(321p) Generalization of a Viscosity Model for Liquid Mixtures Using an Artificial Neural Network

The two binary interaction parameters of a previously developed deterministic viscosity model for liquid mixtures [Macías-Salinas et al., Fluid Phase Equilibria, 210, 319 (2003)] were generalized by means of a stochastic approach based on the use of artificial neural networks (ANNs). The final four-layered feed-forward ANN was trained by using a database containing 272 learning instances of the most representative binary mixtures (non-polar/non-polar, non-polar/polar, polar/polar and aqueous systems) at different conditions of temperature and pressure. After a detailed analysis, 15 relevant ANN descriptors were identified to adequately represent the relationship between the interaction parameters and the nature of the corresponding mixture. These descriptors included: the temperature, the pressure, two asymmetric ratios due to differences in molecular size and shape between the two compounds forming the mixture and 11 additional descriptors relative to the overall chemical structure of the mixture. The training of the ANN yielded root-mean-square errors between the interaction parameters obtained by fitting the viscosity model to experimental data and those calculated from the ANN of 0.963% for the learning set and 2.61% for the validation set. The experimental liquid viscosities of various binary mixtures were remarkably well represented by the deterministic model using interaction parameters predicted by the ANN at different temperatures and pressures thus confirming the suitability of the ANN-based approach as a generalization tool.