GNGTS 2019 - Atti del 38° Convegno Nazionale

GNGTS 2019 S essione 3.1 561 patches. Fig.1 shows two examples of snapshot interpolation through the proposed U-net with k set to 12. The snapshots in the top rows display the ground truth (i.e. the actual propagating wavefield), while the bottom rows the corresponding U-net outputs. In particular, the snapshots highlighted in red show the k−3 interpolated snapshots. The overall Signal-to-Noise Ratio (SNR) achieved on the testing dataset for the interpolated snapshots via U-net is always above 40 dB. Fig. 1 - Two examples of pairs of original and interpolated wavefields. As for the autoencoder based compression strategy, the training was performed on a set 10000 snapshots ( 75% training set and 25% validation set), with 50 epochs of Adam optimization an (batch size set to 32). The performances in terms of Compression Ratio (CR) and SNR are driven by the regularization weight for the loss function enhancing sparsity of the hidden representation and by the threshold for the hidden representation values. Given a desired SNR, we heuristically selected the best configuration that maximizes the CR, obtaining a linear link between the increase of CR and the decrease in SNR. In Fig. 2 we show an example of autoencoder reconstruction results for different values of SNR and achieved compression ratio: ratios superior to 100 still guarantee an SNR superior to 25 dB. Conclusions. We studied a strategy for CNN-based wavefield compression. Specifically, we studied two different procedures: a convolutional autoencoder used to separately compress snapshots, and another CNN architecture (i.e. a U-net) that interpolates wavefields in time. Fig. 2 - Examples of compressed snapshots for different compression ratio (CR). From left to right, Original snapshot, CR=113 (SNR=25.3 dB) and CR=140 (SNR=25.2dB)