par Castellanos, Luisa
Président du jury Coussement, Axel
Promoteur Parente, Alessandro ;Simar, Aude
Publication Non publié, 2026-04-02
Thèse de doctorat
Résumé : With the ongoing climate crisis, it is key to redesign the energetic sector, as well as the hard-to-abate industries. In such tasks, it is important to consider the role of reacting flows for industrial applications, as well as combustion processes. This produces issues for efficient design of technologies, principally due to difficulties in obtaining a priori measurements of performance, and the high cost and limitations when it comes to gathering experimental data. More recently, the implementation of machine learning techniques has become of interest for combustion science. Principally, such techniques promise for dimensionality reduction, obtaining smaller representations of systems holding many variables, and facilitating the development of alternative models; additionally, it is possible to implement advanced regression techniques to obtain models that approximate the values of future thermochemical states, greatly reducing the computational costs of a priori approximations of engines. Regardless of such advantages, it is important to keep in mind that there is still work to do regarding efficient and correct implementation of machine learning techniques for combustion science. More in specific, once dimensionality reduction techniques are implemented, their physical interpretation is not necessarily straightforward, a phenomenon that is stronger in cases in which non-linear reduction techniques are implemented, which are ideal since they allow to keep chemical kinetics non-linearities. Furthermore, the development of surrogate models is greatly affected by error propagations, since every produced approximation is composed of the right thermochemical state, plus a disturbance. Such disturbances are likely to increase once the surrogate model feeds itself iteratively, or in other words, works in autoregressive form. Lastly, it is important to consider the dense training datasets that are required for developing surrogate models with enough accuracy, a practical limitation of machine learning techniques in many scientific fields.In this work, advances are made while encouraging the use of sparse datasets; improvements are made in the topics of interpretable machine learning, discussing the implementation of Time-lag Autoencoders for the analysis and reduction of chemical kinetics, allowing to obtain non-linear reduced representations which are physically interpretable. To emphasize contexts with great lack of data and easing the optimization process, the technique of Gappy-Autoencoder is introduced, enabling to analyze chemical mechanisms and reduce dimensionality in high-sparse data contexts, while keeping physically interpretable reductions. Additionally, the concept of Gradient-based Clustering is explored, which aims to group sets of solutions for chemical kinetics ODEs considering their time derivatives' behavior, under the rationale that, if two derivatives are similar, also both original functions will be. The meaning of such technique becomes more insightful when considering the gradients of a Combustion Progress Variable, a monotonic function that measure the development of combustion, and represents its dynamics. A proof of concept is given using a gradient-based clustering of a typical Progress Variable behavior, enabling for better accuracy and projection of future thermochemical states. Such a concept can be useful for future applications in matters of transfer learning and efficient development of models. Lastly, this work also analyses some core cases of time integrators development, principally for the case of Neural Networks. Such experiments, lead to a reconsideration in the typical training paradigms that are followed in the literature, and provides insights in better practices for surrogate models’ development.

Documents en relation

DI-fusion