Investigating the dependence of shear wave velocity on petrophysical parameters

The dependence of the shear wave velocity ( Vs) of water-saturated reservoir rocks on petrophysical parameters was investigated. Two general regression neural network (GRNN) models were developed to predict V-s of sandstones, shaly sands, and carbonate rocks as a function of the compressional velocity (V-p), grain density (rho(g)), clay content, porosity (phi), permeability (k), and the cementation exponent (m) at a fixed effective stress and frequency. A set of 59 sample measurements of clean and dirty sandstones and carbonate rocks was used to train and test the GRNN models. The first model predicts V-s as a function of phi, k, rho(g), m, and the clay content expressed as a percentage. This GRNN model was trained using two data sets consisting of 80% and 70% of the data, respectively. The truncated portions of the data were used as blind test sets to validate the GRNN model. The analysis reveals that the porosity, clay content, grain density, permeability, and cementation exponent are essential variables for capturing the variance of the shear wave velocity. Porosity appears to be the most important variable for estimating V-s, and the grain density is the second most important variable. The GRNN model was able to estimate V-s from both the training and the blind test data sets within an average absolute error of approximately 6%. Models from the literatufe that predict V-s as a function of porosity and clay content appear to be specific to particular lithologies, such as sandstones or carbonates. Modeling V-s by including parameters such as grain density, cementation exponent, and permeability in addition to porosity and clay content appears to capture the dominating physics that affects V-s. By including this comprehensive set of input parameters, we have developed a generalized model to predict the shear wave velocity V-s value for all types of lithology, such as carbonates and clean and dirty siliciclastics. A second GRNN model predicts V-s as a function of only V-p. This model was also trained using two data sets that consisted of 80% and 70% of the data, respectively. Similar to the first model, the truncated portions of the data were used as blind test sets to validate this GRNN model. The second GRNN model was able to estimate V-s as a function of V-p from the training sets within an average absolute error of approximately 4% while the average absolute error was 3% for the blind test data sets. The compressional velocity alone appears to be sufficient to predict the shear wave velocity. (C) 2016 Elsevier B.V. All rights reserved.