Multiscale neural decoding and weighting for reparameterized seismic full-waveform inversion

ABSTRACT Seismic full-waveform inversion (FWI) stands as a cornerstone technique in subsurface imaging, offering unparalleled insights into subsurface structures. However, conventional FWI methods often face challenges related to nonlinearity and ill-posedness, particularly in handling inaccurate initial models and uncertainty of observations. To address these issues, a novel approach termed multiscale neural decoding and weighting for seismic FWI (MNDW-FWI) was proposed. Our method leverages a multiscale decoding neural network to effectively reparameterize velocity models and capture diverse scales of subsurface features. Furthermore, a flexible weighting mechanism was introduced to assign distinct weights to individual branches of the decoder networks, thereby enabling tailored emphasis on different scales of the velocity models. The integration of these advancements into a recurrent neural network-based FWI (RNN-FWI) framework yields significant improvements in accuracy and robustness. By incorporating the MNDW strategy and automatic differentiation (AD) technique, our approach offers superior performance in capturing intricate subsurface structures from an inaccurate initial model and mitigating the effects of the presence of noise and the absence of low-frequency components in seismic records. Through comprehensive experimental evaluations on synthetic examples with noisy and limited bandwidth observations, the efficacy of our proposed method in enhancing the accuracy and flexibility of inversion was demonstrated.