A Feature-Boosted Convolutional Neural Network for Full-Waveform Inversion

Convolutional neural network-domain full-waveform inversion (CNNFWI) is a powerful technique for reconstructing high-resolution subsurface parameters by iteratively updating network parameters. This approach eliminates the need for a large training dataset as required by data-driven inversion methods and avoids artificially deriving the gradient of the objective function with respect to the inversion parameters in the conventional adjoint-state inversion method. However, CNNFWI remains an ill-posed inverse problem, with the potential of the optimization process converging to local minima. To produce a favorable reconstruction, we develop a feature-boosted CNN for inversion. This work focuses on two key innovations. First, instead of adding a regularization term to the loss function, our approach employs differential operators on multichannel feature maps to enhance the boundary features of geological structures. Second, we incorporate coordinate attention to improve feature representation by considering both spatial location dependencies and channel correlations. The combination of these two modifications enhances the inversion performance, particularly in accurately depicting the shape and velocity of structures. Moreover, we investigate the upsampling (Us) operation within the network architecture responsible for transforming low-resolution feature maps into high-resolution outputs, identifying the optimal Us method suited for CNNFWI. To examine the impact of neural network architecture components on inversion, synthetic experiments are conducted with both an anomaly model and the overthrust model. We performed a comprehensive comparison with previous CNNFWI and conventional FWI methods. The numerical results conclusively demonstrate the effectiveness of our approach in achieving superior subsurface velocity models. Finally, we synthesize time-lapse seismic data with the well-known Kimberlina model to confirm the potential of our framework in providing a high-quality description of reservoir changes due to carbon injection.