Prediction of phase equilibrium properties for complicated macromolecular systems by HGALM neural networks

Traditional error back propagation is a widely used training algorithm for feed forward neural networks (FFNNs). However, it generally encounters two problems of slow learning rate and relative low accuracy. In this work, a hybrid genetic algorithm combined with modified Levenberg–Marquardt algorithm (HGALM) was proposed for training FFNNs to improve the accuracy and decrease the time depletion comparing to the traditional EBP algorithm. The FFNNs based on HGALM were used to predict the binodal curve of water–DMAc–PSf system and protein solubility in lysozyme–NaCl–H2O system. The results would be used for guiding experimental researches in preparation of asymmetry polymer membrane and optimization of protein crystal process.