Modeling sonic logs in oil wells: a comparison of neural networks ensembles and kernel methods

Abstract

Oil well logs are frequently used to determine the mineralogy and physical properties of potential reservoir rocks, and the nature of the fluids they contain. Recently we reported an exploratory use of neural network ensembles for modeling these records. We showed that ensembles are clearly superior to linear multivariate regression as modeling technique, revealing an underlying nonlinear functional dependency between the correlated variables. In this work we use kernel methods to develop nonlinear local models relating Sonic logs (transit time of compressional waves) with other commonly measured properties (Resistivity and Natural Formation Radioactivity Level or Gamma Ray log). The kernel considered is conceptually simple and numerically robust, and allows to obtain the same performance as neural networks ensembles on this task.