Elastic full waveform inversion based on neural network reparameterization and optimal transport theory

Abstract Elastic full waveform inversion (EFWI) is a cutting-edge, high-resolution seismic imaging method. However, it faces several challenges, including cycle skipping, due to its sensitivity to the initial model and low-frequency data, and instability caused by sparse acquisition systems. To address these issues, we propose a deep learning (DL) and optimal transport (OT) based EFWI framework (DLOT-EFWI). In this framework, an autoencoder is used to simultaneously reparameterize the P-wave (vp) velocity model, S-wave (vs) velocity model, and density (ρ) model. The model update is transformed into the update of trainable neural network parameters. Meanwhile, OT theory is employed to match the predicted and observed seismic data, which further reduces cycle skipping, alleviates the sensitivity of EFWI to the initial model and low-frequency data, and enhances the ability to characterize complex subsurface structures. Numerical experiments demonstrate that, even under challenging conditions such as the poor quality initial model, missing low-frequency data, and sparse acquisition, the proposed DLOT-EFWI method can provide high-accuracy and high-resolution results for vp, vs, and ρ compared with conventional EFWI methods.