This thesis investigates the application of Data Driven Reduced Order Models (DDROMs) within a Data Assimilation framework to estimate the temperature, pressure, and veloc ity fields of a Magneto-hydrodynamics system. In magnetic confinement fusion, liquid metals are one of the most promising options for dissipating heat inside the reactor. Since the liquid metal is electrically conductive, it is strongly influenced by the external mag netic field, inducing Lorentz forces that depend on the coupling with the fluid velocity. Due to the strong nonlinearities and multiphysics of the governing equations, high-fidelity simulations are computationally expensive, and the solutions are highly sensitive to in put parameters, such as the intensity and orientation of an external magnetic field. In this context, DDROMs are a promising alternative to simulate specific scenarios, based on the knowledge of high-fidelity solutions for other parameter combinations and online measurements from sparse sensors, as they provide an accurate, reliable, and real-time state estimation, allowing for online monitoring and control of experimental facilities. The techniques employed in this work are the Generalized Empirical Interpolation Method and the Proper Orthogonal Decomposition, the latter is combined with either the Deep Oper ator Networks or a SHallow REcurrent Network (SHRED) to map measurements into the latent dynamics. The major contribution of this work is the application and comparison of the aforementioned methods in two parametric benchmark cases, varying a single or double parameter, i.e, the intensity of the magnetic field and/or the inclination angle. The results establish the superiority of SHRED over the other tested DDROM techniques for state estimation problems, across all sensor configurations and generalization modes. In the two-parameter case, when both the angle and intensity of the magnetic field vary together, SHRED reconstruction errors are always lower than 5% for the velocity, temper ature, and pressure fields, establishing SHRED state estimations as accurate and reliable even when the system dynamics are subjected to highly nonlinear, multi-parametric vari ations.
Questa tesi studia l’applicazione dei Data Driven Reduced Order Models con approcci di Data Assimilation, per stimare i campi di temperatura, pressione e velocità che carat terizzano un sistema magneto-idrodinamico. Nella fusione a confinamento magnetico, i metalli liquidi costituiscono la scelta migliore per dissipare il calore generato all’interno del reattore. Essendo il metallo elettricamente conduttivo, sarà fortemente influenzato dal campo magnetico esterno, il quale indurrà forze di Lorentz nel liquido a seconda dell’accoppiamento con il campo di velocità. In particolare, a causa della forte non lin earità e della natura multifisica delle equazioni governanti, le simulazioni ad alta fedeltà risultano computazionalmente onerose e le soluzioni ottenute sono altamente sensibili ai parametri di ingresso, come per esempio l’intensità e l’orientazione del campo magnetico esterno. In questo contesto, i DDROM rappresentano un’alternativa promettente per simulare scenari specifici, basandosi sulla conoscenza di soluzioni ad alta fedeltà per al tre combinazioni di parametri, e su misure online provenienti da sensori sparsi, potendo fornire una stima dello stato del sistema precisa, affidabile, in tempo reale, permettendo il monitoraggio e controllo di impianti sperimentali. Le tecniche impiegate in questo la voro sono la GEIM, la TR-GEIM e la POD combinata con DeepONet o con SHRED per mappare le misure alla dinamica latente. Il principale contributo di questo lavoro è l’applicazione e il confronto dei metodi sopra citati a due casi studio, nei quali viene stu diata la dipendenza da un singolo o doppio parametro, l’intensità del campo magnetico e/o l’angolo di inclinazione, dimostrando la superiorità del POD-SHRED in termini di precisione nella stima dello stato. Per quanto riguarda il caso doppio parametrico, lad dove sia l’angolo che l’intensità del campo magnetico esterno variano assieme, l’errore associato alla ricostruzione SHRED è sempre inferiore del 5% per i campi di pressione velocità e temperatura, dimostrando la precisione e l’affidabilità della stima dello stato anche quando la dinamica del sitema è altamente non lineare a seguito della variazione dei parametri.
State estimation in magnetohydrodynamics: from interpolation to recurrent networks
Scardino, Claudio
2025/2026
Abstract
This thesis investigates the application of Data Driven Reduced Order Models (DDROMs) within a Data Assimilation framework to estimate the temperature, pressure, and veloc ity fields of a Magneto-hydrodynamics system. In magnetic confinement fusion, liquid metals are one of the most promising options for dissipating heat inside the reactor. Since the liquid metal is electrically conductive, it is strongly influenced by the external mag netic field, inducing Lorentz forces that depend on the coupling with the fluid velocity. Due to the strong nonlinearities and multiphysics of the governing equations, high-fidelity simulations are computationally expensive, and the solutions are highly sensitive to in put parameters, such as the intensity and orientation of an external magnetic field. In this context, DDROMs are a promising alternative to simulate specific scenarios, based on the knowledge of high-fidelity solutions for other parameter combinations and online measurements from sparse sensors, as they provide an accurate, reliable, and real-time state estimation, allowing for online monitoring and control of experimental facilities. The techniques employed in this work are the Generalized Empirical Interpolation Method and the Proper Orthogonal Decomposition, the latter is combined with either the Deep Oper ator Networks or a SHallow REcurrent Network (SHRED) to map measurements into the latent dynamics. The major contribution of this work is the application and comparison of the aforementioned methods in two parametric benchmark cases, varying a single or double parameter, i.e, the intensity of the magnetic field and/or the inclination angle. The results establish the superiority of SHRED over the other tested DDROM techniques for state estimation problems, across all sensor configurations and generalization modes. In the two-parameter case, when both the angle and intensity of the magnetic field vary together, SHRED reconstruction errors are always lower than 5% for the velocity, temper ature, and pressure fields, establishing SHRED state estimations as accurate and reliable even when the system dynamics are subjected to highly nonlinear, multi-parametric vari ations.| File | Dimensione | Formato | |
|---|---|---|---|
|
2026_03_Scardino_thesis.pdf
accessibile in internet per tutti
Descrizione: Thesis
Dimensione
30.25 MB
Formato
Adobe PDF
|
30.25 MB | Adobe PDF | Visualizza/Apri |
|
2026_03_Scardino_Executive_Summary.pdf
accessibile in internet per tutti
Descrizione: Executive Summary
Dimensione
2.5 MB
Formato
Adobe PDF
|
2.5 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/253211