Pay Zone Determination by Applying Automatic Machine Learning (AutoML) in the Mckee Field, Taranaki Basin, New Zealand

Abstract

Reservoir characterization is a critical objective in order to understand the subsurface geology and develop geological models that support sustainable and economic oil and gas exploitation. By measuring and analyzing electrical responses of the penetrated successions, it is possible to infer properties about rock matrix and fluid content, among others, that are indirectly related to petrophysical properties of the rocks. Machine Learning (ML) is a data science tool that fits models that are then used to identify patterns or similarities among observations in a dataset in order to make predictions about unobserved data. The main objective of this study is to build a supervised classification model applying AutoML (Automatic Machine Learning) to identify pay zones using information derived from well logs. Twenty wells from the Mckee field, Taranaki Basin, New Zealand were selected containing a basic set of electric logs. For the response variable, reservoir pay properties were estimated. The classification model was first trained with a proportion of the observed data for the independent and response variables and then tested against the remained data set not observed by the model. Model definition and implementation were done with the AutoML function from H2O open source, and model performance was assessed by means of the confusion matrix method. The results indicate that XGBoost (Extreme Gradient Boosting) is the best model for the classification of the Mckee field and that the correct classification of the pay properties is high in all tested wells. This suggests that AutoML procedures can be a valuable tool for the exploration and assessment of the geological properties of the area, helping to reduce operating costs by optimizing the decision-making time during the well evaluation phase as well as the initial petrophysical evaluation to understand reservoir characteristics.