Innovative data-driven methods for clustering and flow control in reacting flows
- Autores: Adrián Corrochano
- Directores de la Tesis: Soledad Le Clainche Martínez (dir. tes.)
- Lectura: En la Universidad Politécnica de Madrid ( España ) en 2025
- Idioma: inglés
- Programa de doctorado: Programa de Doctorado en Ingeniería Aeroespacial por la Universidad Politécnica de Madrid
- Materias:
- Enlaces
- Tesis en acceso abierto en: Archivo Digital UPM
- Resumen
The present Doctoral Thesis develops new algorithms for developing Reduced Order Models (ROMs) in reacting flows. These novel algorithms pursue different objectives, which are the reduction of the dimensionality of big databases, the clustering of variables with similar dynamics and the control of the reacting flow to increase the efficiency and diminish the emission of pollutants.
These objectives are enclosed in the global goal for the European Union to become climate-neutral by 2050. With this purpose, the study of reacting flows has become very popular in the research community. However, modelling combustion is very challenging from both experimental and numerical insights. On the one hand, experiments are complicated due to the small time scales, the complex reaction driving their dynamics and the high cost of specialised materials, between other reasons. On the other hand, high-quality numerical simulations are also challenging. The numerous species involved, the turbulent-chemistry interaction and the wide range of magnitudes and time scales, make the numerical simulations highly expensive in terms of computational time and memory usage.
ROMs have raised to address this problem, as they can model the main flow dynamics at reduced computational cost. In the present Doctoral Thesis, the ROMs developed are based on Higher Order Dynamic Mode Decomposition (HODMD). This fully data-driven method, well-known in the fluid dynamics' community, decomposes a database into modes which move in time with a given frequency. These modes are patterns which describe the main flow dynamics. The first task is to adapt this algorithm to analyse reacting flows' databases. For this purpose, first, the database is preprocessed following the classical preprocessing techniques generally used in machine learning. The performance of the mentioned algorithm has been assessed with different quality indicators, measuring the error made on the calculations and the level of reduction of the dimensionality carried out. After that, the possibility to couple this algorithms with other typically used in combustion, as Principal Component Analysis (PCA) is also analysed with good results, demonstrating the robustness of the method for developing ROMs.
Once adapting the main algorithm used in the Doctoral Thesis for combustion databases, two different algorithms have been developed based on HODMD with two different and clear purposes. The first one, named hierarchical HODMD (h-HODMD), is a data-driven algorithm able to perform clustering of variables based on their dynamics. The algorithm iteratively performs HODMD and extracts a cluster of variables from the database, associating them a set of HODMD modes, frequencies and amplitudes. First, the algorithm has been used to reduce the reconstruction error, outperforming HODMD in the recreation of combustion databases. Secondly, the clusters have been analysed from the kinetics perspective, providing good results, as it made separated clusters for the nitrogen oxides and the oxy- and peroxy radicals, with different time scales for their dynamics.
The second algorithm developed aims to achieve flow control. With this purpose, the algorithm calculates the so-called non-linear structural sensitivity. This concept identifies the regions of the flow more prone to changes in the main dynamics. The algorithm was first developed for laminar non-reactive flows, and its performance was then compared to the classical theory. Due to its potential, the algorithm was then applied to more complex cases. The algorithm has been found to be useful for non-Newtonian fluids and combustion. In this latter case, the control of the flow is translated into an increase in the combustion efficiency and the reduce of the nitrogen oxides emissions.
