Predicting the Shear Viscosity of Carbonated Aqueous Amine Solutions and Their Blends by Using an Artificial Neural Network Model

In
the present work, a neural network (NN) model based on quantitative
structure–viscosity relationship was implemented for predicting
the shear viscosity of CO2-loaded and CO2-free
aqueous amine solutions and their blends. A total of 1682 amine +
CO2 + water viscosity data sets for primary, secondary,
and tertiary amines and 220 data points for further accuracy examinations
were used. Molecular mechanic methods with CHARMM + CFF force fields
were utilized in order to optimize, simulate, and extract the required
molecular structure properties. Then, weighted nearest neighbor feature
selection algorithm was used for selecting the most influencing descriptors,
while cascade-forward NN (CFNN) model was applied for prediction purposes.
For generality examinations, CO2-loaded aqueous systems
of 3DMA1P(1) + EAE(2), MEA(1) + PZ(2), DMA2P(1) + MEA(2), DEAE(1)
+ PZ(2), and CO2 + pure water were used to find the solution
viscosities, and comparisons were made against experimentations, which
showed the quite robustness of the proposed model for the systemsm
which were completely unaware of the trained model. Comparison between
the values of average relative deviation of the NN model and the most
important semiempirical viscosity models showed that CFNN model outperforms
the other alternatives.

本研究构建了基于定量结构-粘度关系的神经网络（Neural Network, NN）模型，用于预测负载二氧化碳与无二氧化碳的胺水溶液及其混合体系的剪切粘度。本研究共纳入1682组针对伯胺、仲胺、叔胺的胺+二氧化碳+水体系粘度数据集，以及220个用于后续精度验证的数据点。研究采用结合CHARMM力场与CFF力场的分子力学方法，对分子结构进行优化、模拟并提取所需的分子结构属性。随后采用加权最近邻特征选择算法筛选出影响性最强的分子描述符，并应用级联前馈神经网络（Cascade-forward Neural Network, CFNN）模型开展粘度预测工作。为验证模型的泛化性能，本研究选取负载二氧化碳的3DMA1P(1)+EAE(2)、MEA(1)+PZ(2)、DMA2P(1)+MEA(2)、DEAE(1)+PZ(2)以及纯二氧化碳水溶液体系，计算其溶液粘度并与实验值进行对比。结果表明，所提模型对于训练集未覆盖的体系同样具备优异的鲁棒性。通过对比本神经网络模型与主流半经验粘度模型的平均相对偏差值，结果显示级联前馈神经网络模型的预测性能优于其他同类模型。