GNGTS 2015 - Atti del 34° Convegno Nazionale

GNGTS 2015 S essione 3.1 17 3a and 3d. In addition, we note a fair similarity between the rock-physics models derived by the empirical SR and GAmethods and the one resulting from the TRPM. All these three RPMs predict similar Vs increases with the decreasing of shale content and similar Vs increases as the porosity decreases. Conversely, the rock-physics model obtained by the NN method is substantially different from the other ones: it shows an un-physical Vs decrease for a shale content less than 40-50%, approximately. We interpret this fact as an overfitting problem that usually affects the NN method (see van der Bann and Jutten, 2000). In particular, in the Fig. 3 – Graphical representations of the rock-physics models derived by step-wise regression (SR), Neural Network (NN), theoretical rock physics model (TRPM) and genetic algorithms (GA) are, respectively, shown in (a), (b), (c) and (d). These surfaces represent the Vs variations as a function of the shale content and the porosity, keeping fixed the depth and the water saturation to 2700 m and to 50%, respectively. (e) and (f) Results of the blind test and the corresponding correlation coefficients. See the text for additional comments.