« Retourner aux résultats de recherche

• Citer

Development of physics-based reduced-order models for reacting flow applications

par Aversano, Gianmarco
Président du jury Coussement, Axel
Promoteur Parente, Alessandro ;Gicquel, Olivier
Publication Non publié, 2019-11-15

Thèse de doctorat

• Accès en ligne
• Détails
• Contenu
• Statistiques

Résumé :

Modern society will have to meet its energy demands while ensuring low or virtually zero emissions in order to meet future challenges associated to air pollution, climate change and energy storage. Very often, renewable sources cannot be directly employed because of their intermittent nature and because many applications such as transport and other in- dustrial processes require high energy densities. Therefore, novel storage solutions for the energy that renewable sources contribute to produce is necessary and the transformation of this energy into chemical compounds represents the best choice in order to meet the aforementioned demands, which requires novel combustion technologies, such as Moder- ate and Intense Low-oxygen Dilution (MILD) combustion, to be efficient and fuel-flexible. In order to develop such technologies, several studies are being proposed and terabytes of data collected as more and more experiments and high-fidelity simulations are carried out. However, there are two main challenges to this: the huge amount of data available makes it hard for the researcher to distinguish useful from redundant data, with the risk that useful information might stay hidden; the production process of these data-sets re- quires substantial resources as combustion process are multi-physics, multi-scale and thus require high-fidelity computationally-intensive simulations and experiments over a wide range for their operating conditions or input parameters. Digital twins and Artificial Intel- ligence (AI) are shaping the fourth industrial revolution by building data-driven models that make use of machine learning. It makes sense then to extend this approach to combus- tion applications in order to alleviate the two aforementioned issues: the use of machine learning techniques can help automate the process of data interpretation as well as pro- vide a low-dimensional representation of the high-dimensional data produced by either experiments or simulations; they can speed up the data production process by building reduced-order models that can foresee the outcome of a certain simulation with reduced or negligible computational cost. Besides, such reduced-order models are the foundations for the development of virtual counterparts of real physical systems, which can be employed for system control, non-destructive testing and visualization.With the final objective being to develop reduced-order models for combustion appli- cations, unsupervised and supervised machine learning techniques were tested and com- bined in the work of the present Thesis for feature extraction and the construction of reduced-order models. Thus, the application of data-driven techniques for the detection of features from turbulent combustion data sets (direct numerical simulation) was inves- tigated on two H2/CO flames: a spatially-evolving (DNS1) and a temporally-evolving jet (DNS2). Methods such as Principal Component Analysis (PCA), Local Principal Compo- nent Analysis (LPCA), Non-negative Matrix Factorization (NMF) and Autoencoders were explored for this purpose. It was shown that various factors could affect the performance of these methods, such as the criteria employed for the centering and the scaling of the original data or the choice of the number of dimensions in the low-rank approximations. A set of guidelines was presented that can aid the process of identifying meaningful physical features from turbulent reactive flows data. Data compression methods such as Principal Component Analysis (PCA) and variations were combined with interpolation methods such as Kriging, for the construction of computationally affordable reduced-order models for the prediction of the state of a combustion system for unseen operating conditions or combinations of model input parameter values. The methodology was first tested for the prediction of 1D flames with an increasing number of input parameters (equivalence ra- tio, fuel composition and inlet temperature), with variations of the classic PCA approach, namely constrained PCA and local PCA, being applied to combustion cases for the first time in combination with an interpolation technique. The positive outcome of the study led to the application of the proposed methodology to 2D flames with two input parameters, namely fuel composition and inlet velocity, which produced satisfactory results. Alterna- tives to the chosen unsupervised and supervised methods were also tested on the same 2D data. The use of non-negative matrix factorization (NMF) for low-rank approximation was investigated because of the ability of the method to represent positive-valued data, which helps the non-violation of important physical laws such as positivity of chemical species mass fractions, and compared to PCA. As alternative supervised methods, the combination of polynomial chaos expansion (PCE) and Kriging and the use of artificial neural networks (ANNs) were tested. Results from the mentioned work paved the way for the development of a digital twin of a combustion furnace from a set of 3D simulations. The combination of PCA and Kriging was also employed in the context of uncertainty quantification (UQ), specifically in the bound-to-bound data collaboration framework (B2B-DC), which led to the introduction of the reduced-order B2B-DC procedure as for the first time the B2B-DC was developed in terms of latent variables and not in terms of original physical variables.