An empirical investigation of neural networks, evolution strategies, and evolutionary trained neural networks and their application to some chemical engineering problems

Abstract

Evolutionary algorithms and neural networks have been successfully used to solve difficult problems in various domains. Researchers and practitioners have applied them as single paradigms or in combination with each other. Here the utility of CI methods in Chemical Engineering is investigated. The performance of neural networks and evolutionary algorithms and combinations of them on real engineering problems is shown. An encoding of chemical compounds is proposed that allows the application of both paradigms and establishes a basis for comparisons. Solutions found by CI methods are presented that compare to the best physically motivated methods known so far and even outperform them in several ways. During the design process of chemical plants the knowledge how chemicals react with each other ist very important. For this reason there is a need for calculation methods which are able to predict thermodynamic properties. In this work, properties under consideration concern either pure components where the heat of vaporization has to be predicted or mixtures where the heat of mixing should be predicted.