Wavefield compression for seismic imaging via convolutional neural networks

Full Waveform Inversion and Reverse Time Migration are usually based on the adjoint state method and rely on the ready availability of wavefield snapshots. Therefore, standard algorithms store the full source wavefield on disk. This makes Full Waveform Inversion and Reverse Time Migration techniques particularly demanding in terms of disk input/output performance. To face this issue, a common solution is to compress wavefield information in order to reduce input/output operations overhead. In this paper we propose a couple of wavefield compression methods based on Convolutional Neural Networks (CNNs). Specifically, a convolutional autoencoder is trained to compress wavefield snapshots and a specifically designed U-net is trained to reconstruct (i.e. interpolate) the wavefield in the temporal dimension. Results show that these promising techniques could help decreasing storage needs for wavefield snapshots. This makes it possible to balance the required signal-to-noise ratio and compression gain.