Efficient preconditioning of hp-FEM matrix sequences with slowly-varying coefficients: An application to topology optimization

https://doi.org/10.1016/j.cma.2017.04.027Get rights and content

Highlights

  • We previously introduced a preconditioner that has proven effective for hp-FEM discretizations of various challenging elliptic and hyperbolic problems.

  • By building on top of our construction and previous numerical experiments, we develop a preconditioner-update strategy that allows us to handle matrix series arising from problems with slowly-varying coefficients.

  • We investigate the performance of the precondition along with the effectiveness of the update strategy in context of topology optimization of an acoustic cavity.

Abstract

We previously introduced a preconditioner that has proven effective for hp-FEM discretizations of various challenging elliptic and hyperbolic problems. The construction is inspired by standard nested dissection, and relies on the assumption that the Schur complements can be approximated, to high precision, by Hierarchically-Semi-Separable matrices. The preconditioner is built as an approximate LDMt factorization through a divide-and-conquer approach. This implies an enhanced flexibility which allows to handle unstructured geometric meshes, anisotropies, and discontinuities. We build on our previous numerical experiments and develop a preconditioner-update strategy that allows us to handle matrix sequences arising from problems with slowly-varying coefficients. We investigate the performance of the preconditioner along with the update strategy in context of topology optimization of an acoustic cavity.

Introduction

In this work, we apply the construction introduced in [1] to discretizations of problems with strong material discontinuities and time-varying coefficients. The core of the construction rests on the observation that, for a large class of problems, the dense Schur complement matrices that arise in the nested dissection method display rank-deficient off-diagonal blocks. In the case of well-behaved elliptic problems, this property can be traced to the separability of the underlying Green’s function, see, e.g., [2], [3]. To the contrary, this argument does not apply to wave-propagation problems, and, in the high-frequency limit, this property indeed ceases to hold, see [4]. Nevertheless, in the case of moderate frequencies, the off-diagonal blocks of the Schur complements can, for all practical purposes, be treated as low-rank. The intuitive explanation of this phenomenon is that, when considered on sufficiently small subdomains, the solution of wave problems resembles that of elliptic problems. By exploiting this property, approximate matrix decompositions can be constructed cheaply, and turn out to be excellent preconditioners.

A number of compressed-rank formats have appeared in the literature. Hierarchical matrices, or H-matrices, were first defined in the seminal work of Hackbusch, see [5]. Subsequently, the subclass of H2-matrices was introduced in [6]. Those matrices are attractive because allow traditionally expensive operations to be carried out within linear complexity, see, e.g., [7] for a detailed discussion. Hierarchically Semi-Separable (HSS) matrices were proposed by Chandrasekaran et al. in [8], and are closely related to H- and H2-matrices. Fast algorithms for their manipulation have been proposed, among others, by Sheng et al. [9], Xia et al. [10], Martinsson [11], and Gillman and Martinsson [12].

As described in detail in [1], the preconditioner construction combines a classical method, i.e., an LDMt factorization, with an approximation scheme for the resulting dense Schur complements. In fact, this approach is well established in the literature, see, e.g., Grasedyck et al. [13], and Xia et al. [14]. While our construction was largely inspired by the work of Gillman and Martinsson done in the context of finite difference discretizations, see [15], its main novelty is that we exploited the geometric flexibility provided by the LDMt factorization to accommodate for finite elements discretizations and unstructured meshes. In fact, while the main body of the literature deals with finite difference approximations, in the context of finite element discretizations we are only aware of the work by Aminfar and Darve, see [16], [17]. Although their approach is similar since dense frontal matrices are approximated by low-rank compression, they do not take advantage of the geometric partitioning versatility allowed by the LDMt factorization. In the context of finite difference approximations, the recent work of Ho and Ying, see [18], addresses the case of highly discontinuous coefficients and moderately indefinite problems.

In this work we describe the application of the preconditioner to high-contrast, time-varying Helmholtz problems arising in the context of acoustic Topology Optimization, see [19]. Topology Optimization is an iterative method that creates highly optimized designs by determining a distribution of material that fulfills a specific task. Typically, the method requires 100 through 1000 iterations to recover a locally optimal and physically admissible design. Consequently, the governing equations for the problem under consideration must be solved a large number of times for a slowly changing material distribution. When systems of several millions of degrees of freedom are considered, as is often the case for real world applications, their solutions through traditional direct techniques become infeasible. Naturally, this raises the interest in using highly scalable parallel iterative techniques. For physical problems governed by the Helmholtz equation, like acoustic, electromagnetic and structural vibration problems, no general scalable parallel iterative techniques currently exist. Therefore it is of interest to investigate the performance of the preconditioner in the context of Topology Optimization.

