GNGTS 2019 - Atti del 38° Convegno Nazionale

728 GNGTS 2019 S essione 3.3 brine sand and gas sand. These deviations disappear after the normal score transformation (Fig. 1d). In the synthetic seismic inversion example, we simulate a signal-to-noise ratio of 2 in the observed data, whereas an angle range of [0, 30] degrees and a 55-Hz Ricker wavelet are used to compute the seismic data. In the MCMC inversion we use 40 different chains running for 10000 iterations each and with a burn-in period of 5000: 20 chains run at T =1, while the remainder at logarithmically spaced temperature values. We consecutively perturb the elastic properties or the facies configuration at ten different time positions before the likelihood evaluation. Fig. 2 - Estimated elastic properties provided by the analytical (a) and the MCMC (b) inversion for the synthetic inversion test. The black lines represent the true property values, the red lines are the estimated mean models, whereas the colormap codes the estimated posterior pdf . The rightmost plots show a comparison of the observed (black) and the predicted (red) seismic data computed on the estimated a-posteriori mean models. The analytical and MCMC approach yield similar predicted elastic profiles (see Figs. 2a and 2b), although the MCMC inversion often provides slightly superior prediction intervals as demonstrated by the coverage probability values (not shown here for the lack of space). The differences between the outcomes of the two approaches can be clearly appreciated by comparing the facies classification results (Fig. 3). Indeed, just a visual inspection of the estimated facies models and the associated posterior pdfs confirms that the MCMC method (Fig. 3b) outperforms the analytical inversion (Fig. 3a) as it estimates a maximum-a-posteriori (MAP) facies solution with a closer match with the actual facies profile especially below 935