Mitigating elastic effects of acoustic full waveform inversion for VSP data via deep learning

He Liu, Luping Qu, Daniel O. Trad, Kristopher A. Innanen

Full waveform inversion (FWI) is a valuable technique for estimating high-resolution subsurface physical property models. It has become a potent tool for time-lapse seismic inversion. Due to the high resolution feature of underground structures, VSP data have to used to monitor changes in reservoir profiles during activities like injection and production, as well as for long-term CO₂ storage. Nevertheless, FWI faces a significant computational burden, elastic FWI takes much more efforts than acoustic FWI, and time-lapse FWI typically involves at least two FWI computations. Source-encoding strategies can be employed to accelerate the inversion, however, they will eventually introduce crosstalk noise in the inversion results and this phenomena is more obvious in elastic cases.Acoustic full waveform inversion is usually the first choice for velocity model building due to its efficiency and robustness. However the recorded field data always contain elastic effects due to such as PS and SP-wave conversions, even in marine acquisitions. In this work, we adopt a deep learning approach to mitigate the elastic effects in VSP data. We train convolutional network to map elastic shot gathers into their acoustic counterparts, and perform acoustic FWI using the pseudo-acoustic shot gather. Our experiments show that the transformed acoustic data can match well with the direct simulated acoustic data. And the inversion results also show improvement compared with the inversion result by acoustic FWI using elastic data.