Shear wave velocity prediction using bidirectional recurrent gated graph convolutional network with total information embeddings

Abstract

Shear wave velocity is an essential elastic rock parameter for reservoir characterization, fluid identification, and rock physics model building. However, S-wave velocity logging data are often missing due to economic reason. Machine learning approaches have been successfully adopted to overcome this limitation. However, they have shortcomings in extracting meaningful spatial and temporal relationships. We propose a supervised data-driven method to predict S-wave velocity using a graph convolutional network with a bidirectional gated recurrent unit (GCN-BiGRU). This method adopts the total information coefficient to capture non-linear dependencies among well-log data and uses graph embeddings to extract spatial dependencies. Additionally, the method employs a bidirectional gated mechanism to map depth relationships in both upward and backward directions. Furthermore, the prediction performance is increased by an unsupervised graph neural network to handle outliers and the generation of additional features by the complete ensemble empirical mode decomposition with additive noise method. Finally, the GCN-BiGRU network is compared with Castagna’s empirical velocity formula, support vector regression, long-short-term memory (LSTM), GRU, and BiGRU methods over the North Sea open dataset. The results show that the proposed method performs better predicting S-wave velocity than the other ML and empirical methods.