(661d) Using Neural Networks to Interpret Transient Kinetic Data

The kinetics of catalytic reactions are governed by the reaction mechanism and rate parameters of elementary steps. Approaches to kinetic analysis can be broadly classified in to "bottom up" and "top down", where bottom-up approaches seek to establish rate constants and reaction mechanisms from atomic-scale models, and top-down approaches utilize measured kinetic data. One challenge in top-down approaches is the issue of over-fitting, since kinetic models contain many parameters, but steady-state kinetic data is typically controlled by only a few rate-limiting steps, meaning that multiple models are capable of fitting measured data. The use of transient kinetic and operando spectroscopic data can help overcome this limitation by probing more elementary steps over a wider range of reaction conditions. However, transient and operando techniques pose challenges for fitting kinetic models, including measurements at varying time resolutions, uncertain calibration factors, and transport considerations. In addition, the underlying kinetic models can be challenging to fit to transient kinetic data owing to numerical stiffness, noisy data, and algebraic constraints. This talk will provide an overview of using neural networks as a basis set for solving the differential algebraic equations that govern mean-field micro kinetic models, and illustrate how this enables more robust fitting of rate parameters to transient kinetic data. Case studies on synthetic data will be used to evaluate the accuracy and limitations of the approach in handling stiff systems and noisy data, and the results of applying neural networks to data collected from a temporal analysis of products (TAP) reactor will be compared to a physics-based fitting technique. The results indicate that these "kinetics-informed neural networks" provide a robust and flexible route to fitting interpretable kinetic models to transient kinetic data, but reveal a number of challenges that remain for extracting reliable reaction mechanisms and intrinsic rate parameters from measured kinetic data.