Matched field inversion of broadband data using the freeze bath method

This paper presents a statistical freeze bath method for geoacoustic inversion that emphasizes the search for a distribution of models that fit the data well. Contrary to simulated annealing optimization, the freeze bath samples the multidimensional model parameter space at constant freeze probability, corresponding to fixed temperatures for each parameter. The sampling process uses a heat bath algorithm as a Boltzmann sampling tool to carry out a global search over the model parameter space. The conventional heat bath algorithm is modified to sample on a fuzzy grid in order to access the entire range of parameter values. The inversion provides a set of good models that indicates how well the model parameters are constrained by the data, and reveals the degree of correlation between parameters. The efficiency of the search process is improved by reparameterizing the original model parameters to a new set based on the eigenvectors of the model covariance matrix. The inversion performance of the freeze bath is demonstrated using simulated data for a simple geoacoustic model. The method is applied to shallow water broadband data obtained during the Haro Strait geoacoustic tomography experiment to estimate a geoacoustic profile at the site.