Deep Learning-based Self-supervised Multi-parameter Inversion

Summary The quantitative interpretation of geological structures relies on multi-parameter models (MPMs) inversion. However, conventional full waveform inversion that matches simulated seismic data to observed seismic data cannot accurately obtain high-resolution MPMs because of the implicit inter-parameter coupling relations in the multi-parameter wave equation. Additionally, conventional supervised deep learning approaches that require a significant number of annotated labels cannot predict precise MPMs, as only a limited number of sophisticated synthetic MPMs are available as labels. To address this issue, we propose a self-supervised multi-parameter inversion (SS-MPI) to provide high-resolution MPMs from the prior first-arrival-based tomography and reflection-based migration image. SS-MPI creates representative MPMs from the prior information as pseudo-labels to pre-train the deep learning algorithm, which then predicts MPMs as feedback to update these training pseudo-labels iteratively. Synthetic examples of elastic and anisotropic models indicate that SS-MPI outperforms the conventional elastic full waveform inversion (EFWI) and delivers highly accurate and high-resolution MPMs.