An efficient linear elastic FEM solver using automatic local grid refinement and accuracy control

https://doi.org/10.1016/j.finel.2012.12.007Get rights and content

Abstract

This paper presents an overview of an efficient FEM solver with an accuracy control strategy using automatic local grid refinement. Cheap accuracy estimators are briefly described and some concepts concerning a multigrid FE solver are presented. Several (2D/3D) numerical examples demonstrate the efficiency of the proposed strategy in terms of computation time and memory requirement. Some large scale calculations are presented, up to 200 million degrees of freedom. A particular implementation care has been devoted to ensure both a linear time complexity of every intermediate step and a minimal memory requirement.

Highlights

► Efficient Finite Element multigrid solver with an accuracy control strategy. ► Automatic local grid refinement. ► Linear time complexity and minimal memory requirement. ► Examples up to 200 million degrees of freedom are presented.

Introduction

Industrial problems lead to larger and larger finite element models reaching several hundred million degrees of freedom (dof). This paper presents some developments concerning an efficient finite element algorithm and its implementation.

The main features are:

  • a controlled accuracy level which is imposed a priori (before the calculation). Over-accurate (hence too costly) or under-accurate (hence possibly unsafe) solutions are avoided.

  • a fast solver and a quasi-minimal problem size. Very large problem sizes are targeted.

  • an automatic 3D mesher. The mesh is automatically built and reflects the mechanical problem. The meshing step is faster, safer and requires less expertise.

To achieve these objectives, an algorithm based on several discretisation levels is used, leading to a locally adaptative multigrid solver.

During the process of numerically analysing a physical problem, three main error sources are encountered: a modeling error (not discussed here), a discretisation error and a numerical error (which has to be reduced as fast as possible to be smaller than the discretisation error).

The discretisation error appears when a continuous problem is reduced to a set of discrete equations. The discretisation error depends on how the mesh adapts itself to the mechanical problem (element size, type, geometry, boundary conditions, loads, etc.). The design of an optimal mesh has been a particular point of interest for several decades: see Thompson et al. [1], [2], Shephard [3], Babuska et al. [4], [5], [6], [7], Southwell [8], [9].

To build an optimal mesh, hence ensuring a certain level of accuracy, error estimators are used. Under the generic term of error estimators, different techniques to estimate the discretisation error are grouped (including error indicators). The dynamic mesh construction based on global or local error estimators is the basic concept of adaptive techniques.

One of our goals is to obtain a quasi-uniform precision over the entire domain. To achieve this accuracy, several strategies are available, for example:

  • the h-method: element subdivision.

  • the p-method: interpolation function enrichment.

  • the hp-method: a combination of the two previous cases.

In this work, an h-method is used with a local error indicator; to obtain a targeted accuracy at a minimum cost, the mesh size has to vary over the domain. Several ways to build an adapted mesh have been used. Initially, when a solution was considered as locally inaccurate, the geometry was entirely remeshed to improve the solution quality in the concerned areas. Only conforming elements were used and element sizes were diminished continuously from coarse regions to fine regions; for example see Kornhuber and Krause [10], Aubry et al. [11] and Bornemann et al. [12]. In this work, a more recent and local non-conforming meshing technique is used.

Numerical simulations in various fields (fluid mechanics, solid mechanics, electromagnetism, etc.) lead to large systems of equations. Since the 1970s, Brandt [13], [14], [15] has developed multilevel methods to address the solution of such large systems. Venner and Lubrecht [16] applied and developed these techniques for contact mechanics. Adams [17] proposed an adaptation of these techniques to finite element computations. Multigrid techniques are often presented from the sole point of view of efficient solvers. However, some cited references explicitly propose error estimators.

One of the major contributions of this paper is to clearly demonstrate the efficiency reached with the combination of two multigrid technique advantages: an efficient convergence of numerical errors and a simple and intrinsic error estimation based on the comparison of solutions obtained on different discretisation levels. The multigrid solver can be seen as an iterative multiscale strategy in order to automatically capture different scales (structure, high gradient zones, local non-linearities, etc.) which is coupled with an error estimator in order to locally refine to reach a given level of accuracy, see Becker et al. [18] and Bastian and Wittum [19].