The paper is organized as follows. In Section 2 we present an improved analytical apparatus that allows us to better characterize the Schur complements as solution operators. In Section 3 we recall the construction of the preconditioner in order to make this work self-contained. Section 4 is devoted to numerical results. Finally, in Section 5, we draw conclusions from this work, and point towards future directions of research.

Section snippets

Analytical apparatus

We provide an insight into the rank-structure of the Schur complements that arise in the construction of the preconditioner. Let A be a finite element discretization of a boundary value problem posed on a domain Ω with boundary Γ=Ω. Since A is sparse, we can reorder its degrees of freedom (dof’s) to expose the following block-structure, and define the (aggregated) sub-matrices A(k): A=A(1)iiA(1)ibA(2)iiA(2)ibA(1)biA(1)bbA(1,2)A(2)biA(2,1)A(2)bb,A(k)=A(k)iiA(k)ibA(k)biA(k)bbk=1,2.The diagonal

Preconditioner construction and update

The construction relies on a variant of the well-known nested dissection algorithm, see George [23], and extends the work of Gillman and Martinsson, see [15], developed in the context of finite difference approximations.

Numerical results

As anticipated in Section 1, Topology Optimization is an iterative method, used mainly for PDE’s constrained optimization problems, to create highly optimized designs for specific purposes. The objective is to determine a distribution of material that fulfills a specific task in a locally optimal manner, without the need to enforce any a priori restriction on the design topology, see [19]. Although the use of gradient-based techniques results in a significant reduction in the number of

Conclusions and outlook

In this work, we review the construction of the preconditioner proposed in [1], sharpen our theoretical understanding of the methodology, and provide accurate cost estimates for different scenarios. The main novel contribution is a strategy to apply the preconditioner to time-varying discretizations. Specifically, we discuss an application to Topology Optimization of an acoustic cavity, which requires multiple solutions of a notoriously challenging Helmholtz problem with strong material

Acknowledgment

The work of R.E. Christiansen was financially supported by Villum Fonden through the research project Topology Optimization–the Next Generation (NextOpt) .

References (30)

  • MartinssonP.G. et al.

    A fast direct solver for scattering problems involving elongated structures

    J. Comput. Phys.

    (2007)
  • DühringM.B. et al.

    Acoustic design by topology optimization

    J. Sound Vib.

    (2008)
  • GattoP. et al.

    Efficient preconditioning of hp-fem matrices by hierarchical low-rank approximations

    J. Sci. Comput.

    (2017)
  • BörmS.

    Approximation of solution operators of elliptic partial differential equations by H- and H2-matrices

    Numer. Math.

    (2010)
  • XiaJ. et al.

    Superfast multifrontal method for large structured linear systems of equations

    SIAM J. Matrix Anal. Appl.

    (2009)
  • B. Engquist, H. Zhao, Approximate Separability of Green’s Functions for High Frequency Helmholtz Equations,...
  • HackbuschW.

    A sparse matrix arithmetic based on H-matrices. Part I: Introduction to H-matrices

    Computing

    (1999)
  • HackbuschW. et al.

    On H2-matrices

    Lect. Appl. Math.

    (2000)
  • BebendorfM.

    Hierarchical Matrices

    (2008)
  • ChandrasekaranS. et al.

    A fast ULV decomposition solver for hierarchically semiseparable representations

    SIAM J. Matrix Anal. Appl.

    (2006)
  • ShengZ. et al.

    Algorithms to solve hierarchically semi-separable systems

  • XiaJ. et al.

    Fast algorithms for hierarchically semiseparable matrices

    Numer. Linear Algebra Appl.

    (2010)
  • MartinssonP.G.

    A fast randomized algorithm for computing a hierarchically semiseparable representation of a matrix

    SIAM J. Matrix Anal. Appl.

    (2011)
  • GillmanA. et al.

    A direct solver with O(n) complexity for integral equations on one-dimensional domains

    Front. Math. China

    (2012)
  • GrasedyckL. et al.

    Domain decomposition based H-LU preconditioning

    Numer. Math.

    (2009)
  • Cited by (0)

    View full text