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

Surface Tension of Liquid Organic Acids: An Artificial Neural Network Model

“…As an alternative to the previous methods, some authors have used artificial neural networks (ANNs) for the study of the surface tension and other thermophysical properties [ 16 , 30 , 31 , 32 , 33 , 34 , 35 , 36 , 37 , 38 ]. ANNs take their inspiration from biological neural networks, with the mathematical model of a single neuron and the way in which neurons are interconnected (i.e., the “architecture”), leading to a framework for different machine learning algorithms.…”

“…Very recently, Faúndez et al [ 38 ] reviewed the work on ANN models in the correlation and prediction of liquid properties, pointing out their capabilities and limitations. For the advantages, they highlighted their ability to work with incomplete information, as well as their capacity to learn and relate variables (taking care that sometimes the models mostly memorize and lose any predictive capability).…”

“…In addition, a non-linear activation function establishes the relationship between the input and output of a given neuron. The sigmoid function is used in this work, as commonly done in feed-forward networks [ 16 , 21 , 34 , 35 , 38 ].…”

“…Here, the programmer’s experience and ability are essential. Thus, it is necessary to try different architectures and analyze the results carefully [ 38 ]. It is possible to find that the model is over-fitted leading to large unacceptable errors.…”

“…It is possible to find that the model is over-fitted leading to large unacceptable errors. This can be avoided by dividing the dataset into three subsets: Training, test, and prediction [ 32 , 33 , 38 ]. While the training set is used to establish the main parameters of the model (bias, weights, etc.…”