The current work focuses on the application of the Recursive Projection Method (RPM) to accelerate and stabilize time-stepping linear and non-linear solvers. The location of the eigenvalues of the Jacobian matrix with respect to the unit disk plays a fundamental role in this context: those lying outside the disk cause divergence, while those clustered to its boundary can affect the efficiency of the scheme in reaching a steady state. The idea behind the method is to apply a Newton iteration only in the subspace spanned by the eigenvectors associated with these eigenvalues while keeping the original fixed-point iteration in the stable subspace, which actually is not of concern. The RPM relies on the construction of an orthonormal basis, which can be carried out through the so-called Krylov criterion. Anyway, it suffers from some limitations, on the basis of which the AEP criterion is put under analysis and used as an alternative, being also able to provide an estimation of the eigenvalues without the need to solve the eigenvalue problem in a second stage. A procedure to maintain the accuracy of the basis is illustrated and applied as a fundamental tool to provide good results. The first test is performed on the linear advection-diffusion equation, the stability properties of which can vary depending on the setting of the advection and diffusion coefficients. The RPM is found to accelerate convergence when the original fixed-point scheme is stable, and to warrant convergence when it is unstable. In this context, different settings of the main algorithm parameters are explored to test the robustness of the method and to understand the potential correlation between these settings and the behaviour of the solution. The accuracy of the estimation of the unstable eigenvectors is explored, together with its impact on the efficiency of the method in the stabilization procedure. The focus is then brought to the non-linear Bratu equation, which has two solution branches, a stable one and an unstable one. The RPM still provides good results in terms of acceleration and stabilization, but its application is more critical and less robust. Again, several parameters combinations are investigated.
Il presente lavoro si concentra sull’applicazione del Recursive Projection Method (RPM) allo scopo di accelerare e stabilizzare schemi time-stepping lineari e non lineari. La posizione degli autovalori della matrice Jacobiana rispetto al disco unitario gioca un ruolo fondamentale in questo contesto: quelli al di fuori del disco portano a divergenza, mentre quelli posizionati al bordo possono ridurre l’efficienza dello schema nel raggiungere uno stato stazionario. L’idea alla base del metodo è di operare con il metodo di Newton solamente nel sottospazio relativo a questi autovalori, tenendo invece lo schema di origine a punto fisso sul sottospazio stabile, che non richiede particolare attenzione. L’RPM si basa sulla costruzione di una base ortonormale, che può essere attuata attraverso il criterio di Krylov. Tuttavia quest’ultimo ha diverse limitazioni, sulla base delle quali è stato proposto il criterio AEP e utilizzato come alternativa, essendo inoltre in grado di fornire una stima degli autovalori senza la necessità di dover risolvere il problema agli autovalori in un passaggio successivo. Una procedura per il mantenimento dell’accuratezza della base viene illustrata e applicata in quanto fondamentale per fornire buoni risultati. Il primo test viene effettuato sull’equazione lineare di advezione-diffusione, le cui proprietà in termini di stabilità possono variare in base al settaggio dei coefficienti di advezione e diffusione. L’RPM riesce ad accelerare la convergenza dello schema originale quando esso è stabile, e a garantire la convergenza quando è instabile. In questo contesto, vengono esplorate diverse impostazioni dei parametri principali dell’algoritmo, in modo da testare la solidità del metodo e comprendere l’eventuale correlazione tra il settaggio dei parametri e l’andamento della soluzione. Viene inoltre analizzata l’accuratezza della stima degli autovalori, insieme al suo impatto sull’efficienza del metodo nella stabilizzazione. L’attenzione viene poi portata sull’equazione non lineare di Bratu, la cui soluzione è composta da due rami, uno stabile e uno instabile. L’RPM fornisce ancora buoni risultati in termini di accelerazione e stabilizzazione, ma la sua applicazione risulta essere più critica e meno robusta. Anche in questo caso, vengono indagate diverse combinazioni di parametri.
Acceleration and stabilization of time stepping linear and non linear solvers through the recursive projection method
ESPOSITO, GIANPIETRO
2021/2022
Abstract
The current work focuses on the application of the Recursive Projection Method (RPM) to accelerate and stabilize time-stepping linear and non-linear solvers. The location of the eigenvalues of the Jacobian matrix with respect to the unit disk plays a fundamental role in this context: those lying outside the disk cause divergence, while those clustered to its boundary can affect the efficiency of the scheme in reaching a steady state. The idea behind the method is to apply a Newton iteration only in the subspace spanned by the eigenvectors associated with these eigenvalues while keeping the original fixed-point iteration in the stable subspace, which actually is not of concern. The RPM relies on the construction of an orthonormal basis, which can be carried out through the so-called Krylov criterion. Anyway, it suffers from some limitations, on the basis of which the AEP criterion is put under analysis and used as an alternative, being also able to provide an estimation of the eigenvalues without the need to solve the eigenvalue problem in a second stage. A procedure to maintain the accuracy of the basis is illustrated and applied as a fundamental tool to provide good results. The first test is performed on the linear advection-diffusion equation, the stability properties of which can vary depending on the setting of the advection and diffusion coefficients. The RPM is found to accelerate convergence when the original fixed-point scheme is stable, and to warrant convergence when it is unstable. In this context, different settings of the main algorithm parameters are explored to test the robustness of the method and to understand the potential correlation between these settings and the behaviour of the solution. The accuracy of the estimation of the unstable eigenvectors is explored, together with its impact on the efficiency of the method in the stabilization procedure. The focus is then brought to the non-linear Bratu equation, which has two solution branches, a stable one and an unstable one. The RPM still provides good results in terms of acceleration and stabilization, but its application is more critical and less robust. Again, several parameters combinations are investigated.File | Dimensione | Formato | |
---|---|---|---|
Thesis_new.pdf
accessibile in internet per tutti
Descrizione: Corpo di tesi
Dimensione
1.28 MB
Formato
Adobe PDF
|
1.28 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/10589/192595