par Procacci, Alberto
Président du jury Parente, Alessandro
Promoteur Coussement, Axel
Publication Non publié, 2024-11-29
Président du jury Parente, Alessandro
Promoteur Coussement, Axel
Publication Non publié, 2024-11-29
Thèse de doctorat
Résumé : | In future years, combustion technologies must go through a process of radical change if we hope to stop and reverse the process of climate change. New combustion technologies are needed with the emergence of zero-carbon fuels such as hydrogen or ammonia, which have drastically different physical properties than traditional fossil fuels. In addition, the development of new combustion devices is driven by the need of reducing pollutant emissions while maximizing combustion's efficiency and safety. This radical evolution of combustion technologies, combined with the tight time frame imposed by climate change, means that we cannot rely on traditional development methods. What we need is tools that can assist in the rapid development of new prototypes by identifying the combination of design parameters associated with optimal operating conditions. Such a tool must be capable of providing real-time estimations of the combustor's state, so that it can be deployed for combustion control, and it must be able to integrate available measurements to improve its predictions. Relying exclusively on experimental diagnostics is not an option, since providing direct access to the combustion zone is extremely cumbersome and expensive. Numerical simulations do not have this limitation, and they can be used to explore in detail the reactive region. However, numerical simulations of practical combustion systems are computationally very expensive and cannot be employed for fast estimation of the flame state or real-time control. Luckily, the recent advancements in machine learning research offer an array of tools to solve this problem. Data-driven reduced-order models are particularly attractive because they combine the high resolution of numerical simulations with the instantaneous state estimation provided by direct measurements. Reduced-order models are built by combining data compression and regression methods. The data compression step is employed to reduce the complexity of the regression problem, while regression is used to provide an estimation of the system's state for each combination of input parameters. The overarching goal of my work has been the development of reduced-order modelling techniques for combustion systems. Research was focused on linear compression methods, i.e. the proper orthogonal decomposition and its frequency-segregated variant multi-scale proper orthogonal decomposition. As for the regression method, the Bayesian technique known as Gaussian process regression has been employed to predict the system's state. Finally, to integrate numerical information with the experimental measurements, a combination of sparse sensing and data assimilation has been employed. Sparse sensing is a mathematical technique that exploits the intrinsic sparsity of physical phenomena to infer the system's state from few measurements. Data assimilation is employed to compare and adjust the estimations coming from the reduced-order and the sparse sensing models.In particular, the multi-scale proper orthogonal decomposition has been employed for the analysis of the temporal behavior of turbulent premixed and stratified flames issued from the Cambridge/Sandia burner. This was done to understand the impact of combustion on flow instabilities generated by the burner's configuration, such as the Bénard Von-Kármán and the Kelvin-Helmholtz instabilities. The results show that combustion suppresses the Bénard Von-Kármán instability due to the presence of the bluff body, while it does not affect the Kelvin-Helmholtz instability generated by the shear layer between inner and outer flow.The same set of flames was analyzed using sparse sensing to quantify the amount of information overlap between the velocity and chemiluminescence field. Sparse sensing was employed to predict the velocity field from the chemiluminescence signal, and the good level of accuracy reached in the prediction suggests a high level of correlation between the two signals.In a different setting, sparse sensing was employed to estimate the temperature field of a semi-industrial furnace from few temperature measurements. The accuracy of the prediction obtained using sparse sensing is comparable to the one obtained with numerical simulations, meaning that the sparse sensing model can be employed as a real-time surrogate model of the furnace.Another option to build a surrogate model is to couple the proper orthogonal decomposition with the Gaussian process regression. This methodology was applied to build a reduced-order model of the Cabra flame, which could predict the system's state given a combination of the Partially Stirred Reactor (PaSR) model's coefficients. The PaSR model takes into account the flame's turbulence/chemistry interactions, and its coefficients have to be tuned for different configurations. To identify the optimal set of PaSR coefficients, the reduced-order model was paired with a sparse sensing model which uses as input the experimental measurements of the Cabra flame. The two models were assimilated using the Kalman filter, and the optimal PaSR coefficients were identified using the differential evolution algorithm.The proper orthogonal decomposition can be employed also as a tool to perform data fusion, i.e. to build a model capable of predicting different quantities with vastly different spatial resolutions. This concept was demonstrated by building a reduced-order model of a semi-industrial furnace based on experimental measurements of species concentrations, temperature and chemiluminescence. The results have shown that this model can accurately predict the different signals, thus demonstrating the potential of this methodology to build surrogate models of reacting systems based on experimental information. Lastly, sparse sensing was employed to solve the computed tomography of chemiluminescence (CTC) inverse problem. The CTC's objective is to reconstruct the three-dimensional distribution of the chemiluminescence signal given a number of projections. By exploiting the proper orthogonal decomposition's data compression, it is possible to reduce the number of projections required to reconstruct the three-dimensional chemiluminescence field. This methodology was applied to the large eddy simulation of the Adelaide jet in hot coflow, and the results have shown a good reconstruction accuracy with only a single projection. Overall, this work demonstrates that machine learning provides a rich tool set for the development of reduced-order models of combustion systems. In the future, the objective will be to apply these techniques to build digital twins of real combustion systems. |