The local mesh refinement algorithm uses non-conforming meshes. The entire problem does not need to be remeshed at every step of the refinement strategy to incorporate a layer of ‘transition’ elements as in Wieners [20] and Schmidt and Siebert [21]. In the proposed strategy, hanging nodes (also referred later as non-coinciding nodes) are used to define the fine problem, see Jones and Jimack [22] or Janssen and Kanschat [23]. Becker and Braack [24] study two algorithms with conforming and non-conforming meshes; non-conforming techniques are found to be more efficient. Those two techniques are also developed for finite differences, for example see Wise et al. [25] where adaptive multigrid techniques using non-conforming nodes are used for a strongly anisotropic physical problem. Brown and Lowe [26] studied adaptive multigrid techniques for gravitational wave simulations and especially interpolation/restriction order influence on convergence speed.

Cavin et al. [27], [28] developed these methods for linear elastic finite element calculations (static and dynamic). Biotteau et al. [29], [30], [31] or Ekevid et al. [32] integrated non-linear material properties. Although Cavin et al. or Biotteau et al. demonstrated the potential of the method with respect to the accuracy control, the programming environment (Cast3M, [33], [34]) was not efficient enough to completely reveal the algorithm performance in terms of computational cost. The interlevel communication, data structure, etc. were not suited to this multilevel purpose. Indeed, multigrid techniques make intensive use of several discretisation levels and an efficient communication between these different levels is required. The performance gain is conditioned by a limited computing cost of this interlevel communication. This is especially the case for the interpolation/restriction operators. The implementation has to be made at a sufficiently low level to ensure that both data structures and operator implementations are very efficient. Basically, a linear complexity evaluation and a linear memory requirement have been ensured.

A dedicated multilevel finite element implementation has been developed from scratch to demonstrate the performance of the proposed method. This paper describes the main features of the code and describes some obtained results.

Section snippets

Theoretical framework

Very classical static linear elastic isotropic homogeneous and small displacement assumptions are used, for instance see Hughes [35]. The discrete problem is reduced to a set of linear equations where K is a stiffness matrix, u is the nodal displacements and f is the applied load:Ku=f

The algorithm presented in this paper merges an accuracy control, an efficient multilevel solver and an automatic 3D mesher. To construct an automatic mesh refinement which ensures the prescribed accuracy, three

Efficiency

The method efficiency is based on three distinct aspects:

  • An efficient solver.

  • A localised problem, thus a quasi-minimal problem size.

  • A dedicated implementation.

These three aspects are further developed in this section.

Results

Multigrid solver for global refinement: Some 2D results are presented up to 200 million dof using 1 CPU (2.0 GHz), 110 GBytes of RAM and less than 4 h of computing time. Below, some 3D calculations on more complex parts are shown. These 3D calculations used up to 167 million dof using 1 CPU and 228 GBytes of RAM in less than 42 h.

Localisation: The huge gain in memory requirement and computation time is illustrated. For the maximum accuracy tested, the developed code requires 1 day and 80 GBytes of

Conclusion

An efficient FEM solver based on a multigrid strategy has been developed. A local refinement technique is used to dynamically construct a mesh which ensures a prescribed accuracy. An accuracy indicator evaluation is cheap and is naturally derived from the refinement strategy using comparisons between different discretisation levels. The implementation reconciles contradictory objectives: memory requirement and computational cost (complexity). Several numerical examples covering a very large

References (52)

  • M.S. Shephard

    Finite Element Grid Optimization, ASME Special Publication PVP-38

    (1979)
  • I. Babuska et al.

    Reliable error estimation and mesh adaptation for the finite element method

  • I. Babuska et al.

