A dual-decoder U-Net network with ringed residual for full waveform inversion

Full-waveform inversion (FWI) is a critical data-fitting process in subsurface geology. While deep learning (DL) has achieved significant success in FWI, existing methods often struggle with domain alignment inconsistencies, inaccurate mapping from time-series seismic data to velocity models, and data incompleteness. To address these challenges, we propose DDRA-Net, a novel DL-based FWI model that integrates cyclic residual architectures with attention mechanisms. Specifically, the model utilizes a U-Net backbone with dual decoders to simultaneously estimate velocity values and delineate stratigraphic boundaries. A lightweight ringed residual unit is introduced to leverage residual propagation and feedback strategies, thereby enhancing feature extraction and mapping accuracy. Furthermore, the Convolutional Block Attention Module (CBAM) is incorporated between the encoder and decoders to refine features across channel and spatial dimensions. Finally, an efficient staged training strategy is employed to optimize data fitting. Extensive experiments on four synthetic OpenFWI datasets demonstrate that DDRA-Net quantitatively and qualitatively outperforms state-of-the-art DL-based FWI methods, offering superior precision in characterizing complex subsurface structures.