AI Summary & Key Takeaways
AI InsightCompared with many existing ML models, the developed models offer competitive accuracy, improved reproducibility, and easier deployment, providing a practical and efficient tool for rapid MMP estimation in compositional screening and miscible EOR design.
📌 3 Key Takeaways
Abstract
Accurate minimum miscibility pressure (MMP) estimation is essential for designing and screening miscible gas injection EOR projects, particularly CO 2 and N 2 flooding. Empirical correlations provide quick estimates but are often inaccurate due to limited calibration ranges, oversimplified compositions, and poor representation of nonlinear gas-oil interactions. This study develops explicit machine-learning-based models: neural networks (NNs) and support vector regression (SVR), to estimate CO 2 and N 2 MMPs in petroleum reservoirs. A dataset of 586 samples from the literature, including reservoir temperature, compositional descriptors, and injected gas properties, was compiled and preprocessed. NN models used a feed-forward back-propagation architecture with the Levenberg-Marquardt algorithm, while SVR used a linear kernel for its explicit representation. The models’ performance was evaluated using statistical indices. The NN models achieved AARD, MSE, and R 2 values of 0.1360, 3.260×10 -3 and 0.9768, respectively, for CO 2 , and 0.1395, 3.260×10 -3 , 0.9752 for N 2 . The SVR models achieved AARD = 0.1326, MSE = 0.0444, and R 2 = 0.9522 for CO 2 and AARD = 0.1407, MSE = 0.0422, and R 2 = 0.9835 for N 2 . These statistical indices indicate that the developed models’ predictions were consistent with the measured MMP dataset. The trained models were expressed explicitly in mathematical form using weights, biases, support vectors, and dual coefficients, enabling direct application in engineering calculations and software. Sensitivity analysis identified reservoir-injected gas temperature ratio (T R ) and heavy-end molecular weight (MW C5+ ) as the most influential variables, with H 2 S showing relatively low importance. Compared with many existing ML models, the developed models offer competitive accuracy, improved reproducibility, and easier deployment, providing a practical and efficient tool for rapid MMP estimation in compositional screening and miscible EOR design.