Deep Learning for Low-Frequency Extrapolation of Multicomponent Data in Elastic FWI

Full waveform inversion (FWI) strongly depends on an accurate starting model to succeed. This is particularly true in the elastic regime: The cycle-skipping phenomenon more severe FWI compared acoustic FWI, due short S-wave wavelength. In this paper, we extend our work extrapolated (EFWI) by proposing synthesize low frequencies of multi-component seismic records, and use those "artificial" seed frequency sweep FWI. Our solution involves deep learning: separately train same convolutional neural network (CNN) two training datasets, one with vertical components horizontal particle velocities, extrapolate data. architecture CNN designed a large receptive field, either kernels or dilated convolution. Numerical examples Marmousi2 show that 2-4Hz data from band-limited above 4Hz provide good models for P-wave velocities. Additionally, study generalization ability proposed over different physical models. For test data, collecting dataset simulation shows better extrapolation accuracy than simulation, i.e., smaller gap.