VIF-Net: Interface completion in full waveform inversion using fusion networks

Deep learning full waveform inversion (DL-FWI) distinguishes itself from traditional physics-based methods for its robust nonlinear fitting, rapid prediction, and reduced reliance on initial velocity models. However, existing end-to-end deep learning approaches often neglect the reconstruction of layer interfaces and faults. In this article, we propose a two-stage DL-FWI approach named Velocity Interface Fusion (VIF). The first stage comprises two subnetworks: VIF-Velocity (VIF-V) generates the intermediate velocity model, and VIF-Interface (VIF-I) predicts velocity model interfaces. They have the same UNet++ architecture and an optional Fourier transform-based preprocessing module. Their main difference lies in the binary class-balanced cross-entropy loss tailored for VIF-I. The second stage is fulfilled by a fusion subnetwork with a limited downsampling encoder-decoder structure. This network refines the intermediate velocity model using the predicted interfaces to reconstruct the final model. A dynamic learning strategy combining warm-up and cosine annealing is employed to train all three subnetworks jointly. Our method is evaluated on two SEG salt and four OpenFWI datasets using four metrics in comparison with three popular DL-FWI methods. Results demonstrate its superior performance in interface completion and reconstruction. The source code is available at https://github.com/ FanSmale/VIF-dev.