Robust full-waveform inversion based on automatic differentiation and differentiable dynamic time warping

Abstract Full waveform inversion is a methodology that determines high-resolution parameters. The widely used L2-norm misfit function has local minima if the low wavenumber components are not accurate. Suffering from cycle skipping problem, solution of will be trapped in minima. Dynamic time warping aims to find an optimal alignment between two signals, which more robust measure avoid cycle-skipping challenges. However, discontinuity makes conventional dynamic distance unsuitable for inversion. We introduce soft as function, differentiable inverted can converge accurate global minimum. compare convexity and show wider range with different shift amplitudes. It alleviate half-wavelength limitation L2-norm. calculate gradient using automatic differentiation technique minibatch strategy then analyse paths smooth A significant parameter γ Soft-DTW tending L2-norm, generates new recommend small ensure distance. Numerical examples effectively reconstruct deep velocity parameters BG Compass Marmousi models noise robustness lower dependence on initial model.