Machine learning model for design of SMB processes and its application
to separate rebaudioside A and stevioside
Abstract
Several machine learning algorithms were used to simulate the simulated
moving bed (SMB) process, with the sugar separation of rebaudioside A
and stevioside and enantioseparation of 1,1’-bi-2-naphthol racemate as
case studies. It was found the random forest (RF) model and the deep
neural network (DNN) model give satisfactory accuracy with MAEs lower
than 0.19% (RF) and 0.08% (DNN). Then these two models were used to
optimize the operation conditions for maximizing the feed flowrate under
specific purity requirements. The RF model failed to give a set of
operation conditions better than the training dataset. But the DNN model
gave flowrates about 10% higher than the highest values in the training
datasets, for both sugar separation and enantioseparation systems.
Finally the optimized operation conditions for sugar separation were
verified experimentally, with the final purities of rebaudioside A and
stevioside being 99.2% and 98.8% respectively.