A newly developed tridimensional neural network for prediction of the phase equilibria of six aqueous two-phase systems

Abstract

A newly developed tridimensional neural network (TRDNN) has been utilized to represent the phase equilibria of six polyethylene glycol (PEG)-inorganic salt aqueous two-phase systems (ATPSs). 18 data sets totaling 108 experimental data in the temperature range (298.2–318.2K) were categorized into training, test and validation sets in order to teach the model about the input-output relationships and validate its predictive ability. The optimal configuration of the model was found to be {5, [3,4,5], 3} and the system error for the training process was determined as 0.0055. Results indicate that the TRDNN model has better prediction performance as compared to the two-dimensional model. The standard deviations corresponding to three data sets for the TRDNN model were 0.0057, 0.0068 and 0.0055, while those for the two-dimensional model were 0.0065, 0.0078 and 0.0062, respectively. Moreover, it incorporates the molecular weight of polymer, salt type and temperature in one model and can reflect the effects of these factors on the phase behavior of these ATPSs correctly.