GNGTS 2024 - Atti del 42° Convegno Nazionale

Session 3.3 GNGTS 2024 Fig. 3 (a) The lefmost shot of the observed data in black, compared against the data generated using one of the startng models (red); (b) Comparison between the same observed shot gathers (in black) and the data generated from the posterior mean model (in red); (c, d) Comparison of two seismic traces of the same shot of the observed data (green), the predicted data (red) and the inital data (dashed blue). Conclusions In this study we introduced a computatonally efcient Bayesian elastc FWI, leveraging a GB- MCMC sampling technique along with a DCT compression applied to both the model and data spaces. The adopted MCMC strategy addresses the cycle-skipping problem afectng the local FWI approach, utlizing the local gradient and Hessian informaton of the posterior density to guide the sampling towards the most promising regions of the model space. This results in a signifcant reducton in computatonal burden of the probabilistc approach compared with standard gradient- free MCMC algorithms. We demonstrated the efcacy of the GB-MCMC elastc FWI applied to a feld dataset, acquired in Grenoble. Afer pre-processing the seismic data in order to make it comparable with the data generated using the elastc forward modelling, fve Markov chains were employed to numerically assess the PPD, each one startng from very simple inital models. The predicted posterior mean model accurately replicated all the vertcal velocity variatons evidenced by the available borehole data. In additon, our model predicton is also capable to closely match the observed seismic data, afrming the applicability and reliability of the proposed approach that can also be conveniently used to defne a startng point for a subsequent step of local inversion, aimed at enhancing the velocity model’s resoluton and further minimize the diference between predicted and observed data. Our ongoing research focuses on optmizing the overall computatonal efciency of our inversion procedure through the integraton of deep learning techniques.