Optimization of Fructose Dehydration over Zeolite Catalysts Using Machine Learning

The dehydration of biomass-derived fructose is an important reaction to produce the platform chemical, 5-hydroxymethyl furfural (HMF), which could be further processed into biofuels and other green derivatives. This reaction has several side products, making the identification of selective catalysts challenging. Therefore, it has been extensively studied in experiments over various zeolite catalysts. However, no comprehensive analysis has been done to understand the relationships between the yield and rate of the reaction and the properties of different zeolite frameworks.

In this work, we developed a machine learning model to predict the yield and the rate constant of fructose dehydration to HMF and identify potentially better zeolites for this reaction. Experimental data from 25 research papers were collected. 12 features representing the catalyst and solvent properties, and reaction conditions were selected to train the model. Various regression algorithms were used to predict the two labels separately. A 5-fold cross-validation error scaled by the range of each label is used as an evaluation metric.

Random Forest regression model was found to have the lowest error. When trained on the dataset, the model had cross-validation errors of around 14 % and 20% for the prediction of yield and rate constant, respectively. Descriptors such as the TPD Ammonia, Si/Al ratio and BET surface area of the catalyst, as well as total dipole moment of the solvents were found to be the most important features to making the predictions.

The results showed that the conditions leading to a higher rate constant don’t correspond to a higher yield because they promote side reactions as well. The model was used to predict the performance of potential zeolites in the IZA database. Zeolites such as Metal-Organic Zeolite (MOZ) and Linde Type L (LTL) are found to show better performances compared to the current zeolites used for this reaction. We propose that transfer learning can be combined with this model to predict the performance of reactions with similar mechanisms such as Xylose dehydration.