Compositional modeling of solution gas–oil ratio (Rs): a comparative study of tree-based models, neural networks, and equations of state

Accurate knowledge of crude oil pressure-volume-temperature (PVT) properties is essential for both industrial and academic applications. However, traditional experimental methods for determining these properties, particularly the solution gas-oil ratio (R_s), are time-intensive and costly. In this study, advanced compositional models were developed using a broad range of machine learning (ML) techniques to predict R_s efficiently and reliably. A comprehensive database of 1,154 data points was utilized for modeling. Among the tested models, the extra trees (ET) algorithm demonstrated superior performance, achieving an average absolute percent relative error (AAPRE) of approximately 3%, indicating its high reliability for R_s prediction. Additionally, R_s was estimated using seven different equations of state (EoS). Systematic graphical and statistical evaluations revealed that the Schmidt-Wenzel (SW) EoS was the most accurate among the conventional methods, with an average error of approximately 11%. The robustness of the ET models was validated across various temperature ranges, with detailed trend analysis confirming their ability to accurately capture the physical relationship between R_s and pressure. A relevancy factor analysis quantified the influence of each input parameter on model outputs, whereas the Leverage technique identified outliers and defined the parameter ranges for optimal algorithm performance. While the ML models achieved high predictive reliability, their computational demands and complexity may limit deployment in resource-constrained environments and decision-critical applications. Nevertheless, this study represents a significant advancement in R_s predictive modeling, providing robust, scalable, and cost-effective tools for academic and industrial applications.