Misleading Results on the Use of Artificial Neural Networks for Correlating and Predicting Properties of Fluids. A Case on the Solubility of Refrigerant R-32 in Ionic Liquids

The ability and limitations of Artificial Neural Networks (ANN) for correlating and predicting properties of fluids are discussed. Common mistakes found in applications of ANN to fluid property estimation presented in the literature are analyzed and a useful and appropriate way of presenting results of these applications is proposed. As a study case the use of ANN for correlating and predicting the solubility of difluoromethane (R-32, C2H2F2), in seventeen ionic liquids is performed and discussed. The dependent variable is the solubility of R-32 and the independent variables are the temperature (T) and pressure (P), in addition to properties that identify each Ionic Liquid, such as critical temperature (Tc), the critical pressure (Pc), the mass connectivity index (λ), the acentric factor (ω), the mass of the cation (M+) and the mass of the anion (M−). Six cases, combining these independent variables in different forms were analyzed. The study shows that although some cases of the proposed artificial neural networks model give good results for training and testing, none of them gives acceptable results for prediction. Finally, some indications and suggestions are given to correctly apply ANN and to present the results in the literature.