• Citer

Data assimilation for the development of lifelong learning digital twins of combustion systems

par Donato, Laura
Président du jury Coussement, Axel
Promoteur Parente, Alessandro ;Galletti, Chiara
Publication Non publié, 2025-08-28

Thèse de doctorat

• Accès en ligne
• Détails
• Contenu
• Statistiques

Résumé :

The EU Climate Action targets a 55% reduction in greenhouse gas emissions by 2030 (compared to 1990 levels), aiming for a carbon-neutral society by 2050. Achieving these goals will require a significant contribution from renewable energy sources, which are essential for enhancing energy efficiency, enabling fuel flexibility, and reducing pollutant emissions. Accordingly, the growing urgency of the climate crisis calls for innovative solutions that accelerate the transition to sustainable energy systems.Although electrification and integration of renewable energy have made considerable progress in many sectors, heavy industry remains a major contributor to global greenhouse gas emissions due to its energy intensity and dependence on combustion-based processes. As these sectors are particularly difficult to decarbonize, there is a pressing need for advanced tools that can support the development and deployment of alternative low-carbon fuels, such as hydrogen and ammonia. To enable this transition, it is essential to develop new, reliable combustion technologies—underpinned by a deep understanding of combustion systems—to drive technological progress and ensure environmental sustainability.However, the complexity of combustion phenomena, especially under real industrial conditions, presents substantial modeling and experimental challenges. Traditional Computational Fluid Dynamics (CFD) methods, though powerful, are often limited by their high computational cost and long simulation times. These limitations restrict their applicability for large-scale parametric studies or real-time decision-making. This has driven the research community to investigate low-cost and multi-fidelity alternatives, namely physics-based reduced-order models and data-driven models, commonly referred to as Digital Twins (DTs). Still, numerical models can be affected by uncertainties, which may cause them to drift away from reality.At the same time, direct experimental investigations are often constrained by safety concerns, high costs, and the scale of industrial systems, resulting in sparse and noisy data. The combination of experiments and simulations is therefore key to advancing our knowledge of combustion processes and developing reliable and truly predictive computational approaches.In this context, the integration of Data Assimilation (DA) techniques and Machine Learning (ML)-based Digital Twins offers a promising path forward. These tools can enhance physical modeling by fusing experimental and numerical data, reducing uncertainties, and enabling faster, more accurate predictions. Specifically, DA consists of mathematical techniques in which observations and a numerical model are combined to generate more accurate predictions and provide an optimized estimate of the system’s state compared to the one that could be obtained using the data or the model alone.The overarching goal of this thesis is to investigate and develop such advanced data-driven DA frameworks to improve the understanding, modeling, and optimization of combustion processes for low-carbon fuels in industrial applications.In this context, DTs are built upon Reduced-Order Models (ROMs), which approximate the behavior of high-fidelity simulations at a significantly lower computational cost. These ROMs are constructed using dimensionality reduction and regression techniques, and they serve as the predictive core of each DT. Throughout the thesis, DA is applied to update these ROMs in real time, improving their accuracy and reliability.Fulfilling this objective leads to the development of various related activities:(i) A DA approach is applied for parameter estimation of combustion model coefficients characterizing a lifted methane–air flame, the so-called Cabra flame, which serves as a benchmark case. The results identify residual sources of uncertainty in the numerical model, by revealing discrepancies between the optimized model and the experimental data.(ii) A DA framework is developed to integrate in situ experimental data and update a digital twin model of a combustion furnace. The proposed setting involves the use of the Kalman Filter algorithm to adjust the prediction of the digital twin model by incorporating the uncertainties associated with both the model’s predictions and the experimental measurements. Specifically, the study examines the model's performance in predicting temperature profiles and NO concentrations in the combustion chamber with a fuel mixture ranging from pure methane to pure hydrogen. This methodology demonstrates enhanced digital twin accuracy, underscoring the importance of the Kalman gain in balancing prediction and measurement uncertainties.(iii) A DA scheme is implemented to efficiently incorporate high-fidelity experimental data, such as Laser Raman Scattering (LRS), into the digital twin of a combustion furnace. The measurements are projected in the low-dimensional manifold via a sparse sensing approach, and they are assimilated with the numerical prediction using the Kalman filter algorithm. This approach continuously enhances the digital twin model with newly available experimental data.(iv) A dynamic DA strategy is designed for updating a dynamic digital twin model developed for a stagnation-point reverse-flow combustor fed with an NH3/H2 mixture, which is vital for advancing low-carbon combustion technologies. This work implements the Ensemble Kalman Filter algorithm, which is able to update the model by assimilating experimental data, improving tracking of time-dependent variations and capturing uncertainties in the system states.Overall, this work demonstrates how established DA techniques can significantly improve the predictive accuracy and reliability of combustion system models. By bridging experimental and numerical data across different setups and operating conditions, the research underscores the potential of DA-powered Digital Twins in supporting the design, monitoring, and optimization of next-generation low-carbon combustion technologies.