ChE PhD Dissertation Defense: Sevy Harris

Name:
Sevy Harris

Title:
Automated Methods for Improving Microkinetic Models through Uncertainty Quantification and Sensitivity Analysis

Date:
07/27/2026

Time:
11:00:00 AM

Committee Members:
Prof. Richard West (Advisor)
Prof. Steve Lustig
Prof. Qing Zhao
Prof. Franklin Goldsmith

Location:
222 Hayden Hall

Abstract:
Microkinetic modeling is an invaluable technique for investigating new fuels in pursuit of cleaner combustion devices. These models, specifying a network of reacting species and elementary reactions, use physically meaningful parameters to predict how systems will react over a wide range of conditions. Automated mechanism generators, like Reaction Mechanism Generator (RMG) can systematically build microkinetic models by reacting species together according to predefined templates. These are necessary to handle the complexity of combustion mechanisms, which can have hundreds of intermediate species and thousands of reactions, but the first result is rarely accurate enough to be useful. Human intervention is generally required to analyze the model, identify important parameters, and then improve their values with experimental results or quantum chemistry calculations. This work introduces a highly automated workflow for automatically building and improving mechanisms. It uses automated uncertainty and sensitivity analysis to rank the most important parameters to improve upon, and then uses quantum chemistry calculations to compute the most important thermokinetic parameters. The workflow is applied to propane and butane models for ignition delay, demonstrating the strengths and weaknesses of this method of automated model improvement.

The automated uncertainty analysis portion of the improvement workflow was implemented by earlier work of Gao, Liu, and Green. They assigned uncertainties to every input parameter by looking up the source RMG used to estimate the value (e.g. a trusted library entry from the literature, or a less trusted rate rule estimation), and then applying a default uncertainty for that source. They accounted for certain parameter correlations by keeping track of parameters estimated with the same source. However, the previous implementation did not allow for the underlying sources themselves to be correlated, and some new additions to the RMG database include correlation data. The previous implementation also applied the same default uncertainty to all the underlying data in the RMG database, even though certain values are known to be higher quality than others. This work expands and improves upon the existing RMG uncertainty framework to incorporate correlated source data and to apply more specific uncertainty assignments based on information already available in the database. It also extends the uncertainty suite to handle surface-phase mechanisms and enables the easy export of uncertainty covariance matrices so that RMG users can conduct global uncertainty analysis entirely outside the RMG framework. A case study of propane in a jet-stirred reactor shows how these extensions enable global Monte Carlo and Sobol uncertainty analyses, with error bars that are an improvement on previous estimates.

Ammonia is being considered as a possible alternate fuel to traditional hydrocarbons, but when it reacts inside a stainless steel reactor, the nitrogen will, under certain conditions, form nitrides that degrade the steel and shorten its lifespan. A multiscale model of the nitridation of stainless steel was built in collaboration with Mitsubishi Heavy Industries (MHI) to predict the conditions under which nitridation occurs. In the first stage of the multiscale model, the thermodynamics of the reacting surface species are computed using machine learning models. In the second stage, those thermodynamic parameters are used to build a microkinetic model of species reacting on the steel surface. In the third stage, surface concentrations from the second stage inform a transport model that computes how nitrogen and oxygen diffuse into the steel bulk. The end result is a model which translates microscale phenomena into macroscale observables that can be compared to experiments by MHI.

Sevy Harris is a PhD candidate in Chemical Engineering at Northeastern University. She received her BS in Electrical Engineering in 2014 from Ohio State and her MS in Electrical Engineering from Stanford in 2016. She worked at Microsoft for three years, designing flexible printed circuits and later writing test software to characterize a VR depth camera. She joined the Computational Modeling group at Northeastern in 2020 to study kinetic modeling and how these models can be used to investigate alternate fuels for cleaner combustion.