Low-frequency generation and denoising with recursive convolutional neural networks

The absence of low frequencies in seismic data is a major roadblock to the inversion of low wavenumbers with full waveform inversion. We propose to use a recurrent convolutional neural network to denoise existing low frequencies and generate artificial ones from high frequency seismic data. The neural network iteratively uses the same set of filters to halve the central frequency of a gather at each iteration. Under the presence of white noise, the network can lower the frequency content of a seismic gather up to 64 times. The method performs well on real marine data and generates denoised gathers with central frequencies from 40 to 0.64 Hz. The method is readily applicable to hierarchical full waveform inversion, allowing the use of very crude starting models without being affected by cycle skipping. Presentation Date: Wednesday, October 14, 2020 Session Start Time: 9:20 AM Presentation Time: 11:00 AM Location: Poster Station 3 Presentation Type: Poster