Abstract
Development of a kinetic model is one of botlneck lead time in designing a chemical process. To shorten the lead time, a rata-driven predictive model synthesis rate was developed using machine leaning. In this study, a multi-layer perceptron (MLP) or support vector machine (SVM) was adapted to the machine learning model, which reaction rates from reaction condition e.g. temperature and partial pressure of substrates. In this study, the training dataset was synthesized by the Hougen-Watson model for ammonia synthesis, which is know as one of the most accurate kinetic model for the reaction. The hyperparameters of the above two machine learning model were optimized by grid search. It was found that the MLP with three hidde layers, 700 nodes per hidden layer and ReLU as the activation function indicated the highest accuracy, and the SVM with RBF kernel was competitive to the MLP model. From the simulation of kinetics by the machine learning models, it was found that the models underestimated the reaction rates for higher temperature. From this result, it was considered that preprocessing data based on temperature dependence of a chemical reaction is required to improve the accuracy of the models. Temperature dependence of a chemical reaction generally descrive as Arrhenius law: therefore taking inverse of the temperature and taking logarithm of the reaction rates in the input data were conducted. Preprocessing the input data based on the method, it was shown that accuracy of the machine learning models improved.
