Student Name: Anant Girdhar

Advisor: jeff.jagoda@aerospace.gatech.edu

Milestone: PhD Thesis Final Examination (Defense)

Degree Program: Aerospace Engineering

Title: A Process-Based Interpretable Data-Driven Framework for Local Chemical Kinetic Model Comparison

Abstract: Simulation of reacting flow in practical combustors require reduced chemical kinetic models in order to remain computationally tractable. However, comparisons between detailed and reduced chemistry models are frequently based on global observables and therefore do not clearly explain where local discrepancies arise or what physical and chemical processes are responsible for them. When treated as high-dimensional vector spaces, combustion datasets can be partitioned and described compactly through clustering and dimensionality reduction techniques. These techniques are typically used to create transportable Reduced Order Models instead of for analysis. Additionally, they are typically used in conjunction with state-space based representations of the data which do not provide mechanistic insight into local processes. This dissertation develops a process-based, interpretable, data-driven framework for comparative model analysis using Non-negative Matrix Factorization (NMF) applied to species-level chemistry and diffusion contributions. A bluff-body stabilized combustor is used as a motivating benchmark. A freely propagating laminar premixed flame and an opposed-jet premixed flame are used as controlled canonical configurations for method development and comparison. A 17-species reduced methane-air chemical kinetic model is compared against the detailed 53-species GRI Mech 3.0 to demonstrate the methodology. A state-space based L-PCA analysis provides a baseline description of thermochemical structure, while the process-space NMF analysis yields additive factors that are directly connected to local transport and chemistry. Reaction path analysis is used to validate the decomposition against known chemical pathways. The results show that process-based decompositions provide a useful complementary view of chemical model discrepancies and can organize similarities and differences between detailed and reduced mechanisms more clearly than global comparisons alone. In addition, the analysis suggests an emergent interpretation based on the relative alignment of chemistry and diffusion feature vectors, which may offer a useful lens for discussing flame-zone structure. Overall, this work shows that process-based unsupervised decompositions can serve a useful additional layer of analysis for comparing detailed and reduced chemistry models in premixed flames. 

Date and time: 2026-04-15, 10 am

Location: Gary Beringause Community Room, Georgia Tech Police Department

Committee:
jeff.jagoda@aerospace.gatech.edu (advisor), School of Aerospace Engineering
Dr. Lakshmi Sankar, School of Aerospace Engineering
Dr. Andrew Medford, School of Chemical and Biomolecular Engineering
Dr. Turab Zaidi, School of Aerospace Engineering
Dr. Christopher Jeffery, Los Alamos National Laboratory
,