Shear wave velocity prediction based on bayesian-optimized multi-head attention mechanism and CNN-BiLSTM

Abstract

Shear wave velocity (VS) is one of the fundamental geophysical parameters essential for pre-stack seismic inversion, rock mechanics evaluation, and in-situ stress assessment. However, due to the high cost of acquiring VS log data, it is impossible to carry out this logging project in all wells. Thus, it is extremely necessary to develop an efficient and reliable VS prediction method. Deep learning methods have distinct advantages in data inversion, but different neural network have their own characteristics. A single-structured neural network has inevitable limitations in VS prediction, making it challenging to effectively capture the nonlinear mapping relationships of multiple parameters. Therefore, an integrated VS prediction model was proposed based on analyzing the applicability of classical neural networks. This new model, denoted as Bo-MA-CNN-BiLSTM, combines a Bayesian-optimized and multi-head attention mechanism (Bo-MA) with a convolutional neural network (CNN) and a bidirectional long short-term memory network (BiLSTM). It can effectively capture spatio-temporal data reflecting geophysical characteristics from log data, and the integration of the multi-head attention mechanism enhances the rational allocation of weights for log data. Bayesian optimization is utilized to determine the values of hyperparameters, overcoming the subjectivity and empiricism associated with manual selection. Actual data processing demonstrates that the new model achieves higher accuracy in predicting VS than applying CNN, LSTM, BiLSTM, and CNN-LSTM individually. The application results of well log data not involved in training indicate that, compared to other classical models, this new model exhibits optimal evaluation metrics. Especially for strongly heterogeneous formations, the predicted results demonstrate significant superiority, verifying the generalization ability and robustness of the proposed model.