A data-driven approach to predict 3D dynamic fracture propagation during hydraulic fracturing in geo-energy exploitation

Hydraulic fracturing is a crucial technology to boost the productivity of various types of geo-energy, including geothermal, hydrocarbon and gas hydrate. Accurate and efficient prediction of hydraulic fracturing fracture propagation is of utmost importance for the development of geo-energy. However, traditional simulation methods are time-consuming and computationally demanding, while existing neural networks are ill-suited for complex, multi-factor hydraulic fracturing scenarios. To address these challenges, we propose the Att-FC-U-Net network model, which integrates Fast Fourier Transform (FFT) and parallel convolutional (PC) structures into the U-Net framework. By considering spatiotemporal variations, Att-FC-U-Net enables accurate prediction of 3D fracture propagation in different stages of horizontal well fracturing. Additionally, transfer learning is employed wto expedite convergence. The model is trained on a dataset comprising 19,200 instances of fracture propagation images, encompassing crucial data such as wellbore perforation, natural fracture, pumping schedules and reservoir properties. Remarkable results are achieved, with a cross-entropy loss (CE) below 0.0001 and an F1 score exceeding 0.93. These findings underscore the enhanced prediction efficiency of our method and provide valuable insights for optimizing fracturing design and predicting fracture morphology, making it a valuable contribution to the field.