Re-establishment of gradient in frequency-domain elastic waveform inversion

To obtain solutions close to global minimum in waveform inversion, the gradients computed at each frequency need to be weighted to appropriately describe the residuals between modeled and field data. While the low-frequency components of the gradients should be weighted to recover the long-wavelength structures, the high-frequency components of the gradients need to be weighted when the short-wavelength structures are restored. However, the conventional elastic waveform inversion algorithms cannot properly weight the gradients computed at each frequency. When gradients are scaled using the pseudo-Hessian matrix inside the frequency loop, gradients obtained at high frequencies are over-emphasized. When the gradients are scaled outside the frequency loop, gradients are weighted by the source spectra. In this study, we propose applying weighting factors to the gradients obtained at each frequency so that gradients can properly reflect the differences between the true and assumed models satisfying the general inverse theory. The weighting factors are composed by the backpropagated residuals. Numerical examples for the simple rectangular-shaped model and the modified version of the Marmousi-2 model show that the weighting method enhances gradient images and inversion results compared to the conventional inversion algorithms.