Surrogate-Based Hydraulic Fracture Propagation Prediction Using Deep Neural Network Proxy

ABSTRACT: Reliable prediction and uncertainty quantification of hydraulic fracture propagation behavior are crucial aspects of unconventional reservoir development, however, their implementations are largely restricted by the huge computational cost of the high-fidelity model simulations. We present a deep autoencoder neural network proxy workflow that allows us to rapidly visualize subsurface fracture morphology (e.g., fracture width, fracture height and fracture length) given pumping schedule and rock properties (e.g., Poisson ratio, Young module, leakage coefficient, etc). Nevertheless, the training process becomes inefficient due to the large output dimensionality. To overcome this issue, an autoencoder network is initially employed to uncover lower-dimensional features that effectively represent high-dimensional fracture morphology images. In our methodology, the regression model is trained to forecast lower-dimensional features instead of directly constructing original fracture image. During the prediction phase, the trained regression model initially predicts the latent variables based on wellhead pumping data, formation characteristics, and rock geomechanic properties. Subsequently, these latent variables are utilized as inputs for the trained decoder to generate fracture morphology images. In addition, a fast self-developed full-physics fracturing simulator is used to efficiently simulate a large set of training samples related to the fracture morphology. Some preliminary results based on a simple illustrative example, i.e., single-fracture propagation, are used to demonstrate the accuracy and efficiency proposed approach. The proposed deep-learning-based visualizations model should be a fast emulator for quantitative characterization and monitoring of hydraulic fracturing stimulation. Further research is required to verify its applicability to field-scale reservoir models with complex fracture network. 1 INTRODUCTION Horizontal wells hydraulic fracturing is a key technology for unconventional oil and gas development (Daneshy et al, 2011; Barati et al, 2014). In order to improve the degree of reserve utilization, North American oil companies are continuously exploring volume transformation technologies to reduce well spacing and fracture spacing. Onsite fiber-optic temperature monitoring and acoustic monitoring indicate that in the case of small spacing renovation, there are issues such as uneven liquid injection in multiple clusters of cracks, and the proportion of liquid injection in each cluster is closely related to factors such as perforation parameters and stress distribution in each cluster (Wheaton et al, 2016). To improve the balanced expansion of multi cluster fractures and optimize the design of multi cluster fracturing schemes, it is urgently necessary to establish an efficient simulation method for multi fracture propagation coupled with wellbore perforation fracture (Chen et al, 2022, 2022).