Advanced machine learning techniques for predicting solution gas-oil ratios in petroleum reservoirs: A comprehensive study and new empirical correlation

The precise assessment of crude oil properties is crucial for various tasks in reservoir evaluation, including estimating reserves, enhancing oil recovery, predicting oil reservoir performance, designing pipelines and production equipment, and simulating reservoirs. Therefore, the reservoir fluid properties, especially the solution gas-oil ratio (GOR or Rs), which depends on bubble point pressure (Pb), reservoir temperature (Tr), gas specific gravity (γg) and API are of great importance. On the other side, due to the time-consuming and expensive nature of pressure-volume-temperature (PVT) laboratory models, accurate experimental correlations are significant in petroleum engineering. Therefore, a cheap, accurate, fast model is needed to estimate these parameters. In this study, to develop a comprehensive and reliable model, 1309 data points were collected from crude oil samples in different geographical locations. The development of a new correlation was presented with high accuracy. Also, intelligent models based on neural networks such as Multilayer perceptron neural networks (MLP), Radial basis function network (RBF), Extreme gradient boosting (XGBoost), Adaptive boosting (AdaBoost), and Gradient boosting with categorical features support (CatBoost), were used to predict Rs. To assess the performance of the models, statistical and graphical error analyses, sensitivity analyses, group errors, model trends, and a leverage approach were employed. The CatBoost model delivered the highest predictive accuracy, with an AAPRE of 8.13%, surpassing competitors like XGBoost (10.51%) and AdaBoost (24.78%). In addition, the new correlation developed achieved an AAPRE of 16.40%, offering a marked advancement over conventional empirical correlations that typically range from 17.24% to 64.88% in error. These findings demonstrate the superior precision and versatility of methods, underscoring its innovative value for estimating Rs across varied reservoir environments. Factorial sensitivity analysis showed that Pb had the greatest impact on Rs with 25.4%.