This paper outlines the study results of machine learning application to automate well correlation. Algorithms are designed to address challenges such as geological field complexity, number of wells and geological horizons, and quality of log data, that are usually associated with well correlation
The Kansas state oil and gas fields, USA has more than 350000 wells and produced over 5 billion barrels of oil, with an estimated 11 billion barrels of oil still remaining underground. Nearly 100 wells from Thomas, Scott, Logan, Barton, Stafford and Wichita Counties are selected for this study.
Petrophysical data such as Gama Ray, Neutron, and Density logs, considered most sensitive for lithology and Geological Cores descriptions are used as input. Sedimentary facies variation from limestone, shale, sandstone and coal were used as indicators to separate stratigraphic sequences.
The approach followed is a 1D adaptation of panoramic stitching using feature vectors generated from a 1D Convolution Autoencoder. Convolutional autoencoders learn a compressed representation of input by first compressing the input using encoder and decompressing it back using decoder to match the original output This autoencoder is a neural network trained to reproduce input image in output layer, similar to facial recognition techniques. It identifies a series of matching points between a pair of well logs and trained to pick and match geological well tops.
In order to demonstrate the performance of algorithms in term of accuracy, speed and to address the range of uncertainties, more than 30 wells were used for training. The study results have demonstrated high level of constancy in automatically subdividing the reservoir units of different wells consistent with stratigraphy and consistent across wells even though distance separating for each individual well is considerably more than one kilometer.