A data mining approach for automatic classification of rock permeability

Abstract

Reliable estimates of porosity can be obtained from different types of geophysical well logs. However, obtaining in situ permeability estimates is still a major challenge in the geosciences. This work aims to evaluate the application of data mining techniques to NMR logs for rock permeability classification, thus far tested only on laboratory data. For this study, we used a petrophysical database from two Brazilian Pre-salt wells located in the Santos Basin, a formation notoriously difficult to characterize, mainly due to its diversity and complexity. Six classification algorithms were evaluated (k-NN, NB, C4.5, RF, SMO, and MLP) according to their ability to estimate the permeability of rocks in four distinct classes (low: <1 mD, intermediate: 1–10 mD, high: 10–100 mD, and excellent: >100 mD). The predictive performance of the algorithms was compared to the behavior of two traditional permeability estimators. With an accuracy of 66%, the Naïve Bayes algorithm, combined with two preprocessing steps – unsupervised discretization and attribute selection – achieved the highest predictive performance. That mark surpassed the accuracy obtained by Kenyon and Timur-Coates estimators by 154% and 106%, respectively, providing evidence for the superiority of the data mining technique to recognize permeability classes based on NMR logs. Classification experiments employing NMR logs in conjunction with conventional logs were also conducted, but this log combination was not able to best the predictive result based solely on the NMR log data.