Direct inversion of 3D seismic reservoir parameters based on dual learning networks

Tight sandstone has become an important area in gas exploration. In this study, we propose a 3D seismic reservoir parameter inversion method for tight gas-bearing sandstone reservoirs using dual neural networks. The first network referred to as the inversion network, receives seismic data and predicts reservoir parameters. At well locations, these predictions will be validated based on actual reservoir parameters to evaluate errors. For non-well locations, synthetic seismic data are generated by the application of rock physics forward modeling and seismic reflection coefficient equations. The errors are then calculated by comparing synthetic seismic data with actual seismic data. During the rock physics forward modeling, pseudo reservoir parameters are derived by perturbing the actual reservoir parameters, which are then used to generate pseudo elastic parameters through the modeling. Both the actual and pseudo parameters are then used to train the second network, referred to as the rock physics network. By incorporating the rock physics network, the method effectively alleviates issues such as gradient explosion that may arise from directly integrating rock physics computations into the network, while the inclusion of pseudo parameters enhances the network’s generalization capability. The proposed method enables the direct inversion of porosity, clay content, and water saturation from pre-stack seismic data using deep learning, thereby achieving quantitative predictions of reservoir rock physical parameters. The application to the field data from tight sandstone gas reservoirs in southwestern China demonstrates the method has the good capability of indicating the gas-bearing areas and provide high resolution.