Prediction of Dynamic Viscosity and Sensitivity Study of Potassium Amino-Acid Salt Aqueous Solutions by an Artificial Neural Network According to the Structure

Abstract

Aqueous solutions of potassium amino acid salts show promise for capturing carbon dioxide. Accurately predicting their viscosity is fundamental in the design of new processes. Indeed, high viscosity leads to low mass transfer kinetics and significant pressure drops. The higher the viscosity, the larger the contactor's size, and accurate correlation can be useful for contactor design. Moreover, knowing the influence of different groups of molecules on viscosity can help select the best molecule. Nowadays, the artificial neural network (ANN) opens up new possibilities. The main purposes of this study are to build an ANN to model the viscosity of nine aqueous solutions of different potassium amino acid salts and to determine the influence of functionalized groups on dynamic viscosity. This network was built with 330 data points at several temperatures (288.15−353.15 K) and concentrations (0.25 mol/kg solvent to 9.4 mol/kg) from three different research teams. Dynamic viscosity is correlated to the concentrations, temperature, and structure of the anions, with an average absolute relative deviation (AARD) of 1.42%. Sensitivity analysis demonstrates that dynamic viscosity increases with the length of the anions and the concentrations and decreases with the temperature. Furthermore, carboxyl groups (COO − ) and phenyl groups increase dynamic viscosity (sensitivity influence COO − < phenyl). Measurements of dynamic viscosity were conducted for the valinate potassium aqueous solutions, an amino-acid salt not used in the ANN establishment. The AARD between the predicted values and the experimental results is 9.03% for the valinate potassium aqueous solutions.

Intelligent Soft Measurement of Catalyst Activity during a Polyethylene Reaction Based on Mechanistic Modeling and Deep Learning

“…Improving the accuracy of the training algorithm for this network can reduce the mean square error and make it as small as possible . The result output of MLP is shown in eq . γ j k = F k true( prefix∑ i = 1 N k − 1 w i j k γ i false( k − 1 false) + β j k true) where denotes the contribution value of neuron j in layer K and the bias weight of neuron j in layer K are γ jk and β jk , respectively, and W ijk denotes the connection weight …”

“…where denotes the contribution value of neuron j in layer K and the bias weight of neuron j in layer K are γ jk and β jk , respectively, and W ijk denotes the connection weight. 31 The structure of MLP neural network is shown in Figure 8. To ensure the accuracy of the ANN results, a trusted database is created.…”