Reservoir multiparameter prediction method based on deep learning for CO₂ geologic storage

Time-lapse seismic difference refers to the comprehensive response of fluid saturation, pore pressure, and porosity. However, the contribution of different parameters to the seismic response is difficult to distinguish. The high-precision prediction of these reservoir parameters is of great significance in CO₂ geologic storage and oil and gas development. Therefore, a simultaneous time-lapse reservoir multiparameter prediction method based on a multitask learning network is proposed. Combined with CO₂ geologic storage monitoring, the process of generating training data is described, involving numerical simulation, petrophysical models, and seismic forward modeling. Moreover, the Hertz-Mindlin formula, which considers pressure changes, is used to establish the relationship between formation elasticity and physical parameters in CO₂ storage. The effects of fluid saturation, pressure, and porosity on P- and S-wave velocities and densities are analyzed, and the amplitude-variation-with-offset response characteristics of fluid saturation, pressure, and porosity changes are discussed. In total, 4700 and 300 sets of reservoir parameters and seismic angle gather data are used for network training and testing, respectively. The prediction results of synthetic and field data find that the time-lapse reservoir multiparameter prediction method based on multitask learning can effectively distinguish changes in each parameter and simultaneously obtain high-precision prediction results of fluid saturation, pressure, and porosity. Once the network is constructed, the prediction will take only a few seconds, which will promote the further development of the CO₂ geologic storage theory and technology.