Permeability Prediction in Marine-Eolian Sediments Using Multivariate Analysis and Nonparametric Regression

Yumei Li 
Department of Statistics, University of Wyoming
Laramie, WY
liyumei@uwyo.edu

Estimating rock permeability from well logs is an important yet difficult task. The three most frequently used techniques in the petroleum industry are empirically determined models, multiple regression, and artificial neural networks (ANN).

In this project, multiple linear regression, generalized additive model (GAM) and alternating conditional expectations (ACE) are used to analyze log and core data from the marine-eolian sediments in the Upper Tensleep Formation, Teapot dome, Wyoming. The quality of the prediction is evaluated by cross validation and compared by examining MSE (mean squared error). Five facies represented are sand dune, interdune, sand sheet, shallow marine and sabkha. The wireline logs used are gamma ray, density, neutron porosity and resistivity.

The approach follows a four-step procedure. First, log data are classified into electrofacies by cluster analysis and lithofacies are determined from core and field sections. Second, all above techniques are investigated for correlations between permeability and well logs in the training set. Third, ACE is employed within each electrofacies and lithofacies in the training set. Fourth, the derived correlations are applied in holdout samples and performances of methods are compared by examining MSE.

Analysis of the predicted values shows a narrower distribution than was observed in the original data set for all methods employed. In random samples, there is not significant difference between the linear regression and nonparametric regression. However, in well samples, ACE within electrofacies appears to outperform the other techniques.