(645e) Data Driven Discovery of Reaction Pathways for Understanding Catalytic Cracking of Supercritical Dodecane in the Presence of ZSM-5

Modeling catalytic chemistry with non-model molecules is an immense challenge due to the vast number of potential intermediates, products, and pathways. Group-type approaches developed using chemical intuition are often used to deal with this complexity by limiting the number of species tracked. Data-driven reaction pathway discovery enabled by machine learning provides an alternative to chemical intuition for analyzing complex catalytic reaction schemes. The objective of this work is to evaluate data-driven pathway discovery for an application of practical interest and in the absence of the thousands (or more) of data points generally needed for machine learning.

In this work we developed a data-driven method for reaction pathway discovery of supercritical dodecane cracking on ZSM-5 in the presence and absence of water. The experimental data set used consisted of 6 time points with concentrations of >20 species recorded at each time point. Dendrograms were used to identify potential groups for use in reaction pathway models. Data was interpolated using a cubic spline, a piece-wise linear model, and a linear model to represent the data fed to the dendrogram algorithm. All of these methods resulted in the identification of the same groups, indicating the robustness of the approach despite the limited size of the data set. Interestingly the data driven approach identified qualitatively different groups than previous work based on chemical intuition. The identified groups were then organized into a manifold of reaction pathway models using an automated procedure that fit rate constants for all chemically feasible combinations to the available data. The performance of these models was evaluated based on their ability to capture the data quantitatively and qualitatively. This work establishes a simple two-step method for data-directed discovery of complex reaction pathways that has the potential to be useful even for data sets of realistic size (i.e., <20 data points).