Progressive transfer learning for low-frequency data prediction in full-waveform inversion

To effectively overcome the cycle-skipping issue in full-waveform inversion (FWI), we have developed a deep neural network (DNN) approach to predict absent low-frequency (LF) components by exploiting hidden physical relation connecting LF and high-frequency (HF) data. efficiently solve this challenging nonlinear regression problem, two novel strategies are proposed design DNN architecture optimize learning process: (1) dual data feed structure (2) progressive transfer learning. With structure, not only HF data, but also corresponding beat tone fed into relieve burden of feature extraction. The second strategy, learning, enables us train using single evolving training set. Within framework set continuously evolves an iterative manner gradually retrieving subsurface information through physics-based module, progressively enhancing prediction accuracy propelling process out local minima. synthetic numerical experiments suggest that, without any priori geologic information, predicted sufficiently accurate for FWI engine produce reliable velocity models free artifacts.