Open Access

ARTICLE

Models for Predicting the Minimum Miscibility Pressure (MMP) of CO₂-Oil in Ultra-Deep Oil Reservoirs Based on Machine Learning

Kun Li¹, Tianfu Li^2,*, Xiuwei Wang¹, Qingchun Meng¹, Zhenjie Wang¹, Jinyang Luo^1,2, Zhaohui Wang¹, Yuedong Yao²
1 Huabei Oilfield Company, PetroChina, Renqiu, 062552, China
2 College of Petroleum Engineering, China University of Petroleum (Beijing), Beijing, 102249, China
* Corresponding Author: Tianfu Li. Email:
(This article belongs to the Special Issue: Integrated Geology-Engineering Simulation and Optimizationfor Unconventional Oil and Gas Reservoirs)

Energy Engineering https://doi.org/10.32604/ee.2025.062876

Received 30 December 2024; Accepted 19 March 2025; Published online 24 April 2025

Abstract

CO₂ flooding for enhanced oil recovery (EOR) not only enables underground carbon storage but also plays a critical role in tertiary oil recovery. However, its displacement efficiency is constrained by whether CO₂ and crude oil achieve miscibility, necessitating precise prediction of the minimum miscibility pressure (MMP) for CO₂-oil systems. Traditional methods, such as experimental measurements and empirical correlations, face challenges including time-consuming procedures and limited applicability. In contrast, artificial intelligence (AI) algorithms have emerged as superior alternatives due to their efficiency, broad applicability, and high prediction accuracy. This study employs four AI algorithms—Random Forest Regression (RFR), Genetic Algorithm Based Back Propagation Artificial Neural Network (GA-BPNN), Support Vector Regression (SVR), and Gaussian Process Regression (GPR)—to establish predictive models for CO₂-oil MMP. A comprehensive database comprising 151 data entries was utilized for model development. The performance of these models was rigorously evaluated using five distinct statistical metrics and visualized comparisons. Validation results confirm their accuracy. Field applications demonstrate that all four models are effective for predicting MMP in ultra-deep reservoirs (burial depth >5000 m) with complex crude oil compositions. Among them, the RFR and GA-BPNN models outperform SVR and GPR, achieving root mean square errors (RMSE) of 0.33% and 2.23%, and average absolute percentage relative errors (AAPRE) of 0.01% and 0.04%, respectively. Sensitivity analysis of MMP-influencing factors reveals that reservoir temperature (T_R) exerts the most significant impact on MMP, while Xint (mole fraction of intermediate oil components, including C₂-C₄, CO₂, and H₂S) exhibits the least influence.

---

Keywords

MMP; random forest regression; genetic algorithm based back propagation artificial neural network; support vector regression; gaussian process regression

---