fcei-yy

Building an accurate long-wavelength velocity model of subsurface medium is important in many fields of earth science, such as exploration geophysics, seismology, etc. Frequency controllable envelope inversion (FCEI), which is a kind of full waveform inversion (FWI) method, is able to build such a long-wavelength velocity model. However, when starting from a very coarse model, FCEI may also fails to converge due to the cycle-skipping problem. To achieve robust velocity model building and reduce the dependence on initial model, we propose the Fourier-reparameterized implicit FCEI (FR-IFCEI) method, which has the following two key features: (1) The implicit neural representation (INR) is used to shift the model space of FCEI from velocity to neural network parameters, which expands the convergence domain and significantly reduces the dependence of FCEI on initial velocity model; (2) The Fourier-reparameterization technique is adopted to enforce INR to output long-wavelength velocity rather than a high-resolution one, reducing the complexity of inversion and significantly accelerating convergence. To assess the performance of FR-IFCEI, we conduct synthetic experiments based on two benchmark models, which are the Marmousi model and the 2004 BP velocity model. The results clearly demonstrate that FR-IFCEI can build the long-wavelength velocity model even starting from a random initial model. Moreover, it significantly outperforms both FCEI and implicit FWI (IFWI) in terms of inversion efficiency and accuracy. This dataset includes all data from the synthetic experiments. Refer to the file list for details.