    Error estimates for adaptive finite element computations

    SIAM, J. Numer. Anal.

    (1978)
  • I. Babuska et al.

    Error estimates for the combined h and p version of the finite element method

    Numer. Math.

    (1981)
  • I. Babuska et al.

    Adaptive Computational Methods for Partial Differential Equations

    (1983)
  • R.V. Southwell

    Relaxation Methods in Theoretical Physics

    (1946)
  • R.V. Southwell
    (1956)
  • R. Kornhuber et al.

    Adaptive multigrid methods for Signorini's problem in linear elasticity

    Comput. Visualization Sci.

    (2001)
  • F. Bornemann et al.

    Adaptive multilevel-methods in three space dimensions

    Int. J. Numer. Methods Eng.

    (1993)
  • A. Brandt

    Multi-Level adaptive solution to boundary value problems

    Math. Comput.

    (1977)
  • A. Brandt, Multi-level adaptative technique (mlat) for fast numerical solution to boundary value problems, in:...
  • A. Brandt

    Multigrid techniques1984 guide with applications to fluid dynamics

    (1984)
  • C.H. Venner, A.A. Lubrecht, Multilevel Methods in Lubrication, Elsevier Tribology Series, vol. 37, Amsterdam, 2000,...
  • M. Adams

    Evaluation of three unstructured multigrid methods on 3D finite element problems in solids mechanics

    Int. J. Numer. Methods Eng.

    (2002)
  • R. Becker et al.

    Adaptive error control for multigrid finite element

    Computing

    (1995)
  • P. Bastian et al.

    Adaptive multigrid methodsthe UG concept

    (1994)
  • Cited by (12)

    • Adaptive Isogeometric analysis for transient dynamics: Space–time refinement based on hierarchical a-posteriori error estimations

      2022, Computer Methods in Applied Mechanics and Engineering
      Citation Excerpt :

      This method is efficient to some extent because the number of elements requested to be calculated keep decreasing roughly. Then, Biboulet et al. [32–34] extended this strategy to be applicable in the 2D and 3D elastoplastic dynamics in the multigrid framework. In addition, space–time discontinuous Galerkin methods and the adaptive space–time mesh strategies for linear elastodynamics were well developed by Abedi et al. [35–37].

    • A general finite element method: Extension of variational analysis for nonlinear heat conduction with temperature-dependent properties and boundary conditions, and its implementation as local refinement

      2021, Computers and Mathematics with Applications
      Citation Excerpt :

      Despite numerical examples in Section 3 demonstrate that general FEM possesses higher accuracy and stability, direct implementation of general FEM to the entire domain replacing conventional FEM is of low efficiency, since calculations for matrices in general FEM (Eqs. (17)–(19)) are more costive and general FEM degenerates to the conventional FEM at point with small-amplitude or zero derivatives. Inspired by the strategies of local refinement in Refs. [38–43], general FEM is alternatively utilized as local refinement. Unlike Refs. [38–43] who majorly implement local refinements to meshes via h-type, p-type, hybrid type or adaptive methods, local refinement via utilizing general FEM is applied to governing equation.

    • An efficient strategy for large scale 3D simulation of heterogeneous materials to predict effective thermal conductivity

      2019, Computational Materials Science
      Citation Excerpt :

      With this method, one can solve the slowness problem of single level Jacobi relaxation due to the presence of low-frequency errors. The work of Biboulet et al. [3] shows the efficiency of using a MG method on the FEM. However, they assembled the stiffness matrix which, is very expensive for large scale problems.

    • Isogeometric local h-refinement strategy based on multigrids

      2015, Finite Elements in Analysis and Design
      Citation Excerpt :

      As a consequence, we will define an error indicator which allows us to have an approximate of the exact error. Furthermore, a keypoint is that it has to be not expensive in term of CPU time (see for instance [18,22,23]). It is shown in Cavin et al. [2] that displacement and energy error indicator are the only ones that behave well in singular problems.

    View all citing articles on Scopus
    View full text