Development of Both Linear and Nonlinear Methods To Predict the Liquid Viscosity at 20 °C of Organic Compounds

Abstract

Experimental values for the liquid viscosity (η) at 20 °C ranging from 0.164 mPa‚s (trans-2-pentene) to 1490 mPa‚s (glycerol) have been collected from literature for 361 organic compounds containing C, H, N, O, S, and all halogens. Multiple linear regression (MLR) and two-layer neural network (NN) modeling (one hidden layer) with back-propagation have been applied to derive prediction methods for log η using nine descriptors as input. The analysis includes different partitionings of the data set into training and prediction sets and different numbers of hidden-layer neurons of the neural networks. For the linear and nonlinear models derived from a training set of 237 compounds, squared correlation coefficients of 0.92 and 0.93 as well as root-mean-square errors of 0.17 and 0.16 log units were achieved for a prediction set of 124 compounds, reflecting a reasonable accuracy for a wide range of chemical structures and viscosity values. However, only the NN model was capable of successfully treating glycerol with the maximum viscosity value, which was not possible with the MLR approach and with any other existing estimation scheme.

Application of a General Computer Algorithm Based on the Group-Additivity Method for the Calculation of Two Molecular Descriptors at Both Ends of Dilution: Liquid Viscosity and Activity Coefficient in Water at Infinite Dilution

“…Later, a quantitative structure-property relationship (QSPR) study [6], founded on 361 compounds and using five molecular structural descriptors including electrostatic and quantum chemical properties, resulted in a correlation coefficient of 0.854 and a standard error of 0.22. The multiple linear regression and artificial neural network (ANN) back-propagation methods, outlined in [4], based on 361 compounds and nine physical and structural descriptors, yielded a correlation coefficient of 0.92 and 0.93, respectively, and corresponding standard errors of 0.17 and 0.16 units. In a later paper [5], the same authors presented slightly better results with a set of 440 compounds, using the same ANN approach and input descriptors, which produced correlation coefficients for the training, validation and test sets of 0.956, 0.932 and 0.884, respectively, with corresponding standard errors of 0.122, 0.134 and 0.148 units.…”

“…Earlier attempts to predict the liquid viscosity coefficient of organic compounds have been developed on a statistical mechanics model based on the square well intermolecular potential [3], or have been carried out applying multiple linear regression and artificial neural network modelling methods using a limited number of descriptors as input [4,5], or are based on a quantitative structure-property relationship (QSPR) approach using a five-descriptor equation [6], or use a combination of partial least-square and QSPR technique starting with 18 mostly experimental parameters, finally ending with a model with nine descriptors [7]. …”

Data and QSPR study for viscosity of imidazolium-based ionic liquids

“…Earlier, QSPR studies on the viscosity of conventional moleculartype organic compounds have been performed, and the results are summarized in Table 5 [70][71][72][73][74][75][76]. As indicated in Table 5, for molecular-type organic compounds, the most important factor that affects the viscosity is generally the intermolecular hydrogenbonding interactions; the dipole moment (i.e., the polarizability mostly arising from polar atoms), which mainly reflects the intermolecular van der Waals interactions, gives remarkable contributions to the viscosity; the steric bulk, shape, and size of a molecule also have some effects on the viscosity.…”

Artificial Neural Network in Drug Delivery and Pharmaceutical Research

“…The average absolute error for the test set in flash point prediction was 11.9 ºC with a 26-36-2 configuration. Suzuki et al developed an ANN model for predicting liquid viscosity at a standard temperature of 20 ºC [95] and subsequently, a temperaturedependent model [96]. The best model showed a root mean square error of 0.148 log units for the test set of 79 compounds and 133 data points.…”