Using a machine learning model for the optimal design of simulated moving bed processes and its application to separate rebaudioside A and stevioside

Abstract

BACKGROUND Solving a mechanistic model of simulated moving bed (SMB) is extremely time consuming, therefore, it is inefficient for online control using the optimization method based on such a mechanistic model. To this end, the machine learning model can be regarded as a potential substitution to accelerate the optimization process. RESULTS Several machine learning algorithms were applied to predict the product purities under various operation conditions using two SMB processes as case studies: i.e., a sugar separation of rebaudioside A and stevioside, as well as an enantioseparation of 1,1′‐bi‐2‐naphthol racemate. The results indicate that the random forest (RF) model and the deep neural network (DNN) model provide a satisfactory accuracy with a mean absolute error (MAE) lower than 0.19% (RF) and 0.08% (DNN), respectively. During the optimization process to maximize the feed flowrate under specific purity requirements, among the two selected models, DNN model showed a better generalization ability than RF model and gave a feed flowrate 10% higher than the highest value in the training dataset, which was consistent with the result obtained by using the mechanistic model. The optimized operation conditions for sugar separation were verified experimentally and the achieved purities of rebaudioside A and stevioside were 99.2% and 98.8%, respectively. CONCLUSION The DNN model was successfully used to substitute the mechanistic model to reach a rapid optimization of SMB processes with an improved efficiency of 103 ~ 104 times. © 2021 Society of Chemical Industry (SCI).