• Citer

Thermo-hydraulics of cryogenic liquid storage

par Duque Morgado Marques, Pedro
Président du jury Parente, Alessandro
Promoteur Scheid, Benoît ;Mendez, Miguel Alfonso
Publication Non publié, 2025-12-02

Thèse de doctorat

• Accès en ligne
• Détails
• Contenu
• Statistiques

Résumé :

This thesis focuses on the experimental characterization and modeling of cryogenic liquid storage, under static and dynamic sloshing scenarios, and in terrestrial gravity and reduced-gravity environments. The experimental work centers on two objectives. The first is to characterize the coupling between sloshing dynamics and the resulting thermodynamic behavior in a pressurized cryogenic tank. The second focuses on the problem of cryogenic storage as a whole, and consists of developing a generalized scaling framework for heat and mass transfer in both static and dynamic scenarios.The first objective is tackled through experiments on-ground and in reduced-gravity conditions. The ground-based tests use liquid nitrogen (LN2) as the working fluid and apply lateral harmonic excitations to trigger sloshing in three wave regimes: planar, chaotic, and swirl. In microgravity, two identical tanks filled with non-cryogenic replacement fluid HFE-7000 are subject to a step drop in gravity: one remains in isothermal conditions, while the other is initially under a thermal stratification and pressurized ullage. Comparing the two tanks allows for assessing the impact of the thermal conditions on the sloshing dynamics and, in turn, how sloshing influences the thermal evolution of the system. In both campaigns, the experimental characterization combines pressure and temperature measurements with optical techniques.To address the second objective, this work scales the governing equations of the tank's thermodynamic state and highlights the key dimensionless groups that drive the dominant physical mechanisms in each scenario of interest, i.e., self-pressurization, active-pressurization, relaxation, and sloshing. The scaling analysis is tested against a broad range of experimental setups from the literature, spanning a wide range of tank sizes, operating conditions, and working fluids, as well as two dedicated experiments, one using HFE-7000 and the other with LN2.The modeling part of the thesis aims to establish a physics-integrated data-driven framework for predicting, and eventually controlling, the thermodynamic state of a cryogenic propellant tank. This part of the work tackles three main research questions. First, it investigates the feasibility of developing a real-time data assimilation framework for estimating closure parameters, i.e., heat and mass transfer coefficients, in a lumped thermodynamic model. The proposed approach combines ideas from traditional data assimilation and multi-environment reinforcement learning, where an agent's training (model assimilation) is carried out simultaneously on multiple environments (systems). The first proof of concept is performed using a simulated environment, for which the ground truth is available. The focus is placed on the viability of the learning process, its convergence, sensitivity to measurement noise, and how the convergence can be accelerated by simultaneously collecting data from multiple environments experiencing different scenarios. Second, the thesis explores the extension of the modeling framework toward control applications through the Reinforcement Twinning (RT) approach. This strategy is applied to a simulated cryogenic tank with a thermal vent system (TVS) under time-varying heat loads. The control problem aims to maintain the tank pressure under nominal conditions while minimizing the boil-off induced by the environmental heat ingress. The RT framework combines a model-based loop, where a digital twin is continuously updated and used to derive control policies, with a model-free reinforcement learning loop that learns directly from system interaction. The two strategies show complementary strengths: the model-based loop is sample-efficient but prone to local minima, whereas the model-free loop, though requiring more training iterations, enables broader exploration of the parameter space.Lastly, the data-driven modeling framework is applied to real-world data from cryogenic tests with LN2. The main objective is to assess the feasibility of deploying this framework in a real-world context, while identifying its limitations and evaluating its learning performance. The experimental database spans a range of operating scenarios, including self-pressurization, active pressurization, relaxation, venting, and sloshing. The digital twin of the system is based on a thermal nodal formulation, with closure parameters represented by neural networks dependent on dimensionless groups linked to the tank’s thermodynamic state. The results are encouraging, with the model showing good agreement with the experimental data across all operating scenarios.