Attili, Antonio

Pillai, Rohit

Peterson, Brian

Arumapperuma Arachige, Geveen

2026-03-18T14:13:46Z

2026-03-18

Pollutants are undesirable by-products of combustion, posing significant challenges in the development of environmentally sustainable energy and propulsion systems. Among these pollutants, soot emissions are of particular concern due to their harmful effects on human health and the environment. As emission regulations continue to tighten, there is a growing demand for a detailed understanding of the physico-chemical mechanisms governing soot formation and evolution. Computational Fluid Dynamics (CFD) is expected to play an increasingly vital role in investigating these complex processes and guiding the design, optimisation, and certification of next-generation low-emission combustion systems. However, modelling soot remains particularly challenging, as it requires a detailed understanding of the complex interactions between turbulence, particle dynamics and chemistry. Despite significant progress, current models are still constrained by computational limitations and often lack the fidelity needed to capture these processes with sufficient accuracy. This thesis aims to systematically study soot and soot precursor formation and evolution in turbulent reacting flows and to use this understanding to explore and assess different modelling strategies. To assess the limitations of the current state-of-the-art soot and combustion models, Large-Eddy Simulations (LES) of a swirl stabilized model aero-engine combustor were performed. The results were validated against the experimental data showing that the simulations capture the gas-phase statistics accurately. The qualitative trends of the spatial distribution of the soot volume fraction are captured fairly well. A quantitative comparison shows that soot volume fraction is reasonably well predicted in the shear layers but significantly underpredicted elsewhere. Spectral analysis revealed that soot dynamics in the shear layer is characterised by relatively high frequency dynamics and are strongly coupled with large-scale turbulent structures (i.e., precessing vortex core). Therefore, these large-scale structures have a significant influence on the evolution of soot in the shear layers. This study highlighted the limitations of certain model components while also providing valuable insights into the relationship between flow structure and soot formation, that can be used to develop better models. It is known that the accuracy of the soot precursor predictions is crucial for the accurate prediction of soot. However, soot precursor formation and evolution in turbulent flames are not well understood. Polycyclic Aromatic Hydrocarbons (PAH), which are typically considered the most important soot precursors, are studied using a series of large-scale three-dimensional Direct Numerical Simulations (DNS) of turbulent non-premixed ethylene/air flames. The simulations employed finite-rate chemistry with a detailed ethylene oxidation mechanism including naphthalene as the PAH species. Three configurations are analysed: two cases have the same Reynolds number with different Damköhler numbers and one has a higher Reynolds number with the same Damköhler number as one of the lower Reynolds number cases. Results show that PAH formation correlates strongly with scalar dissipation rate but is insensitive to the Reynolds number. Additionally, an LES with tabulated chemistry of the higher Reynolds number flame is performed for an a posteriori analysis of the state-of-the-art PAH model. While the LES captures the spatial distribution of PAH with reasonable accuracy, it significantly underpredicts the magnitude of PAH. Furthermore, a new approach to read PAH from a flamelet table based on mean quantities is proposed and shows promising results. To address the shortcomings of the current state-of-the-art PAH model, a novel Machine Learning (ML) model based on a Convolutional Neural Network (CNN) was developed to predict the PAH mass fraction in a turbulent flame. During the development process, the generalisability of the CNN model in an LES context was rigorously tested via an a priori analysis. The study focused on the influence of varying training Reynolds numbers, filter sizes and filter kernels to evaluate the performance of the CNN model to out-of-sample conditions, i.e., not seen during training. The results revealed that the CNN models show good extrapolation performance when the training Reynolds number is sufficiently high. The CNN model was able to leverage the asymptotic nature of high Reynolds number turbulence. When the CNN models are tested on datasets with filter sizes not included in the training process, they exhibit sufficient interpolation capabilities. The extrapolation performance is less precise but still satisfactory overall. This indicates that CNN models can be effectively trained using data filtered with a limited range of filter sizes and then successfully applied across a broader spectrum of filter sizes. The CNN model was then applied to predict the PAH mass fraction in an LES. This a posteriori analysis showed that the spatial structure and the magnitude of PAH was captured with good accuracy by the CNN model. This showed a significant improvement in the results over the conventional transported PAH model. This showed that the CNN has a good ability to learn the complex spatial structure of the turbulent fields which can be leveraged to build high-quality models

https://era.ed.ac.uk/handle/1842/44494

https://doi.org/10.7488/era/7011

en

G. Arumapperuma, N. Sorace, M. Jansen, O. Bladek, L. Nista, S. Sakhare, L. Berger, H. Pitsch, T. Grenga, and A. Attili. “Extrapolation performance of convolutional neural network-based combustion models for large-eddy simulation: Influence of Reynolds number, filter kernel and filter size”. In: Flow, Turbulence and Combustion (2025), pp. 1–30.

G. Arumapperuma, Y. Tang, A. Attili, and W. Han. “Spectral analysis of soot dynamics in an aero-engine model combustor”. In: Proceedings of the Combustion Institute 40.1-4 (2024), p. 105344.

environmental pollution

soot formation

polycyclic aromatic hydrocarbons

machine learning

modelling

Physics-based and data-driven modelling of pollutant emissions in turbulent reacting flows

Thesis

Doctoral

PhD Doctor of Philosophy