A research team led by Alexei Rozhenko, Fahimeh Hadavimoghaddam, Erfan Mohammadian, and Peyman Pourafshary has evaluated data-driven and explainable machine learning models to determine the minimum miscibility pressure of carbon dioxide across a broad spectrum of reservoir temperatures and fluid compositions. Minimum miscibility pressure, commonly abbreviated as MMP, defines the lowest pressure at which injected carbon dioxide creates a single-phase or low interfacial tension front with reservoir oil under operating subsurface conditions. In petroleum engineering and carbon management, precise determination of this boundary is vital for planning enhanced oil recovery projects as well as carbon capture, utilization, and storage initiatives. When gas injection operations underestimate the required pressure, the displacement process loses displacement efficiency, resulting in diminished oil recovery.
Traditional laboratory approaches used to assess minimum miscibility pressure include slim tube tests, rising bubble apparatus, pressure-composition diagrams, and vanishing interfacial tension procedures. While these physical experiments capture complex fluid interactions and phase behavior within porous media, they remain costly, labor-intensive, and inherently discrete, making continuous reservoir simulation difficult without mathematical paradigms. To circumvent experimental delays, the industry has historically relied on empirical correlations formulated by past researchers such as Lee, Yellig and Metcalfe, Orr and Jensen, Holm and Josendal, Alston, Glasø, and Shokir. However, many empirical formulas rely on narrow parameter ranges, limited observational databases, or simplified temperature dependencies, frequently causing severe prediction errors when deployed outside their development conditions. Furthermore, rigorous theoretical models governed by equations of state often impose substantial computational burdens and depend on rigid assumptions.
To overcome these limitations, the authors constructed a comprehensive data bank comprising 910 experimental observations gathered from diverse published literature sources. The compiled database reflects crude oil and carbon dioxide systems incorporating pure carbon dioxide as well as impure gas streams containing components such as nitrogen, argon, hydrogen sulfide, and carbon monoxide. In setting up the computational framework, the researchers compiled 21 physical and chemical attributes, encompassing reservoir temperature, critical temperature, the molecular weight of intermediate and heavy hydrocarbon fractions, and the mole fractions of volatile, intermediate, and light hydrocarbon components. To systematically examine how different combinations of input variables impact predictive capacity, the researchers structured the database into five distinct functional subsets, aligning with conventional compositional parameters established in modern reservoir engineering literature.
Four distinct machine learning algorithms were deployed and compared across the partitioned datasets: extreme gradient boosting (XGBoost), categorical boosting (CatBoost), group method of data handling (GMDH), and genetic programming (GP). XGBoost operates as a tree-based ensemble architecture that coordinates regression trees through iterative minimization of a regularized objective function, using shrinkage and regularization parameters to mitigate overfitting. CatBoost incorporates permutation strategies and target-based categorical encodings to handle complex categorical features effectively. While both gradient-boosting frameworks offer high computational accuracy, their ensemble structures function as complex algorithms that do not yield direct closed-form formulas. In contrast, GMDH builds a polynomial neural network with intermediate quadratic equations solved through least-squares algorithms, and genetic programming harnesses evolutionary principles such as crossover and mutation to generate explicit algebraic correlations.
Across the investigated datasets, the machine learning models achieved high overall predictive performance, with coefficient of determination values ranging from 0.89 to 0.99 for several optimized configurations. XGBoost and CatBoost consistently registered the highest predictive accuracy across all evaluation metrics, including root mean square error, mean absolute error, and residual distribution spans. In residual examinations on the first dataset, XGBoost maintained tightly clustered relative error bounds largely between -0.15 and 0.15, exhibiting the smallest deviation around the zero-error line. CatBoost performed with comparable accuracy, displaying only slight additional dispersion on testing records. Genetic programming exhibited lower relative accuracy and wider residual spreads, yielding errors between -0.4 and 0.3 on the same data configuration, while group method of data handling maintained intermediate performance with useful predictive fidelity.
The structural advantage of group method of data handling and genetic programming lies in their direct mathematical transparency, providing practitioners with explicit equations that can be reviewed and implemented without proprietary software platforms or machine learning source code. In the equations formulated through genetic programming, constant coefficients and input variables are directly evaluated in mathematical expressions without sequential neuron dependencies. In contrast, the group method of data handling generates intermediate polynomial variables calculated sequentially across progressive layers before determining the final target estimate. Because the genetic programming expressions do not require layered iterative calculations, they execute with faster computational evaluation times, whereas group method of data handling provides systematic polynomial approximations that closely approximate underlying fluid interactions.
