Application Of Artificial Neural Networks In Prediction Of Vapour Liquid Equilibrium Data

The associative property of artificial neural networks (ANNs) and their inherent ability to “learn” and “recognize” highly non-linear and complex relationships finds them ideally suited to a wide range of applications in chemical engineering. The present paper deals with the potential applications of ANNs in thermodynamics – particularly, the prediction/ estimation of vapour-liquid equilibrium (VLE) data. The prediction of VLE data by conventional thermodynamic methods is tedious and requires determination of “constants” which is arbitrary in many ways. Also, the use of conventional thermodynamics for predicting VLE data for highly polar substances introduces a large number of inaccuracies. The possibility of applying ANNs for VLE data prediction/ estimation has been explored using the back propagation algorithm Application of ANNs to the VLE predictions of NH3 – H2O and CH4 – C2H6 system is investigated. The results of neural equation of state (NEOS) are compared with popular thermodynamic approaches. The inputs to the net consist of Temperature and Pressure. Liquid and vapour phase compositions are obtained as outputs. These outputs are compared with the predictions obtained by using Peng Robinson equation of state and Wilson activity coefficients. For NH3 – H2O system vapour phase compositions are well predicted by all three approaches but thermodynamic approaches are unable to predict liquid phase compositions. ANNs or NEOS gives good results. For CH4 – C2H6 system, both Peng Robinson EOS and NEOS give good results. From the present work it is concluded that though for simple systems ANNs do not offer additional advantage, they are certainly superior when complex and polar systems are encountered. An heuristic approach to reduce the trial and error process for selecting the “optimum” net architecture is discussed. I TRODUCTIO Man has been fascinated by the capabilities of human brain and has tried to make a computer mimic the way the human brain sorts through the information. ‘Neurocomputing’ is such an attempt, to understand and simulate its functioning. It has been touted as the first known alternative to the programming paradigm that has dominated computing for the past fifty years. Artificial Neural Network (ANN) can thus be cited as ‘algorithmic equivalent’ of the human learning process and information processing scheme at a modest scale. They are pattern recognition architecture which can identify patterns between complex sets of input and output data. These patterns are then used to predict outcomes for fresh inputs. They do not require the specification of correlations which govern process, but are trained on real life data. The chief advantage of ANNs lies in the fact that ANN uses a generic model which covers a wide class of problems. It does not require a fundamental understanding of the process or phenomena being studied and can handle complex and non linear models. Thus they are gaining a rapid interest within engineering, medical, financial and various other fields. They have thus made strong advances in area of continuous speech recognition, classification of noisy data, market forecasting, process modelling, fault detection and control. In present work, application of neurocomputing for estimating VLE data has been explored. Conventional thermodynamic techniques for VLE data estimation of mixtures are tedious and have a certain amount of empiricism by way of determining mixture “constants” using arbitrary mixing rules. ANNs, on the other hand, help such predictions and eliminate the need for determining these constants by finding the functional relationship all at once. ANNs also offer the potential to overcome the limitations of existing equations-ofstate (EOS) in determining VLE data for highly polar systems. The ammonia-water (NH3 H2O) and the methane-ethane (CH4 C2H6 ) systems were studied and the results are presented. The methodology used and the advantages and the limitations of this approach have also been discussed. Proceedings 25th European Conference on Modelling and Simulation ©ECMS Tadeusz Burczynski, Joanna Kolodziej Aleksander Byrski, Marco Carvalho (Editors) ISBN: 978-0-9564944-2-9 / ISBN: 978-0-9564944-3-6 (CD) Various types of Nets are formed by combination of different algorithms and activation functions. The Multilayer Perceptron (MLP) model is capable of mapping data that are non linear and complex. Thus, an MLP model using Back Propagation Algorithm (BPA) based on sigmoidal activation function has been found to be most suited for chemical engineering applications and is used in present work.