A NEURAL-NETWORK MODEL FOR PREDICTION OF PHASE-EQUILIBRIA IN AQUEOUS 2-PHASE EXTRACTION

A model based on a feedforward back-propagation neural network was emp loyed to predict the phase equilibrium diagram of the aqueous two-phas e systems. The PEG/potassium phosphate/water system (pH 7) was selecte d as the model system to demonstrate the point of interest. A variety of molecular weights (MW) of PEG systems including PEG 600, 1500, 3400 , 8000, and 20 000 were considered for training the patterns in order to estimate the systems with PEG MW of 400 and 1000. After the optimal architecture of the network was investigated and finally determined, the extrapolated and interpolated simulations by this model exhibited an excellent agreement with experimental data. The characteristics of the phase diagram such as the binodal curve and tie lines were illustr ated in precision in all trials. The model can associate the dependenc e of PEG MW with the subtle shift of the corresponding phase diagrams over the test MW range. All the equilibrium data of the PEG/potassium phosphate systems with continuously variable PEG MW ranging from 20 00 0 to 400 could be predicted by the model. The results indicated the ap plicability of the neural network model as a design-oriented technique for optimization of extraction condition. The neural network model sh ould be a potent means to deal with more complex models such as PEG/de xtran systems and partition of proteins in aqueous two-phase systems.