Using Machine Learning to Predict Lithostratigraphic Facies

Abstract

This paper aims to use rapidly growing machine learning applications in geology to predict vertical layers in rock based on properties. These layers in rock with similar chemical and physical properties are referred to as facies. Understanding the underlying strata and various facies informs geologists about the nature of a particular area. The order and nature of the layers in the ground can represent both how the location formed, as well as its evolution over time. This paper takes commonly analyzed wells from a block in the Dutch sector of the North Sea and shows methodology in selected particular models and parameters for prediction. Visual representation of the parameters allows for influence on the facies to be determined. My approach filters through extraneous properties and applies a Butterworth low-pass filter. Because depth is a continuous data parameter that cannot be pieced apart for training data, splitting the training data was an obstacle. However, this problem was circumvented by using a stratified k-fold split. Six different models of supervised learning were directly compared both visually and analytically. Results from these comparisons from the F02-1 well indicate that a K-Nearest-Neighbors model is most accurate and should be used by lithostratigraphic drillers. Results on the test data yielded a prediction accuracy of 99%, but prediction accuracy is yet to be extensively applied to other wells. Finally, a visual reconstruction of the facies of a nearby F02-3 well presents the results of the application and reveals the geographic history of the North Sea.