Artificial neural networks for the solution of the phase stability problem

The prediction of the thermodynamic properties of multiphase systems is complex, because, besides equilibrium calculations, it involves determination of the number and nature of the phases present in the system (phase stability tests). For a system exhibiting a heterogeneous azeotrope, for example, the problem is to develop methods that can tell whether, for a given overall composition, the system lies inside or outside the binodal surface (two liquid phases in equilibrium or a single stable liquid phase). In this work, the application of artificial neural networks (ANNs) for the solution of the phase stability problem, a classification problem, is proposed. The input-output patterns, required for training the networks, were obtained computationally for the range of temperatures that covers liquid-liquid equilibrium, vapor-liquid-liquid equilibrium, vapor-liquid equilibrium and homogenous liquid and vapor. Hence, the ANN must be able to decide between these five possible regions. Two types of ANNs were tested: feedforward neural networks (FNNs) and probabilistic neural networks (PNNs). The results indicate that each kind of ANN is better for different conditions, the developed ANNs were able to predict correctly the type of equilibrium in more than 99.9% of the cases. (c) 2006 Elsevier B.V. All rights reserved.