Seismic Porosity Prediction in Tight Carbonate Reservoirs Based on a Spatiotemporal Neural Network

Porosity prediction from seismic data is of significance in reservoir property assessment, reservoir architecture delineation, and reservoir model building. However, it is still challenging to use traditional model-driven methodology to characterize carbonate reservoirs because of the highly nonlinear mapping relationship between porosity and elastic properties. To address this issue, this study proposes an advanced spatiotemporal deep learning neural network for porosity prediction, which uses the convolutional neural network (CNN) structure to extract spatial characteristics and the bidirectional gated recurrent unit (BiGRU) network to gather temporal characteristics, guaranteeing that the model accurately captures the spatiotemporal features of well logs and seismic data. This method involves selecting sensitive elastic parameters as inputs, standardizing multiple sample sets, training the spatiotemporal network using logging data, and applying the trained model to seismic elastic attributes. In blind well tests, the CNN–BiGRU model achieves a 54% reduction in the root mean square error and a 6% correlation coefficient improvement, outperforming the baseline models and traditional nonlinear fitting (NLF). The application of the proposed method to seismic data indicates that the model yields a reasonable porosity distribution for tight carbonate reservoirs, proving the strong generalization ability of the proposed model. This method compensates for the limitations of individual deep learning models by simultaneously capturing the spatial and temporal components of data and improving the estimation accuracy, showing considerable promise for accurate reservoir parameter estimation.