Prediction Of Phase Equilibrium Properties In Complicated Macromolecular Systems By Neural Networks

The researches of the phase equilibrium properties in complicated macromolecular systems provide the bases for the simulation and optimization in chemical and process engineering. It is essential to understand the thermodynamic properties in these systems and can be used for guiding the development and optimization for the chemical engineering. The results can be used as a support for the experimental research in asymmetry polymer membrane preparation as well as the simulation of the growth process and the reactor for crystallization.Since the traditional thermodynamic model and polynomial correction method can not be well used to simulate and predict the thermodynamic properties in complicated macromolecular systems, the artificial neural networks (ANNs) have been proven to be an effective way to deal with these problems for their inherent ability to map highly non-linear. In this paper, the ANNs were used to predict the phase equilibrium properties both in polymer membrane formation system and protein crystallization system. Moreover, the traditional error back-propagation method, which was used as the learn algorithm in neural networks, had been improved, and the performances of these improved algorithms had been validated.The main works of this paper could be summarized as follows:1 The ANNs were used to train the database of cloud points in water-DMAc-PSf system, and the well trained networks were used to predict the binodal curve at other temperatures. The results could compensate for the lack of the thermodynamic data, and could be used for guiding the experimental researches in preparation of the asymmetry polymer membranes.2 An artificial neural network was used for simulating and predicting the solubility of lysozyme in bio-macromolecular lysozyme-NaCl-H2O system. The impact factors that affect the protein solubility were analysed and discussed. Moreover, the predicted performances were compared to the traditional thermodynamic models. It was found that a properly selected and trained neural network could give the favorable results for prediction of the lysozyme solubility in lysozyme-NaCl-H20 system, and that the prediction accuracy was improved compared to the traditional thermodynamic models.3 Three modifications which based on the momentum strategy, self-adaptive learning rate coefficient and the modified Levenberg-Marquardt algorithm were used to improve theperformances and overcome the deficiency of the traditional error back-propagation algorithm. The results show that the ANNs based on the momentum and self-adaptive learning rate coefficient can improve the prediction accuracy and accelerate the learning rate, Whereas, the modified Levenberg-Marquardt algorithm shows better performance than the two modified algorithm. In this paper, the back-propagation (BP) networks, which were based on the modified Levenberg-Marquardt algorithm, were used to predict the phase equilibrium properties in complicated macromolecular systems. It could be found that the BP network shows good performances both in the process of training and predicting.4 Introducing the genetic algorithms (GAs) that borrow operations and themes from drawing evolution, a new algorithm that combined the hybrid genetic algorithm with modified Levenberg-Marquardt algorithm (HGALM) was presented. A genetic algorithm possesses the capability to find a global optimum was integrated with the modified Levenberg-Marquardt algorithm which gives a good ability to find the local accuracy solution in the HGALM. The perfoemance of the HGALM was validated with two examples. It can be found that the proposed algorithm can improve the accuracy and decrease the time depletion comparing to the traditional EBP algorithm.