(General session) Rock Engineering / Mining technologies

[3K0101-07-07] [Student presentation: Doctoral course] Full-waveform Inversion for the Processing of Seismic Data Based on Regularized Regression

○Jiahang Li¹, Hitoshi Mikada¹, Junichi Takekawa¹ (1. Kyoto University)

Chairperson : Hashiba Kimihiro (The University of Tokyo)

Keywords:Full-waveform Inversion, Frequency Domain, Source-receiver Wavefield, Regularized Regression

Conventional full-waveform inversion (FWI) obtains the gradient direction for the next update by calculating the residuals of the actual and simulated wavefields. In the present research, we have decided to add a new twist to the commonly used frequency-domain processing flow in the practice of FWI. We first extract the source-receiver wavefield slices from the multiscale frequency inversion strategy, take the wavefield slices out of the FWI for separate processing, and then bring the processed wavefield slices back into the FWI for the conventional FWI processing. The implementation of the separate processing will significantly improve the inversion accuracy of the FWI. Therefore, our algorithm concentrates on regularized regression processing of the source-receiver wavefield slices. Specifically, seismic data containing noise are often sparse, so we can extract the nonzero terms in the sparse matrix and perform regularized regression on their absolute value, then rearrange the normalized nonzero terms to a sparse matrix again. The image gradient of the matrix after being regularized will be reduced; that means the difference between adjacent values of the matrix is reduced, which reduces the proportion of noise components in the seismic data, and improves the generalization ability of the data matrix.
In the numerical test part, the performance of the experiments shows that first, our modified algorithm can get more accurate inversion results at the same frequency and with higher background noise. Second, because the influence of outliers in the frequency slices is optimized, the complexity of the model is reduced; thus, modified FWI can obtain a more precise model profile faster.