GNGTS 2019 - Atti del 38° Convegno Nazionale

756 GNGTS 2019 S essione 3.3 temperature chains ensure exploitation of the high-probability regions. According to stochastic criteria, swaps of models are allowed between chains at different temperatures, and in this context the high temperature chains ensure that low-temperature chains access all the high probability regions while maintaining an efficient exploitation capability. Only the models collected at T = 1 are considered in the computation of the PPD because the models collected at T > 1 sample a biased posterior probability density function. Fig. 1 shows the flowchart of the implemented rjMCMC inversion, whereas Fig. 2 illustrates an example of evolution of the RI models sampled by the algorithm. These models represent the RI values estimated over a 2D stratigraphic horizon. We note that the algorithm starts from RI values very different from the reference model and from a Voronoi tessellation with a number of cells that is not enough to successfully predict the observed data. As the iterations proceed, the sampled RI values get closer to the reference model, whereas the Voronoi tessellation successfully partitions the considered 2D horizon by gathering within the same polygon CDPs with similar AVA response. In other terms, the algorithm successfully recognizes adjacent CDP positions with similar AVA responses and similar relative contrasts in the P- and S-impedance values. Synthetic inversion tests. The synthetic observed AVA data are computed from a reference elastic model constituted by a regularly spaced grid of 160 x 240 CDPs. This reference model (leftmost part of Fig. 3) represents a stratigraphic section of a deltaic depositional area characterized by twisted river channel systems where yellow regions are representative of predominantly sandy channels, nested in blue-colored background shale portions with significant lateral contrasts in the RI and RJ values. The Ursenbach and Stewart equation is applied to the reference model to compute the AVA response for each CDP gather position Fig. 2 - Example of evolution of the sampled RI models from iteration 5 (top left) to iteration 2600 (bottom center) over a regular spaced grid of 100x100 CDPs. The reference model is shown on the bottom right corner.