerating conventional workflows. Virtual optimization over multiple realizations yields three key insights balancing efficiency and safety. First, a quantifiable safe injection window is identified to prevent uncontrolled aperture dilation and shear failure. Second, flow heterogeneity is pronounced, with most conductivity concentrated in dominant fracture clusters, which is critical for leakage assessment. Third, incorporating realistic rough surfaces not only increases effective storage capacity but also enhances mineral trapping, offering dual benefits in security and capacity. The framework consistently outperforms purely data driven models, especially under limited data, providing a robust pathway that integrates prior physics knowledge with site specific monitoring for secure CO2 storage assessment.