Data-Driven Discovery of Reaction Pathways for Modeling Catalytic Cracking of Hydrocarbons

Abstract

Modeling complex chemical reactions is a long-standing challenge in chemical reaction engineering. Group type analysis is commonly used for reducing the complexity of the reaction model while retaining sufficient chemical information for a particular application. This study evaluates a data-driven approach based on mathematical similarity as a new tool for identification of groups for use in group type reaction models. Data for dodecane cracking in the supercritical state over Zeolite Socony Mobil-5 (ZSM-5) was used as a test system. Mathematical similarity analysis of the raw data differentiated as many as five different groups. Adding synthetic data did not affect the groups identified by similarity analysis, indicating that the separation was not limited by the number of data points. Scaling or normalizing the data improved the separation of the similarity analysis. To test the data-identified groups, different types of reaction models were generated systematically, kinetic parameters regressed, and the resulting predictions compared with experimental data. The resulting reaction models consisted either of parallel or sequential reactions. As a general statement, the reaction models consisting of parallel reactions were more accurate than those consisting of sequential reactions. Additional tests showed that user defined groups could be added to those identified mathematically to improve the accuracy of predictions for target species without sacrificing overall accuracy. The similarity approach was applied to a data set consisting of catalytic dodecane cracking in the presence of water added to the reaction mixture (50 wt %). The proposed similarity analysis identified different groups in the presence and absence of water, indicating that the data-driven approach can be used to identify qualitative differences in the reaction pathways. It is therefore demonstrated that data-driven identification of group types represents a useful new tool for development of group type models.