Low-Frequency Extrapolation by Deep-Learning for Cross-Well Full-Waveform Inversion – Case Study From the Aquistore CO 2 $\rm CO_{2}$ Storage Site

Abstract

Full-waveform inversion (FWI) of seismic data is a powerful method for estimating high-resolution models of the subsurface. An accurate initial model and low-frequency data are necessary to avoid cycle skipping and perform a successful FWI. In the absence of this information, FWI is likely to fail due to convergence in local misfit minima. With the recent advancements in artificial intelligence, studies have shown that absent low-frequency data can be extrapolated using deep learning (DL). These studies have been mostly focused on surface seismic data whose frequency content is different from cross-well data. In this study, we assess the use of DL for low-frequency extrapolation for a cross-well survey that was done at the Aquistore ${\mathrm{CO}}_2$ storage site in Saskatchewan. This assessment includes both numerical and field data examples. We extrapolate the low frequencies to increase the bandwidth of the acquired data at the Aquistore site and perform FWI. We evaluate the efficiency of this method by comparing the results with obtained velocity models from the conventional multiscale FWI. Our results for the Aquistore data show that the proposed strategy leads to an accuracy improvement of 39% and 20% in the model and data domains, respectively.