Date of Award
Doctor of Philosophy (PhD)
Combustion processes are ubiquitous to human technological development and provide many benefits such as large-scale power generation for electricity and transportation along with residential and commercial heating for manufacturing, cooking, and warmth. However, these various processes can also have harmful effects on human health and the environment via emission of CO2 and other pollutants such as NOx and particulate matter (PM; often in the form of soot). For these reasons, there is a continued need for controlling, improving, and optimizing combustion processes. Modeling of these processes provides powerful insights into system-level dynamics and their control. Due to the size and complexity of industrial-scale combustion systems, there is a pressing need for the development of computationally-inexpensive models that can accurately predict gaseous and PM emissions. The research described in this dissertation addresses this need by: 1) careful evaluation of existing soot formation models for application in oxygen-enriched flames, 2) development of new, robust soot modeling capabilities with improved accuracy for flames outside of the normal fuel/air condition, and 3) production of a large data set for the development of machine learning-based algorithms for predicting pollutant emissions under a wide range of combustion operating conditions.
Over the past decades, several semi-empirical soot models have been developed for specific applications with unique characteristic timescales and/or validated only under fuel-air combustion conditions. Hence, their universal use, especially under oxygen-enriched combustion conditions, could lead to highly inaccurate predictions. Twelve semi-empirical models (1-step or 2-step) are evaluated based on their ability to respond to changes in stoichiometric mixture fraction (Zst) and strain in a series of ethylene counterflow flames spanning across the sooting-to-non-sooting (yellow to blue) transition. Results show that no existing model is able to predict a blue flame when Zst is increased beyond the experimentally-measured sooting limit.
Motivated by this finding, a novel modeling approach is presented to account for the unique flame characteristics at elevated-Zst environments and their effect on soot formation. This modeling approach is designed to capture both the formation and the reversible processes that occur on the fuel-side of a diffusion flame in a robust yet simple manner and can be utilized in many industrial combustion applications. A new semi-empirical formulation is presented that achieves this goal. In addition, extensions are presented for two widely-used semi-empirical models (Leung-Lindstedt and Moss-Brookes) which would otherwise be inaccurate at these conditions. Upon application of this approach to counterflow flame systems, the predicted soot volume fraction profiles agree well with experimental findings reported by previous studies under low Zst. This improved approach also resulted in the prediction of blue (soot-free) limit conditions in a non-premixed counterflow flame for the first time. Thus, the performance of semi-empirical soot formation models can be dramatically improved when the reversible nature of soot formation at high temperature is considered.
The next goal was to develop a machine-learning based modelling approach for combustion systems. As part of this collaborative effort, a series of experiments were performed using a lab-scale (25 kW) combustor that was operated under varying the fuel and air ratios. Measurements were made of temperature profiles along the reactor wall and gas composition and pollutants (CO, NOx, PM) in the exhaust. A series of tests were performed totaling 60 hours of runtime and 140,000 data points corresponding to each parameter. Findings from these experiments highlight a series of trends in the reactor: low and high primary air flows lead to elevated PM and NOx emission levels, respectively; NOx levels correlate with varying swirl ratios under fixed fuel / air ratio. Based on the generated data set, a model may be developed to accurately predict pollutant levels and subsequently recommend optimized operating conditions for the combustion system.
Rajan Chakrabarty Benjamin Kumfer
Richard Axelbaum, Pratim Biswas, Chun Lou, Jay Turner,