Differentiable Dynamic Time Warping Divergences In Full Waveform Inversion Expanded

Summary

Full-waveform inversion (FWI) aims to obtain accurate subsurface models by minimizing the discrepancy between observed seismic data and modeled data. However, the commonly used L2-norm waveform difference misfit functional is prone to cycle-skipping due to local minima. To address this issue, we propose more effective misfit metrics for FWI by utilizing the soft-dynamic time warping (SDTW) divergence distance and its sharp variant. The proposed methods introduce a hyper-parameter to ensure differentiability of the functional, enabling the use of the adjoint state method for gradient computation in FWI. Unlike conventional SDTW, our divergence-based metrics always yield positive values, reaching their minimum when the modeled trace matches the observed trace. The efficacy of the proposed methods is demonstrated through their application on a field dataset, highlighting their robustness in mitigating cycle-skipping compared to the conventional L2 norm.