Automatic machine learning (AutoML) for petrophysical evaluation: Case study in Sirikit Field Thailand

Abstract

The petrophysical interpretation is critical for assessing the economic justification. However, the existing workflow of petrophysical assessment is time-consuming. This study aims to investigate the petrophysical interpretation utilizing the machine learning algorithms in the determination of lithology classification and reservoir identification from well log data of Sirikit field. This experiment is based on data from 50 deviated wells located throughout the Sirikit main area, which is the main production area of the Sirikit field, containing oil and gas reservoirs from the Yom, Pratu Tao, and Lan Krabue formations. The programming will concentrate on four well log data types: gamma ray, resistivity, density, and neutron log, as well as two interpretation logs: lithology interpretation and fluid interpretation log. The approach is separated into two basic phases, the first of which is to develop an artificial architecture of neuron networks capable of categorizing lithology, namely sandstone and shale. The lithology will then lead to the secondary goal of reservoir categorization, which includes gas-, oil- and water-saturated-sandstones and shale. This research will focus on the extreme gradient boosting (XGBoost) technique developed as a result of automated machine learning (AutoML). The mean squared error (MSE) and customized error measurement (CEM) accuracy on prediction is the main accuracy metrics used to assess the model score. The best lithology prediction receives an average MSE of 2.76 percent and average CEM of 4.27 percent. Furthermore, the best reservoir classification prediction receives an average MSE of 0.17 percent and average CEM of 1.90 percent. Consequently, the algorithm developed in this work help shorten the time required for petrophysical interpretation.