Deep Learning Inversion of Electromagnetic Detection Data for Macroscopic Fractures in Croswell

The correct identification of reservoir fractures is of great practical significance for accurately evaluating the oil and gas reserves of reservoirs and the effectiveness of hydraulic fracturing, especially for macroscopic fractures between wells with an aperture greater than 1cm, which are often the culprit of hydraulic fracturing failure. However, traditional reservoir fracture identification methods face difficulties in feature extraction and illposedness in inversion problems, making it difficult to ensure the accuracy of results. Regression prediction models based on convolutional neural networks have powerful nonlinear data mapping capabilities and can replace traditional geophysical inversion calculations. To address the issue that traditional convolutional neural networks can only handle scalar data while the actual electromagnetic field data is vector valued, this paper proposes a macroscopic fracture identification method for inter-well reservoirs based on complex-valued convolutional neural networks. By using the real and imaginary parts or amplitude and phase data of the observed field as the input to the complex-valued convolutional neural network, the information input to the network is increased, enabling the network to extract more target features and improve the identification ability of reservoir fractures. Through comparative experi-ments based on amplitude scalar data and complex-valued convolutional neural networks, the results demonstrate that the electromagnetic detection data inversion for fracture identification based on complex-valued convolutional neural networks has higher resolution and provides a new approach for the accurate identification of macroscopic fractures in inter-well reservoirs.