Abstract The interaction between natural and hydraulic fractures significantly influences fracture propagation and connectivity, leading to complex fracture networks, yet measurement of these properties remains challenging. This study introduces a physics-informed, probabilistic inversion framework that combines forward geomechanical simulation and Bayesian deep learning to infer NF (nature fracture) properties from fiber-optic strain response data. Our methodology integrates three key components: (1) stochastic discrete fracture network (DFN) generation with parameterized natural fracture distributions, (2) coupled geomechanical-fluid flow simulations using a displacement discontinuity method (DDM) and finite volume fluid modeling, and (3) a deep learning framework for natural fracture property inversion from fiber optic response data. The computational efficiency of DDM, where displacements are only calculated at predefined discontinuities rather than throughout the entire domain, makes large-scale forward modeling feasible, enabling deep learning-based inversion. Sobol sequence sampling was applied to systematically vary orientation, density, and length, generating 2046 simulation cases that capture diverse fracture interaction scenarios. Forward modeling was performed to compute fiber optic responses, capturing variations in arrival time, location, and strain intensity at the monitoring well. The deep learning model was trained on these spatiotemporal strain datasets, leveraging data-driven insights to decode characteristic patterns and accurately infer key natural fracture attributes. The fiber optic response effectively captures the influence of natural fracture properties on fracture propagation. Variations in orientation, density, and length systematically affect arrival time, location, and strain intensity at the monitoring well, enabling the characterization of natural fracture properties based on fiber optic response. The Bayesian neural network achieves strong inversion performance, with coefficients of determination (R2) of 0.911 for orientation, 0.838 for density, and 0.710 for length. Error residuals approximately follow normal distributions, and uncertainty analysis shows that the predicted standard deviations remain below 8° and 6 m for most samples. These results demonstrate the model's robustness and uncertainty-awareness, while offering interpretable inversion outputs. This study, for the first time, applies Bayesian neural networks to invert natural fracture properties from fiber optic response data in hydraulic fracturing. By integrating forward-modeled fiber optic data with a geomechanical model, it quantifies the impact of natural fractures on strain signals. The proposed approach enhances natural fracture characterization, providing a novel method for improving hydraulic fracturing designs in complex systems.
Liu et al. (Mon,) studied this question.