Scholarlink

<p>Precise estimation of minimum miscibility pressure (MMP) is advantageous for CO<sub>2</sub>-based EOR applications and CO<sub>2</sub> capture, sequestration and utilization. However, experimental methods, including slim tube tests, are costly and time-consuming. In addition, conventional empirical correlations generally have a limited range of applicability. Hence, in this research, four types of databased machine learning techniques, namely categorical boosting, extreme gradient boosting, group method of data handling (GMDH) and genetic programming (GP), were used for MMP estimation. To train the model, a dataset with more than 910 rows of data and multiple features, including molecular weight of light, CO<sub>2</sub> stream impurities, intermediate and heavy hydrocarbon components, as well as mole fractions of different elements, was fed to the models. All the models could accurately predict the MMP in various conditions with <i>R</i><sup>2</sup> of 0.89–0.99. Among all methods, XGBoost had the highest accuracy, and GP had the lowest. To shed some light on the models' performance, sensitivity analysis and feature importance were performed for the proposed models, and the most impactful features were identified in each model. During the final phase of the research, the model's effectiveness and applicability were compared to four well-established MMP correlations from the literature. The developed model in the current research performed substantially better (up to 35 %). The vast datasets used in this study resulted in models with a much broader range of applicability than most previous experimental and empirical methods.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Evaluation of explainable machine learning techniques for CO2 minimum miscibility pressure determination in a vast range of temperatures and compositions

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月21日

查看全文
  • Alexei Rozhenko,
  • Fahimeh Hadavimoghaddam,
  • Erfan Mohammadian,
  • Peyman Pourafshary

摘要

Precise estimation of minimum miscibility pressure (MMP) is advantageous for CO2-based EOR applications and CO2 capture, sequestration and utilization. However, experimental methods, including slim tube tests, are costly and time-consuming. In addition, conventional empirical correlations generally have a limited range of applicability. Hence, in this research, four types of databased machine learning techniques, namely categorical boosting, extreme gradient boosting, group method of data handling (GMDH) and genetic programming (GP), were used for MMP estimation. To train the model, a dataset with more than 910 rows of data and multiple features, including molecular weight of light, CO2 stream impurities, intermediate and heavy hydrocarbon components, as well as mole fractions of different elements, was fed to the models. All the models could accurately predict the MMP in various conditions with R2 of 0.89–0.99. Among all methods, XGBoost had the highest accuracy, and GP had the lowest. To shed some light on the models' performance, sensitivity analysis and feature importance were performed for the proposed models, and the most impactful features were identified in each model. During the final phase of the research, the model's effectiveness and applicability were compared to four well-established MMP correlations from the literature. The developed model in the current research performed substantially better (up to 35 %). The vast datasets used in this study resulted in models with a much broader range of applicability than most previous experimental and empirical methods.