Abstract:

Deep learning-based methods for multiple attenuation have been a pivotal research focus in the field of seismic data processing. Traditional supervised neural networks are data-driven and rely on direct end-to-end mapping. The physical interpretability of the network model outputs is limited, and the effectiveness of multiple attenuation is constrained by the quality of labeled seismic data. This paper proposes an improved deep learning method that integrates the time-spatial physical information of 3D seismic data volume as prior knowledge with the neural network output to construct an implicit polynomial of multiple and full-wavefield. In this approach, the output of the neural network is not labeled seismic data but coefficients of polynomial function space. By incorporating prior knowledge into the loss function and minimizing this loss function, an implicit polynomial of multiple and full-wavefield is derived. This approach obviates the matching subtraction process in surface-related multiple attenuation. The results of synthetic and field data demonstrate that the proposed method surpasses traditional end-to-end deep learning methods in terms of efficacy and accuracy in free surface-related multiple attenuation.

Key words: deep learning')">

deep learning, prior knowledge, neural network, multiple attenuation