Fracture Propagation Dynamics Predicted by Data-model-interactive Neural Proxy Model

Abstract: Fracture propagation dynamics in complex geological formations is crucial for understanding cracking mechanism of deep rock and facilitating subsurface reconstruction and resource extraction. However, predominant mechanism-driven numerical model exist several inherent limitations: 1) Over-reliance on empirical formulas and simplified hydraulic fracture propagation model with multi-assumptions restrict its capacity for effectively characterizing the multi-physics coupling in 3-D space, thereby reducing the accuracy of fracture morphology. 2) Computational schemes such as finite element method (FEM) or discrete element method (DEM) involving extensive repetitive calculations, are resource-intensive and exhibit poor temporal efficiency, posing a challenge to engineering requirements. Therefore, a data-model-interactive neural proxy model combining the prior-knowledge from mechanism models and fitting efficiency of deep neural networks, is put forward to depict the fracture propagation dynamics in complex geological formation. Initially, a numerical model for fracture propagation is developed by implementing the 3-D discrete lattice method alongside the elastic-plastic constitutive equation. The coupling of rock deformation and fluid flow is iteratively processed in a stepwise manner to generate a sequence of fracture morphology evolution over time. These mechanism data will provide training samples for the subsequent neural proxy model. Secondly, the efficacy of the neural proxy model is contingent upon the richness and diversity of features presented in the training dataset, necessitating a close approximation of all conceivable scenarios. In light of the irregular spatial distribution of data resulting from the complex geological formation with strong heterogeneity, the Latin Hypercube sampling method is employed to ensure a uniform selection of all conditions, mitigating the potential data imbalance. Furthermore, the integration of numerical results with empirical measurements is employed to train the developed deep-neural networks, fitting high-dimensional mapping relationships among formation physical parameters, engineering parameters, and fracture morphology. Finally, the efficiency and the accuracy of the proposed method are verified by multi-level comparison experiments between real data and simulation results. Our research provides reliable technical support for rapid evaluation of formation fracturing potential in field and guidance of development process.

Keywords: Fracture propagation, Neural proxy model, Deep learning, Numerical simulation, Deep-formation