MAU-net: A multibranch attention U-net for full-waveform inversion

Abstract

Data-driven velocity inversion has emerged as a prominent and challenging problem in seismic exploration. The complexity of the inversion problem and the limited data set make it difficult to ensure the stability and generalization of neural networks. To address these concerns, a novel approach called multibranch attention U-net (MAU-net) is proposed for velocity inversion. The key distinction of MAU-net from previous data-driven approaches lies in its ability to not only learn information from the data domain but also incorporate prior model domain information. MAU-net consists of two branches: one branch uses seismic records as input to effectively learn the mapping relationship between the data and model domains, whereas the other branch uses a prior geologic model as input to extract features from the model domain, thereby guiding MAU-net’s learning process. In addition, three major improvements are made to the model to enhance MAU-net’s utilization of seismic data and handle redundant information. The effectiveness of the improvements is verified through ablation experiments. The performance of MAU-net is demonstrated with the Marmousi model and the 2004 BP model, and it also can be combined with full-waveform inversion to further improve the quality of the inversion result. MAU-net exhibits robust performance on field data through the use of transfer learning techniques, further confirming its reliability and applicability.