Learning Adaptive Coarse Basis Functions of FETI-DP - Kölner UniversitätsPublikationsServer

Klawonn, Axel ORCID: 0000-0003-4765-7387, Lanser, Martin ORCID: 0000-0002-4232-9395 and Weber, Janine ORCID: 0000-0002-6692-2230 (2022). Learning Adaptive Coarse Basis Functions of FETI-DP. Technical Report. Open Access

Preview

PDF
CDS_TR-2022-02.pdf
Download (2MB) | Preview

Abstract

Domain decomposition methods are successful and highly parallel scalable iterative solution methods for discretized partial differential equations. Nevertheless, for many classes of problems, for example, elliptic partial differential equations with arbitrary coefficient distributions, adaptive coarse spaces are necessary to obtain robustness or, in other words, to guarantee a reliable and fast convergence. Adaptive coarse spaces are usually computed by solving many localized eigenvalue problems related to edges or faces of the domain decomposition. This results in a computationally expensive setup of the domain decomposition preconditioner or system operator. In this paper, we suggest to directly learn the adaptive constraints using a deep feedforward neural network and thus completely skip the computationally most expensive part of the setup, i.e., the solution of local eigenvalue problems. We consider a specific adaptive FETI-DP (Finite Tearing and Interconnecting - Dual Primal) approach and concentrate on stationary diffusion problems in two dimensions with arbitrary coefficient functions with large jumps. As an input for the neural network, we use an image representation of the coefficient function which resolves the structure of the coefficient distribution but is not necessarily identical to the discretization of the partial differential equation. Therefore, our approach is independent of the finite element mesh and can, in principle, be easily extended to other adaptive coarse spaces, problems, and domain decomposition methods. We show the robustness of our method for different problems and the generalization property of our trained neural networks by considering different coefficient distributions not contained in the training set. We also combine the learned constraints with computationally cheap frugal constraints to further improve our approach.

Item Type:

Preprints, Working Papers or Reports (Technical Report)

Creators:

Creators	Email	ORCID	ORCID Put Code
Klawonn, Axel	axel.klawonn@uni-koeln.de	orcid.org/0000-0003-4765-7387	UNSPECIFIED
Lanser, Martin	martin.lanser@uni-koeln.de	orcid.org/0000-0002-4232-9395	UNSPECIFIED
Weber, Janine	janine.weber@uni-koeln.de	orcid.org/0000-0002-6692-2230	UNSPECIFIED