Modeling and optimizing the synthesis of isopropyl acetate over niobium pentoxide using experimental design methodology coupled with artificial neural network

Abstract

Experimental design methodology (EDM) was coupled with an artificial neural network (ANN) to predict and optimize the synthesis of isopropyl acetate over Nb2O5. The coupled model was compared to the empirical correlation of EDM, which correlates the conversion of acetic acid in isopropyl acetate (output variable) to reaction temperature, catalyst loading, and molar ratio of acetic acid/2-propanol (input variables). A key advantage of ANN is adding the reaction time to the input variables. The optimal ANN architecture was [5 5 5] with Bayesian regularization backpropagation as the network training algorithm. The experimental conversion has shown a good correlation with the predicted conversion for both training and test datasets, justifying the use of ANN. Statistical metrics showed that EDM coupled with ANN can better predict the conversion of acetic acid when compared to the empirical correlation (R2 = 0.9992 and SSE = 0.0285). Shapley values were applied in the ANN interpretability and ranked the input variables in descending order of importance: reaction temperature reaction time molar ratio of acetic acid/2-propanol catalyst loading.