Application of deep learning-based velocity modeling in aquifer prediction

In this study, an improved U-Net neural network architecture is employed to achieve an efficient mapping from pre-stack seismic records to velocity models through an encoder-decoder structure. The encoder extracts multi-level features using convolution and pooling layers, while the decoder reconstructs the velocity model through deconvolution and upsampling layers. The improved U-Net architecture is applied to velocity model prediction in two geological scenarios: one based on a faulted background and the other featuring aquifers with random geometric shapes. Seismic records used for training data are generated through finite-difference acoustic wave numerical simulations. The training results demonstrate that the deep learning model can accurately predict the location, thickness, and geometry of aquifers, maintaining high accuracy even under complex geological conditions.