GNGTS 2022 - Atti del 40° Convegno Nazionale

524 GNGTS 2022 Sessione 3.3 According to the performance of the different FC-ANNs on both noise-free and noise contaminated synthetic data, we tested the deepest architecture on a real marine seismic dataset. Real inversion tests. The raw data of the Line 12 of the Viking-Graben 2D marine dataset were acquired in the North Sea by Mobil Corporation and publicly released in 1994 (the dataset can be downloaded at: http://s3.amazonaws.com/open.source.geoscience/open_data/ Mobil_Avo_Viking_Graben_Line_12/mobil_avo.html). These data were processed by Berti (2021) through automatic picking techniques (Toldi, 1989) aimed to retrieve the rms velocity profiles for different CMP gathers. Furthermore, Berti (2021) applied shaping regularization (Fomel, 2007) to regularize the Dix inversion and retrieve the interval velocity model for the Viking-Graben dataset. Since automatic pickers are prone to provide erroneous estimates of the rms velocity profiles, we generated a new noise-contaminated synthetic training dataset considering the range of the Viking-Graben rms velocity model. Inverting the automatically picked rms velocity profiles with the analytic Dix inversion resulted in an un-physical interval velocity model. It was characterized by strong lateral velocity inversions and NaN values were present on the right side of the model. Interestingly, the interval velocity model predicted by the FC-ANN was less affected by the presence of noise in the rms velocity profiles. Indeed, the strong lateral interval velocity variations were smoothed, and numerical instability did not occur. Differently, the interval velocity model obtained with shaping regularization resulted in a smooth interval velocity model. Eventually, we compared the interval and rms velocity profiles obtained through shaping regularization and FC-ANN (Fig. 3) to analyse the difference between model-driven (shaping regularization) and data-driven (FC-ANN) techniques in tackling the ill-conditioning of the Dix inversion. Fig. 3 - Comparison of the inversion results on the Viking-Graben dataset from different techniques. a) Panels on the left column are the interval velocity models estimated from analytic solution (top), shaping regularization (middle), and predicted by the fully connected artificial neural network (bottom); panels on the right are the corresponding rms velocity models. b) Velocity profiles extracted from the various velocity models. The left panel shows the interval velocity profiles obtained with different techniques. In black the profiles obtained with the analytic Dix inversion, in red the outcome of shaping regularization, and in blue the prediction of the FC-ANN. The right panel shows the corresponding rms velocity profiles.