Solving Full Waveform Inversion Enhanced by Efficient Progressive Transfer Learning

The full waveform inversion (FWI) is highly non-linear and underdetermined. Due to the limitation of acquisition devices, low-frequency information may be missing from obtained data. The absence of low-frequency information could severely compromise FWI results. To solve these problems, we propose a learning-based approach to extrapolate the low-frequency data using a progressive transfer learning workflow to eliminate the labeling requirement for training. The system is implemented on a differentiable program platform to take advantage of automatic differentiation. The performance of the system is validated by an electromagnetic (EM) inverse problem. Our experiment results show that the proposed method is efficient and effective.