Faster determination of an initial velocity model for full-waveform inversion based on simulated annealing

Abstract

Full-waveform inversion (FWI) is a prominent method commonly used to create detailed velocity models of the subsurface. However, because it relies on gradient methods, it suffers from the limitation of getting trapped in local minima. To avoid this problem, FWI needs to start from an initial velocity model that lies in the same region of convexity as the global minimum. Global optimization methods, such as simulated annealing, can be used to find such an initial velocity model. However, the iterative process of simulated annealing entails high computational cost, first because of the large number of iterations needed to explore the variables space, and second because of the large number of simulations needed to adjust the simulation parameters. Although the first issue is a common concern of previous work, the second issue is usually neglected, relying on trial and error. We introduce an approach to tackle the issues and then apply it within a simulated annealing framework we create. Our experiments using the Marmousi data yielded promising results when compared with previous work. We find that our approach almost eliminated the computational effort to fine-tune several simulation parameters. In addition, the number of iterations needed to explore the variables space is reduced by two orders of magnitude. FWI is able to find a detailed velocity model with high quality when using the initial velocity model generated by the method.