GNGTS 2017 - 36° Convegno Nazionale

GNGTS 2017 S essione 3.1 597 Due to the need of performing a wide exploration of the model space, the computational time of the GA-FWI depends heavily on the number of the unknowns, which is proportional to the number of the grid nodes. As surface wave modeling requires extremely fine grids, the adoption of a two-grid approach for the inversion is nearly unavoidable. With this approach, a fine grid is used in forward modeling, a coarse grid is employed to represent unknowns in the inversion and bilinear interpolation converts the coarse grid into the fine grid. Although we make use of a global optimization method, which is less affected by the notorious local minimum issues, to further limit this problem and to speed up convergence, Fig. 1 - True Vs near-surface model: a) with sharp velocity contrasts and velocity inversions; b) containing lateral velocity variations; c) with an irregular topographic surface; d) to f) show the model predictions by using our two-grid genetic algorithm FWI code, related sequentially to the models in (a) to (c). Fig. 2 - In each row from left to right the figures correspond sequentially to the models shown in Fig. 1a to Fig. 1c. (a) to (c) show the starting models used in FWI with the local optimization method. They are either the same or the smoothed versions of the models shown in Fig. 1d to Fig. 1f. (d) to (f) display the Vs model predictions of FWI with the local optimization method. In (g) to (i) we display chosen Vs profiles for a clearer comparison among talked- about Vs. All the profiles are picked at the lateral distance of 5 m on the Vs models. The gray profiles correspond to the observed models, the blue to the models predicted by two-grid genetic algorithm FWI, the green to the starting models used by FWI with the local optimization method while the red to the finally refined models by FWI with the local optimization method.