GNGTS 2019 - Atti del 38° Convegno Nazionale

GNGTS 2019 S essione 1.3 187 transmitted waves as the observed data, we perform full waveform inversion (FWI) to estimate the subsurface velocities. FWI, originally introduced by Tarantola (1984), differently from other inversion methods (such as traveltime tomography), takes into account the whole information (traveltime, amplitude and phase) of the input data and has a strong potential for reconstructing the detailed velocity field by minimizing the data misfit between the predicted and observed seismograms. In general, the inverse problem is often framed as a local optimization problem making it necessary to compute the gradient of the data misfit with respect to the model parameters. This approach has the advantage of being computationally efficient, but requires the knowledge of a suitable starting model from which to trigger the inversion. A suitable starting model should reproduce the true long-wavelength structures so to avoid the inversion being stuck in local minima of the objective function. It is then obvious that a substantial knowledge of the morphology of the subsurface should be known before the inversion. This may be not a problem in many cases, where well data are available and/or where there already is an established, large-scale, geological reconstruction of the subsurface. In other cases, such a knowledge is not available or conflicting interpretations may exist and thus the choice of a suitable starting model, which in strongly non linear problems, such as the one at our hands, could set the gross structure of the final FWI model, is troublesome. For example, for what concerns the area of the CROP/18 profile, searching the literature (Bertini et al., 2006; De Matteis et al. , 2008; Liotta and Ranalli, 1999, among others) we found at least three different hypotheses (metamorphic aureola, petrophysical factors, brittle-ductile transition,) on the nature of the deep “K-horizon”. To avoid being biased by a possibly-erroneous initial model we cannot use linear or linearized optimization techniques: we propose to employ stochastic optimization algorithms, such as genetic algorithms, that performing a global exploration of the model space are much less hampered by local minima issues and do not require any starting model. Then, once that a low resolution, long-wavelength model has been estimated stochastically, we can move on applying local optimization algorithms to improve the model resolution and produce a detailed image of the velocity structure of the subsurface. A brief outline of the proposed FWI methodology. The inversion strategy we propose consists in performing the FWI in two successive steps: the first employs a global optimization method, genetic algorithms (GA), as the optimization method, while the second step employs a local optimization method based on the computation of the gradient (GB) of the data misfit function. GA do not require a starting velocity model, meaning that the inversion starts from a “population” of many models, randomly selected within a predefined search space. In the case at our hand, the search space varies from 2 km/s to 6 km/s at the surface, and from 4 km/s to 8 km/s at 2.5 km depth.Then, the population is let evolve towards better and better models mimicking genetic evolution. The validity, or the “fitness”, of each model is measured by Fig. 1 - Example of shot gather. The shaded area evidences the transmitted waves.