Screening of Enhanced Oil Recovery Methods Using Supervised Machine Learning Predicated on Range Data

Enhanced oil recovery (EOR) is the tertiary stage of subsurface hydrocarbon extraction from an oilfield. It involves the use of chemical and thermal procedures to manipulate the properties of the in situ rock and/or fluids to mobilize a good fraction of the residual oil left over from prior primary and secondary stages. EOR screening is the first step in the procedure of selecting the best EOR technology that could be applied to a specific reservoir, before the economic analysis step. EOR screening is a challenging multi-criteria decision-making process, considering that are about twenty possible methods to choose from. The conventional pre-defined screening criteria are based on hundreds of successful EOR projects undertaken worldwide, aggregated into fifteen EOR methods (classes) each defined by nine reservoir properties (features). However, the features’ data are given in the form of acceptable ranges of values (minimum, mean and maximum). In this study, we adopted a supervised machine learning (ML) technique capable of creating models for making classifications predictions. We introduce a novel triangular distribution approach to generate a semi-synthetic dataset of 500 data rows (instances) for each EOR method, resulting in a total 7500 rows. This was used to train and test two multiclass classification algorithms – Logistic Regression and Random Forest. The Random Forest model gave an overall classification accuracy of 97% compared to 91% for the Logistic Regression. The ML models created herein are one-step solutions for recommending the most technically applicable EOR method for a new project.