A knowledge of various thermophysical (in particular transport) properties of ionic liquids (ILs) is crucial from the point of view of potential applications of these fluids in chemical and related industries. In this work, over 13 000 data points of temperature- and pressure-dependent viscosity of 1484 ILs were retrieved from more than 450 research papers published in the open literature in the last three decades. The data were critically revised and then used to develop and test a new model allowing in silico predictions of the viscosities of ILs on the basis of the chemical structures of their cations and anions. The model employs a two-layer feed-forward artificial neural network (FFANN) strategy to represent the relationship between the viscosity and the input variables: temperature, pressure, and group contributions (GCs). In total, the resulting GC-FFANN model employs 242 GC-type molecular descriptors that are capable of accurately representing the viscosity behavior of ILs composed of 901 distinct ions. The neural network training, validation, and testing processes, involving 90, 5, and 5% of the whole data pool, respectively, gave mean square errors of 0.0334, 0.0595, and 0.0603 log units, corresponding to squared correlation coefficients of 0.986, 0.973, and 0.972 and overall relative deviations at the level of 11.1, 13.8, and 14.7%, respectively. The results calculated in this work were shown be more accurate than those obtained with the best current GC model for viscosity of ILs described in the literature.
Read full abstract