Bayesian Optimization for Full Waveform Inversion

Full Waveform Inversion (FWI) is a computational method to estimate the physical features of Earth subsurface from seismic data, leading to the minimization of a misfit function between the observed data and the predicted ones, computed by solving the wave equation numerically. This function is usually multimodal, and any gradient-based method would likely get trapped in a local minimum, without a suitable starting point in the basin of attraction of the global minimum. The starting point of the gradient procedure can be provided by an exploratory stage performed by an algorithm incorporating random elements. In this paper, we show that Bayesian Optimization (BO) can offer an effective way to structure this exploration phase. The computational results on a 2D acoustic FWI benchmark problem show that BO can provide a starting point in the parameter space from which the gradient-based method converges to the global optimum.