Seismic Waveform Inversion with Source Manipulation

Abstract

Summary In recent decades, Full-waveform inversion (FWI) has suffered from the cycle-skipping issue, which we found can be mitigated by changing the source signature of the observed data. Compared with a physical source such as the Ricker source, seismic data with the Gaussian source can provide a better landscape of the objective function while improving the gradient's quality in the iterative reconstruction. In the synthetic experiments, we transform band-limited seismic data simulated with the Ricker wavelet into seismic data with the Gaussian source and apply it to FWI. Neural networks are employed to provide an efficient solution to this problem. Numerical experiments on the Marmousi model are conducted to demonstrate the effectiveness of our proposed method.