GNGTS 2022 - Atti del 40° Convegno Nazionale

GNGTS 2022 Sessione 3.3 525 This comparison highlighted that the model-driven technique was mainly driven by the a priori assumptions, providing less consistent solutions with the observed data (automatically picked rms velocity). Differently, the predictions obtained by the data-driven technique were governed by the statistics of the training dataset. In particular, the predicted interval velocity profiles were more blocky and less affected by noise than the ones obtained by using the analytic solution. Furthermore, the rms velocity profiles calculated from the FC-ANN predictions were consistent with those obtained with the automatic picker (Fig. 3). Finally, we migrated the data using the interval velocity model obtained through shaping regularization and the one predicted by the FC-ANN, to fairly compare the quality of the two models. Considering the moderately complex geological setting of the Viking-Graben, we used the pre-stack Kirchhoff depth migration (PSDM) method to generate two migrated seismic sections and extract the common reflection point (CRP) gathers. Interestingly, the focusing of the seismic sections and the analysis of the flatness of CRP gathers highlighted that the shallower portion was better imaged by the interval velocity model predicted by the FC-ANN. Differently, the deeper portion was more correctly imaged using the model obtained with shaping regularization. This difference could be related to the presence of dipping reflectors in the deeper portion of the Viking-Graben, resulting in the violation of the 1D media assumption of the Dix inversion. Conclusions . The potentialities of fully connected artificial neural networks (FC-ANNs) in solving non-linear problems are well known and confirmed by their application in this study to solve the Dix inversion. Considering this very simple inverse problem, a purely data-driven solution could not achieve the same accuracy as the analytic solution in case of noise-free data. Differently, training a FC-ANN on a noise-contaminated labelled dataset allowed to update the internal parameter of the architecture to attenuate the effect of noise in the inference phase. Furthermore, the FC-ANNs were robust with respect to the statistics of noise contaminating the training data. Eventually, including realistic features and noise within the synthetic training dataset highlighted that machine learning models can be effectively applied to real cases when large amounts of labelled training data are not available. The results of this study might be integrated with further research focused on combining machine-learning to standard inversion algorithms, exploiting regularizers learned from training datasets. Moreover, methods aimed to explain which aspects drive the predictions of a machine learning model need to be developed. Acknowledgments. The first author would like to thank Eni S.p.A. for the opportunity to collaborate with the AESI Department. In particular, Doctor Bienati Nicola is gratefully acknowledged for his assistance and supervision during the development of this work. Seismic data migration was performed using the PROMAX ® software of Halliburton/ Landmark, which is gratefully acknowledged. References Berti S.; 2021: Full wavefield velocity analysis and imaging on a 2D marine dataset. M.Sc. Thesis in Exploration and Applied Geophysics, University of Pisa (IT). Dix, C. H.; 1955: Seismic velocities from surface measurements. Geophysics, 20(1), 68-86. Fomel S.; 2007: Shaping regularization in geophysical-estimation problems. Geophysics, 72(2), R29-R36. Koren Z. and Ravve I.; 2006: Constrained Dix inversion. Geophysics, 71(6), R113-R130. LeCun Y., Bengio Y. and Hinton G.; 2015: Deep Learning. Nature, 521, 436-444. Toldi J.A.; 1989: Velocity analysis without picking. Geophysics, 54(2) , 191-199. Yu S. and Ma J.; 2021: Deep Learning for Geophysics: Current and Future Trends. Reviews of Geophysics, 59(3). DOI: 10.1029/2021RG000742.