Optimization of Cu-based Oxide Catalyst for Methanol Synthesis Using a Neural Network Trained by Design of Experiment

Abstract

Cu-Zn oxide catalyst for methanol synthesis was optimized using an activity map by neural network method. The catalyst composition was determined randomly and the training data were measured in a high-pressure HTS reactor using a 96-well microplate system. In this case, the data around the optimum were usually scattered with increasing parameters, despite 95 datasets causing low accuracy of the prediction. To obtain effective datasets for training, design of experiment was employed. After 18 datasets of catalyst composition-activity were designed and measured, a variety of neural networks were constructed. Each maximum was determined by the activity-envelope method and the best neural network was compared with that constructed from the 94 datasets. Design of experiment combined with the neural network was useful to determine the optimum catalyst composition.