Artificial Neural Network Classification of Phase Equilibrium Methods – Part 2

S. Oreški, J. Zupan and P. Glavič

Abstract
A further study of the neural network application for predicting appropriate methods of phase equilibrium on the basis of known physical properties is presented. Kohonen neural networks are used to classify objects into none, one or more possible classes. The classes in the study represent possible methods of phase equilibrium. The trained neural network estimates the reliability of its predictions – the adequacy of individual methods of phase equilibrium for further efficient chemical process design and simulation. The analysis of the preliminary, less accurate results confirms the hypothesis to use Kohonen networks for classification of the classes as a correct one. Therefore, the Kohonen network architecture yielding the best separation of clusters was chosen for further analysis. It has been adapted and the training continued until the conflicting situations were resolved. Out of the several Kohonen networks trained the best one was analyzed. The maps of individual physical properties and the probability maps were obtained for each specific phase equilibrium. The correlation among maps is shown.

This work is licensed under a Creative Commons Attribution 4.0 International License

Keywords
physical properties, phase equilibrium, artificial neural networks, artificial neural network classification, learning procedure