Abstract The present work aims at applying Machine Learning approaches to predict CO2 viscosity at different thermodynamical conditions. Various data-driven techniques including multilayer perceptron (MLP), gene expression programming (GEP) and group method of data handling (GMDH) were implemented using 1124 experimental points covering temperature from 220 to 673 K and pressure from 0.1 to 7960 MPa. Viscosity was modelled as function of temperature and density measured at the stated conditions. Four backpropagation-based techniques were considered in the MLP training phase; Levenberg-Marquardt (LM), bayesian regularization (BR), scaled conjugate gradient (SCG) and resilient backpropagation (RB). MLP-LM was the most fit of the proposed models with an overall root mean square error (RMSE) of 0.0012 mPa s and coefficient of determination (R2) of 0.9999. A comparison showed that our MLP-LM model outperformed the best preexisting Machine Learning CO2 viscosity models, and that our GEP correlation was superior to preexisting explicit correlations.

中文翻译：

---

在高温高压条件下模拟 CO2 的粘度

摘要 目前的工作旨在应用机器学习方法来预测不同热力学条件下的 CO2 粘度。各种数据驱动技术包括多层感知器 (MLP)、基因表达编程 (GEP) 和数据处理组方法 (GMDH)，使用 1124 个实验点实现，覆盖温度范围为 220 至 673 K，压力范围为 0.1 至 7960 MPa。粘度被建模为在规定条件下测量的温度和密度的函数。在 MLP 训练阶段考虑了四种基于反向传播的技术；Levenberg-Marquardt (LM)、贝叶斯正则化 (BR)、缩放共轭梯度 (SCG) 和弹性反向传播 (RB)。MLP-LM 是所提出模型中最适合的，其总体均方根误差 (RMSE) 为 0.0012 mPa s，决定系数 (R2) 为 0。