(306f) Machine Learning Enhancements to Transient Kinetic Models

Transient kinetic models are a powerful way to study, simulate, and forecast the kinetics of heterogeneously catalyzed reactions. However, whether through employing the mean-field assumption within microkinetic models (MKMs) or invoking the quasi-steady state approximation on a proposed reaction network, these models frequently rely on numerous assumptions to glean insights conducive to interpreting experimental kinetic observations. In many cases, these assumptions prove inadequate, failing to account for factors such as lateral adsorbate-adsorbate interactions or dynamically changing catalyst environments.

This study focuses on the development of an innovative transient hybrid kinetic model with the capability to accurately simulate and predict responses of heterogeneously catalyzed reactions to sudden changes in operating conditions, such as partial pressure. Our approach merges the interpretability of transient kinetic models for studying heterogeneous catalytic reactions with the precision and flexibility of data-driven models. The presented approach employs sensitivity-based hybrid modeling with deep neural networks [1]. Here, learning is conducted on critical parameters identified through sensitivity analysis within the kinetic model. The resulting hybrid model showcases superior accuracy and robustness compared to conventional kinetic models. This integration aims to provide a more accurate representation of systems that traverse multiple kinetic regimes with varying surface coverages.

Benefits of the hybrid transient kinetic model are demonstrated for selective propylene oxidation over an industrial mixed metal oxide catalyst under lean/rich cyclic feed compositions. By combining the strengths of both physical and machine-learned approaches, the intricacies of heterogeneous catalyzed reactions in dynamic conditions are better captured and allow for further optimization of the oscillating feed conditions.

References:

[1] Shah, P., Sheriff, M. Z., Bangi, M. S., Kravaris, C., Kwon, J. S.-I., Botre, C., & Hirota, J. (2022). Deep Neural Network-based hybrid modeling and experimental validation for an industry-scale fermentation process: Identification of time-varying dependencies among parameters. Chemical Engineering Journal, 441, 135643.