Hydrocarbon Index Identification in Lateral Section of Horizontal Wells Using Machine Learning

Abstract

Machine learning is a powerful tool that has become an essential part of the oil and gas industry that provides additional insights at limited data availability. In the vertical section, representative samples can be obtained, but in the horizontal section, core and well log data are extremely rare. With machine learning, mineralogy data can be trained to predict Total Organic Conten (TOC) and amount of free hydrocarbons (S1) and implemented in the lateral section of a well. X-Ray Diffraction (XRD) data were used to train the Machine Learning Model to predict TOC & Flowable Hydrocarbon Index (FHI). For training, the vertical section of a well along with data from other wells nearby was used. It was then tested on the horizontal section of the same well to make certain that the model had learned meaningful relationships. The data were pre-processed to remove outliers, clean string variables, and missing values. New features were then added using domain knowledge, as some features are non linearly correlated with TOC. For example, certain minerals correlate with TOC and S1 differently in different flow units. Because of this, we made sure to collect data for training and testing within the same flow unit. The features were then standardized to ensure that the resulting distribution had a mean of 0 and a standard deviation of 1. This was done for two reasons, firstly, it makes sure the Machine Learning algorithm gives equal weightage to all features irrespective of their units, and range of values. Secondly, it helps the algorithm to converge faster, taking less time to train. The data was then fed to the model to train on. Once the model was trained, the hyperparameters of the model were then tuned using a suitable error metric. This was done in order to make certain that the model performs well on test data, and does not overfit on the training data. Results of the trained model were compared to actual measured results from the rock along the horizontal wellbore. The predicted and measured results show a great match based on several metrics such as coefficient of determination, mean absolute error, mean squared error, and root mean squared error. The key observation was the importance of the sample acquisition, sample handling, eliminating human error during measurements in the lab, and data handling when applying the model. In conclusion, when the data is properly clustered using geological understanding it is meaningful to apply machine learning algorithms to find multivariate relationships among different parameters. This paper presents a novel approach of obtaining volumetrics data based on "ground truth" measurements such as saturations along the horizontal wellbore without operational risks, potential downtime, and chances to lose tools.