To evaluate physical consistency and explain the underlying models, the authors conducted sensitivity analyses using relevance factor computations to quantify how changes in individual input variables influenced predicted minimum miscibility pressure values. Across the general database and specific model subsets, critical temperature and reservoir temperature emerged as the two most influential input variables affecting minimum miscibility pressure. Fluid composition fractions also exercised significant control; for instance, intermediate hydrocarbon components and pentane-to-hexane fractions exhibited negative correlation scores of -65 percent and -55 percent, respectively, although local collinearity within specific subsets occasionally altered individual feature rankings.
Performance comparisons against conventional empirical correlations highlighted the marked predictive gains achieved by data-driven algorithms. When evaluated against established literature correlations, including methods published by Alston, Cronquist, Yellig and Metcalfe, Orr and Jensen, Shokir, Kamari, and Fathinasab and Ayatollahi across 224 comparative records, the newly developed algorithms demonstrated superior reliability. In published benchmarks, empirical correlations frequently exhibited high average absolute relative deviations, such as 18.95 percent for Orr and Jensen on pure carbon dioxide streams and 31.79 percent when coupled with the Sebastian adjustment factor on impure streams. In comparative testing against independent data excluded from the initial training bank, the study’s data-driven models outperformed classical empirical formulations by margins reaching up to 67 percent.
Despite the high accuracy documented across the experimental range, the researchers emphasized crucial technical constraints associated with machine learning applications in reservoir studies. Like traditional empirical models, data-driven frameworks are fundamentally bounded by the domain of their training sets and are susceptible to estimation errors when applied to conditions outside their established parameter ranges. Model precision also tends to decrease near extreme boundary values where observational records are scarce, as statistical training algorithms require adequate density to capture underlying distributions. Furthermore, subset evaluations revealed that data quality and feature selection strongly dictate algorithm success; while datasets containing comprehensive compositional features yielded determination coefficients up to 0.99, a reduced three-parameter subset exhibited noticeably lower accuracy, underscoring that machine learning cannot fully compensate for missing physical information.
The study demonstrates that explainable data-driven techniques provide a viable, fast, and accurate alternative to costly experimental testing and rigid empirical correlations in reservoir pressure forecasting. By delivering explicit mathematical equations alongside advanced gradient-boosted ensembles, the research offers adaptable tools for varied engineering contexts, allowing practitioners to balance raw precision against operational reproducibility. The authors note that expanding experimental repositories to encompass additional injection gases, such as methane, along with wider compositional extremes, will be essential for extending these explainable frameworks to a broader variety of miscible gas injection and carbon storage applications.
Subject of Research: Explainable machine learning models for carbon dioxide minimum miscibility pressure prediction
Article Title: Evaluation of explainable machine learning techniques for CO2 minimum miscibility pressure determination in a vast range of temperatures and compositions
Article References: Rozhenko, A., Hadavimoghaddam, F., Mohammadian, E., & Pourafshary, P. (2026). Evaluation of explainable machine learning techniques for CO2 minimum miscibility pressure determination in a vast range of temperatures and compositions. International Journal of Coal Science & Technology, 13(1), Article 77. https://doi.org/10.1007/s40789-026-00908-3
Image Credits: AI Generated
DOI: 10.1007/s40789-026-00908-3
Keywords: carbon dioxide flooding, minimum miscibility pressure, machine learning, extreme gradient boosting, categorical boosting, group method of data handling, genetic programming, enhanced oil recovery, carbon capture utilization and storage, reservoir temperature, critical temperature, hydrocarbon compositions, empirical correlations, sensitivity analysis, data-driven modeling, petroleum engineering
Cite Scienmag News
Teresa Odom. (November 20, 2026). Explainable Machine Learning Accurately Determines Carbon Dioxide Minimum Miscibility Pressure. Scienmag. https://scienmag.com/explainable-machine-learning-accurately-determines-carbon-dioxide-minimum-miscibility-pressure/
Teresa Odom. "Explainable Machine Learning Accurately Determines Carbon Dioxide Minimum Miscibility Pressure." Scienmag, 20 November 2026, https://scienmag.com/explainable-machine-learning-accurately-determines-carbon-dioxide-minimum-miscibility-pressure/. Accessed 20 November 2026.
Teresa Odom. "Explainable Machine Learning Accurately Determines Carbon Dioxide Minimum Miscibility Pressure." Scienmag. November 20, 2026. https://scienmag.com/explainable-machine-learning-accurately-determines-carbon-dioxide-minimum-miscibility-pressure/








