S-wave velocity inversion and prediction using a deep hybrid neural network

The S-wave velocity is a critical petrophysical parameter in reservoir description, prestack seismic inversion, and geomechanical analysis. However, obtaining the S-wave velocity from field measurements is difficult. When no measured S-wave data are available, petrophysical modelling provides the most accurate S-wave velocity prediction. However, because of the complexity of underground geological structures and diversity of rock minerals, the prediction results of petrophysical modelling are easily affected by factors such as the cognition and experience of the modeller. Therefore, the development of novel robust and simple S-wave velocity inversion and prediction methods independent of the modeller is critical. Inspired by ensemble learning and based on the geologic sedimentation law of reservoirs and their characteristics in logging response, an S-wave velocity inversion and prediction method based on deep hybrid neural network was developed by combining the classical convolution neural network (CNN) with the long short-term memory (LSTM) network. Considering the conventional logging data such as acoustic and density as the input in the proposed method, the CNN was used to establish the nonlinear mapping relationship between the input data and S-wave velocity, and the LSTM network was used to integrate the vertical variation trend of the stratum. Thus, intelligent data-driven inversion and prediction of the S-wave velocity were realised. The experimental results revealed that the proposed method exhibited a strong generalisation ability and could obtain prediction results comparable to those of petrophysical modelling with a single-well data set for training. Thus, a novel methodology for robust and convenient S-wave velocity prediction was devised. The proposed method has considerable academic and application implications.