Probabilistic Full-Waveform Inversion of Surface Waves: a Field Data Application

Summary Conventional techniques for inverting surface waves, such as dispersion curve inversion, rely on the 1D assumption of the subsurface structures. Conversely, full-waveform inversion exhibits potential in reconstructing high-resolution subsurface models, even for strong lateral and vertical variations. Nevertheless, applying full-waveform inversion in near-surface seismic scenarios poses significant challenges due to the high non-linearity of the optimization problem and the associated computational burden. Traditional methods employ gradient-based optimization algorithms to minimize an error function, which usually measures the misfit between observed and simulated seismic data. This deterministic approach only yields a "best-fitting" model but fails to incorporate uncertainties in the solution. Framing this inverse problem within a probabilistic framework faces the huge computational demands of the Bayesian approach when dealing with high-dimensional problems. Our proposed solution introduces a gradient-based Markov chain Monte Carlo full-waveform inversion, where the sampling of the posterior distribution is expedited by defining a proposal distribution that is a local approximation of the target posterior and by compressing model and data spaces through the discrete cosine transform. We have tested our algorithm using a real dataset, acquired in Grenoble, showing our obtained results in terms of data matching, uncertainty estimation and agreement with the available borehole data.