Abstract: | We present a source-manipulation formulation formitigating the long-standing cycle-skipping issue in the seismicwaveform inversion. We first extend the set of variables in theoptimization problem to include the source wavelet. It turns theminimization problem into a minimax problem. However, thisapproach is not attractive for practical use owing to its unfeasiblecomputation and data collection cost. We propose a method toapproximate the minimax problem with a constrained optimization,which leads to a so-called source manipulation approach. It enhancesthe convexity of the objective function while improving the gradient'squality in the iterative reconstruction. Neural networks are employedto provide an efficient solution to the induced sub-problem. |