Communication Dans Un Congrès Année : 2024

A Thorough Analysis of Deflation Techniques Applied to CFD: from RPM to BoostConv and Beyond

Résumé

Iterative methods such as Fixed-Point Iterations arising from the discretization of PDEs by standard methods (FE, FV, etc.) offer their share of challenges when applied to large systems due to their poor conditioning. Slow convergence, limit cycle oscillations and, in the worst case, divergence are regularly observed. Deflation-based techniques such as the Recursive Projection Method (RPM), are particularly appealing for tackling these kinds of issues without sacrificing the physical fidelity as they allow for the discrimination in between slow and fast modes, the first one being solved implicitly while the baseline algorithm is still relied upon to solve the latter. More recently, an alternative approach, called BoostConv, has emerged around the same idea, although the connection between the two has received little attention. In this paper, we offer to comprehensively revisit this family of methods for fixed-point iterations, giving rise to a new unified framework that encompasses both approaches and more and brings to light a number of as yet unexplored degrees of freedom, auguring further potential improvements in terms of performance and robustness. The present analysis focuses on three primary elements: choice of a projector, trouble modes recruitment, and rescue mode. A comparison of existing approaches is provided along with the benefits of proposed new variants on a selection of random matrices and CFD cases with different levels of complexity for the sake of completeness.
Fichier principal
Vignette du fichier
Draft_Sanchez_Pinedo_668171876pap_2228.pdf (750) Télécharger le fichier
Origine Fichiers produits par l'(les) auteur(s)

Dates et versions

hal-04920103 , version 1 (30-01-2025)

Identifiants

Citer

J. Kalfoun, G. Pierrot, J. Cagnol. A Thorough Analysis of Deflation Techniques Applied to CFD: from RPM to BoostConv and Beyond. 9th European Congress on Computational Methods in Applied Sciences and Engineering, Jun 2024, Lisbon (Portugal), Portugal. ⟨10.23967/eccomas.2024.271⟩. ⟨hal-04920103⟩
0 Consultations
0 Téléchargements

Altmetric

Partager

More