Séance Séminaire

Séminaire ACSIOM

Model order reduction via domain decomposition and the partition of unity condensed pole interpolation (PU-CPI)

In this talk, we discuss the PU-CPI, first developed in [1], which allows to construct a reduced order model which approximates the eigenvalues of an elliptic operator accurately and efficiently. The PU-CPI is a domain decomposition technique where local subdomain solutions are used to perform model order reduction via Rayleigh–Ritz projections. These local subspaces are constructed independently of each other, using data only related to the corresponding subdomain, resulting in a method that can be fully parallelized. Moreover, communication is only needed at the beginning and at the end, namely for distributing the local data, and for transferring the computed local results at the end of each task, respectively. Numerical examples will also be presented to illustrate the efficiency of the method. [1] A. Hannukainen, J. Malinen, and A. Ojalammi. Distributed solution of laplacian eigenvalue problems. SIAM Journal on Numerical Analysis, 60(1):76–103, 2022.