学术文献库

Nonlinear least squares data-fitting driven by physical process simulation is a classic and widely successful technique for the solution of inverse problems in science and engineering. Known as "Full Waveform Inversion" in application to seismology, it can extract detailed maps of earth structure from near-surface seismic observations, but also suffers from a defect not always encountered in other applications: the least squares error function at the heart of this method tends to develop a high degree of nonconvexity, so that local optimization methods (the only numerical methods feasible for field-scale problems) may fail to produce geophysically useful final estimates of earth structure, unless provided with initial estimates of a quality not always available. A number of alternative optimization principles have been advanced that promise some degree of release from the multimodality of Full Waveform Inversion, amongst them Wavefield Reconstruction Inversion, the focus of this paper. Applied to a simple 1D acoustic transmission problem, both Full Waveform and Wavefield Reconstruction Inversion methods reduce to minimization of explicitly computable functions, in an asymptotic sense. The analysis presented here shows explicitly how multiple local minima arise in Full Waveform Inversion, and that Wavefield Reconstruction Inversion can be vulnerable to the same "cycle-skipping" failure mode.