Uncertainty quantification-driven shear wave velocity prediction using a spatiotemporal Bayesian neural network

Abstract Shear wave velocity (Vs) is a key parameter used for seismic data processing and evaluating oil and gas reservoirs. Conventional logging and machine learning methods are hampered by the provision of fixed Vs predictions, which limits their capacity to effectively quantify the uncertainty of predictions. Therefore, we propose a spatiotemporal Bayesian neural network (STBNN), which exploits the advantages of ensemble learning methods in processing logging data and Bayesian methods in modeling uncertainty. STBNN facilitates both accurate Vs predictions and a robust quantification of uncertainty. We initially perform a feature selection on logging data to identify logging-sensitive attributes as input features, and employ Monte Carlo dropout, a Bayesian approximation method, to determine both epistemic and aleatoric uncertainties. These uncertainties are integrated into the STBNN framework, which reduces prediction uncertainty and enhances accuracy. Subsequently, we assess the model using a logging dataset derived from actual oil reservoirs, and evaluate its generalizability. The trained STBNN model is applied to an untrained exploration area to predict Vs and its associated uncertainties. Finally, we assess the influence of sample size on network epistemic uncertainty and devise a practical mitigation strategy. The results reveal that compared with spatiotemporal neural network and empirical models, the STBNN model achieves a higher accuracy of prediction. We establish that prediction uncertainty is primarily attributable to epistemic uncertainty, with higher uncertainty observed in areas with limited data. This prediction uncertainty can be effectively reduced by increasing the diversity of sample data. The proposed method delivers excellent Vs prediction results, whilst minimizing the risk of errors in the interpretation of logging.