GNGTS 2018 - 37° Convegno Nazionale

GNGTS 2018 S essione 3.3 719 each petrophysical properties at each time sample as an independent model parameter to be optimized. This unfortunate parameterization results in a total number of 414 unknowns. Fig. 1b shows the best predicted model selected from three different inversion runs compared with the actual petrophysical property values. We observe that the predicted petrophysical values are totally different from the true ones and characterized by unrealistic vertical variations, although resulting in a perfect match with the observed data. This indicates that the standard parameterization results in a hopelessly ill-conditioned problem characterized by a wide null- space of solutions. 1DAVA inversion. In this case we consider a non-linear AVA inversion performed on actual well log data. The full Zoeppritz equations are again used as the elastic forward modelling. In this case the unknowns are the Vp , Vs and density values for each time sample and the objective function to minimize is the simple L2 norm misfit between predicted and observed seismic gathers. A S/N ratio equal to 10 is simulated on the observed data. I still consider an angle range between 0-45 degrees and a 45-Hz Ricker wavelet with a sampling rate of 0.002 s as the source signature. To decrease the ill-conditioning, I now include a 1D EPS filter into the optimization kernel, thus obtaining a two-layer optimization framework. In the first layer, FA algorithm works independent of the EPS filter, that is the objective function is evaluated for each individual in the swarm. Then the EPS filter is applied to each individual, and the filtered version is accepted within a Markov process. In other terms the filtered individual is accepted Fig. 2 - AVA inversion: Comparison of the results and data gathers provided by the proposed (a) and the standard approach (b).