par Özden, Aysu
Président du jury Simar, Aude
Promoteur Parente, Alessandro ;Contino, Francesco
Publication Non publié, 2025-10-10
Thèse de doctorat
Résumé : Climate change is driving dramatic shiftsworldwide, including wildfires, droughts, and unusually high temperatures. Without urgent efforts to reduce greenhouse gas (GHG) and pollutant emissions, exceeding the established limits for global temperature rise will become inevitable. A crucial step is to develop innovative solutions that address the link between energy demand and pollutant emissions. Since most global emissions stem from traditional combustion-based processes, finding ways to reduce emissions from these sources will have the greatest impact.Researchers are actively working on novel combustion technologies that can achieve high energy efficiency, offer fuel flexibility, reduce greenhouse gas and pollutant emissions, and ensure robust and reliable operation. However, the design process of such systems is time-consuming and costly, as changes in design parameters directly affect system performance. Evaluating the operational behavior of these technologies heavily depends on extensive experimental campaigns and high-fidelity numerical simulations. Due to the multi-scale and multi-physics nature of combustion systems, achieving both physical insight and robust designs require tools that are accurate yet computationally efficient to explore and predict a wide range of operating conditions and device geometries.Exclusively relying on experimental diagnostics is not practical, as accessing the combustion zone is both technically challenging and prohibitively expensive. From a computational perspective, although computational fluid dynamics tools have advanced significantly in recent years, accurately predicting complex reacting flows—especially those involving detailed chemical kinetics and advanced turbulence-chemistry interactions—still demands immense computational resources, which remains a critical barrier.The growing interest and progress in machine learning algorithms offer a promising pathway to address this challenge. Data-driven reduced order models (ROMs) can replace computationally expensive full-order simulations, accelerating design exploration, optimization, and real-time predictions. However, these models typically require large training datasets generated from high-fidelity simulations, and the associated computational cost still represents a major limitation in their practical use.Multi-fidelity reduced order models provide a solution by striking a balance between accurate but costly high-fidelity data and less accurate but inexpensive low-fidelity data, enabling reasonably accurate system predictions at significantly reduced computational cost. Given the typically large number of quantities of interest and the high dimensionality of numerical domains, training datasets for such models suffer from the curse of dimensionality. To address this, dimensionality reduction techniques are commonly applied to simplify the regression problem. In multi-fidelity frameworks, a shared low-dimensional representation is established across both fidelity levels using manifold alignment techniques. These shared manifolds are then employed to train regression models capable of accurately predicting the system’s full state under new operating conditions.The overarching goal of my work has been to develop a multi-fidelity reduced order modeling framework for combustion systems. Specifically, my research combined proper orthogonal decomposition for dimensionality reduction, Procrustes manifold alignment to establish shared low-dimensional manifolds, and CoKriging-an extension of Gaussian Process Regression- to predict the system’s state.In particular, to address one of the key questions in multi-fidelity modeling—determining the minimum number of high-fidelity data points required for a desired level of model accuracy—I developed incremental sampling algorithms to (i) reduce the model prediction error, and (ii) minimize the uncertainty of the framework. The accuracy of the multi-fidelity framework and the effectiveness of the incremental sampling strategies were first assessed on a furnace operating with a methane-hydrogen fuel blend under Moderate and Intense Low-oxygen Dilution combustion conditions. The results showed that sampling strategies focused on reducing model uncertainty were most effective, and that the multi-fidelity model achieved prediction accuracy comparable to its single fidelity counterpart while reducing the number of training degrees of freedom by approximately half.The same framework and sampling strategies were also applied to ammonia combustion in a stagnation-point reverse-flow combustor to test the framework’s adaptability to different fuels and configurations. In addition, I investigated the effect of CoKriging predictions in extrapolation scenarios — specifically, test conditions outside the original design space. The results demonstrated that the multi-fidelity framework is flexible and effective across different combustion systems and fuels. While prediction accuracy and computational savings were lower in extrapolation cases compared to interpolation, the framework still successfully captured the qualitative behavior of key quantities of interest outside the training range.The training dataset for the ammonia case was initially limited, which motivated expanding the design space through additional simulations. This expansion revealed significant changes in combustion characteristics within the combustor as the equivalence ratio and air inlet temperatures decreased, largely due to ammonia’s narrow flammability limits. To further improve prediction accuracy, a hierarchical clustering technique was introduced to optimize the distribution of training samples. The results showed that this clustering approach not only provided a more structured strategy for model development but also improved prediction accuracy while reducing computational cost by tailoring the training process to localized behaviors within the design space.Finally, a multi-level multi-fidelity reduced order model was developed by integrating experimental measurements as high-fidelity data, while 3D and 2D RANS simulations served as mid- and low-fidelity datasets, respectively. The results demonstrated that incorporating additional high-fidelity experimental data progressively stabilized the model’s response and significantly improved its prediction accuracy compared to the multi-fidelity model trained exclusively on numerical simulations. It was also observed that the framework was highly sensitive to the initial selection of the linked datasets and to the constraints applied when selecting subsequent high-fidelity data points within the design space.Overall, this work shows that multi-fidelity reduced order models provide an effective solution for balancing accuracy and computational cost. The adaptability of this approach to different fuels and combustion systems makes it particularly valuable for the combustion community, as it accelerates design and optimization processes that are otherwise computationally intensive and costly, especially in combustion applications. Future work will focus on improving model accuracy by expanding the training dataset, optimizing hyperparameters, exploring alternativeregression methods, and integrating physical knowledge to develop physics-informed models.

Documents en relation

DI-fusion