Integrating CPI and Core Data into Logistic Regression for Lithofacies Modeling

Abstract

The lithofacies classification research is part of an extended multidisciplinary reservoir characterization and simulation study that has been implemented on the upper shale member/Zubair formation of the southern Iraqi X oil field. This study has been conducted through the Integrated Reservoir Management School (IRMS) at the Basrah Oil Company (BOC). Lithofacies classification is a process to determine rock lithology by analyzing core and well- log data set. Traditionally, lithofacies were classified manually or with the use of some graphing approaches. Many artificial intelligence techniques have recently been adopted to categorize lithofacies. In this work, two robust algorithms were applied to modeling the lithofacies through specific well section (formation), these procedures were adopted and their results were compared to determine more accurate lithofacies classification method. Logistic Boost Regression LBR and Multinomial Logistic Regression MLR were utilized to model the resulting lithofacies as a function of CPI dataset in order to anticipate discrete properties distribution in non-cored depth in wells.

CPI data, which are available for 49 wells in the upper shale Zubair formation, includes: water saturation, porosity (∅_neutron) and volume of shale (V_sh). However, routine core analyses of permeability, porosity and facies are existent for only one well. For that well, the lithofacies types are sand, silty sand and shale. Two supervised statistical learning techniques, LBR and MLR, has been certified to model the discrete lithofacies distribution as a function of the CPI records. The lithofacies classification was then validated through forming the confusion table and computing the accuracy for each method. LBR was observed to be the optimum approach as it led to more accurate lithofacies classification than MLR in clastic reservoirs. The presented workflow demonstrated reasonable facies distribution that leads to strong relationship between porosity and permeability to estimate the petrophysical properties in non-cored wells.

In addition, the posterior lithofacies distribution were plotted to show the probability of spatial distribution and direction of model. These algorithms implemented through R programming a language commonly used in statistical computing by using software packages. Then, these costs for overall data process acquisition could be reduced.