Automatic prediction of shear wave velocity using convolutional neural networks for different reservoirs in Ordos Basin

Abstract Shear wave velocity (S-wave velocity) has great significance for reservoir characterization and can effectively reduce the ambiguity in seismic interpretation. However, owning to its high cost and technical difficulties, it is usually difficult to obtain for the whole region. Therefore, the prediction of S-wave velocity currently has turned into an urgent task. Thus, the S-wave velocity was predicted using 1D convolutional neural network (1D-CNN) with the well logs. In the first step, the traditional Xu-White and Xu-Payne model were used to carry out the inversion of bulk modulus, shear modulus and density, and illustrated the phenomenon for the traditional rock-physics model. Then, 1D-CNN, including one input layer, four convolutional layers, three fully connected layers and one output layer, was proposed, in which three types of well logs (conventional well logs, array induction well logs and gamma ray spectral well logs) sensitive to the S-wave velocity were determined through the well logs analysis. Finally, based on the established 1D-CNN model, the influence of different well logs combination on the S-wave velocity prediction was analyzed, indicating that the addition of different well logs can improve the accuracy of S-wave velocity. And different machine learning methods (MLs), including back-propagation (BP) and support vector regression (SVR), were also compared, showing that the predicted accuracy of 1D-CNN has been improved by 4.2% on average. Two cases for sandstone and carbonate reservoirs also showed that 1D-CNN proposed can achieve higher accuracy and better performance than traditional MLs and rock-physics models. Furthermore, the proposed method can be applied to the other oil and gas exploration fields, improving the exploration accuracy and increasing the production of oil and gas.