Prediction of vapor-liquid equilibrium via neural networks

Abstract

In recent years the need for reliable experimental information for the design and optimisation of industrial processes has been in increasing demand. In this context, neural networks were applied for the prediction of phase equilibrium composition.

A feedforward regression neural network (ANN) with supervised learning was designed. The training was performed using experimental phase-equilibrium data of polydimethylsiloxane + CO₂. The ANN prediction was compared with Sanchez-Lacombe equation’s results. The theoretical values obtained by the latter only consider experimentally measured data, whereas the ANN also contemplates data from the literature.

The design of an ANN with 2, 1 and 30 neurons in its input, output and hidden layers was the best choice to predict the equilibrium CO₂ composition for temperatures between 30 and 80°C. With the increase of the number of neurons in the hidden layer, the prediction of the mass fraction for CO₂ at equilibrium becomes less and less accurate, resulting in behaviours that differ greatly from the experimental points. Therefore, increasing the problem size does not help the correct ANN training.