Bayesian data science and physical principles for kinetic modeling of reaction networks on heterogeneous catalysts
Date
2022
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
University of Delaware
Abstract
Kinetic modeling could play an integral role in determining optimal operating conditions and in discovering physical insights to enable reactor and catalyst improvements. While tremendous progress in modeling has been achieved, significant gaps remain. As we move into renewable and recyclable feedstocks, products and reaction networks are often unknown, kinetic parameters are hard to predict from first principles due to reaction complexity and lack of information on catalyst structure and active sites, and uncertainty pervades parameters, models, and networks. Developing systematic tools to overcome these barriers is essential. ☐ In the first section of this dissertation, we propose a Bayesian approach to kinetic modeling. To simulate chemical reaction networks, one must first discern which reactions occur. We create a methodology for reaction network identification using graph theory to inform the design of experiments. The resulting data is analyzed with Boolean reaction circuits and probabilistic graphical models. These rate-equation-agnostic methods provide Bayesian statistical confidence in the key reactions. In an active learning framework, we repeat this design of experiments to determine the next most informative experiment, and the workflow continues until the reaction network is fully confirmed. We illustrate the methodology on the ethane dehydrogenation reaction network, the CO2-assisted ethane dehydrogenation reaction network, and a reaction network of cross-ketonization of furoic acid and lauric acid to produce precursors for detergents. The last example uses experimental data collected from a parallel study conducted during this investigation. ☐ In the second section, we derive the Modified Energy Span Analysis (MESA) framework that integrates and ultimately unifies microkinetic modeling concepts and the energy span model. MESA rationalizes how individual intermediate and transition state energies contribute to kinetic observables such as turnover frequency, coverages, apparent activation energies, and reaction orders. This model provides mechanistic insights using experimentally measured kinetic metrics. MESA preserves accuracy while addressing the Bayesian concern of containing too many parameters. The framework has been extended to complex reaction networks of parallel reactions, including selectivity. We illustrate MESA’s applicability by evaluating four published models: an ethylene reaction network, a CO oxidation network, and two networks for 2-methylfuran acylation. ☐ In the last section of this dissertation, we propose a new methodology to quantify the uncertainty of kinetic models using propagative Bayesian inference. Demonstrating this methodology on a CO2-assisted ethane dehydrogenation reaction system, we uncover the effects of water on the surface chemistry rate and stability. We develop a Python software package, called the Chemical Kinetic Bayesian Inference Toolbox (CKBIT), which employs statistical techniques with minimal user coding required to estimate kinetic parameters from experimental data. We illustrate CKBIT upon laboratory measurements of ethane dehydrogenation rate data and simulated measurements of a 5-hydroxymethylfurfural reaction network.
Description
Keywords
Kinetic modeling, Pervades parameters, Feedstocks, Bayesian Inference Toolbox, Learning framework