GNGTS 2019 - Atti del 38° Convegno Nazionale

760 GNGTS 2019 S essione 3.3 inversion algorithms. The observed data consists of ten shots separated by 460 m with 192 receivers each and was computed by the finite difference acoustic forward modelling algorithm described in Galuzzi et al. (2017). The seismic inversion was carried out by means of the two grid approach of Sajeva et al. (2016). The coarse grid is defined by 17 vertical nodes separated by 72 m and 10 horizontal nodes separated by 460 m. The fine or modeling grid consists of 48x192 nodes with distances of 24 m. The idea behind this experiment is to verify if the capability of the Wasserstein norm to compute a misfit function with a reduced number of local minima can help the GA inversion to converge with a reduced computational effort to a model containing the long wavelength structure of the true model. Fig. 3 shows the inversion based on this metric. It seems to give a correct distribution of the model velocity anomalies within the low resolution imposed by the the coarse grid. However the main difficulty encountered in this test was to design an adequate processing sequence to apply to the data in such a way to satisfy the requirements of the Wasserstein norm. For instance, the fact that the inverted data must be positive, force the introduction of some processing procedures that can make the convergence towards a satisfactory solution more difficult. Conclusions. This study showed that using the Wasserstein metric as the object function in the first iterations of an inversion process may be a reliable approach to avoid cycle skipping; however a careful processing must be applied to prepare the observed and predicted data for the inversion which are required to be positive. Due to the low resolution that characterizes the objective function of this metric, a possible approach to improve the obtained result can be a successive application of a different norm, such as the L2 norm. Concerning the other metrics that have been tested, even if the instantaneous envelope and phase need approximately the same computational time as the Wasserstein norm, they are more sensitive to the noise present in the data; while the AWI approach (in our implementation) requires a proibitive computational time and does not seem to be as performing as the Wasserstein metric against the cycle-skipping. Fig. 3 - (On the right) Final model obtained inverting with the Wasserstein metric data from a portion of the Marmousi model (on the left). The process was able to detect the presence of low-velocity structures in the depth interval between 200 and 800 m.