Jacobi-accelerated FFT-based solver for smooth high-contrast data

Martin Ladecký, Ivana Pultarová, François Bignonnet, Indre Jödicke, Jan Zeman, Lars Pastewka · 2025

Read Jacobi-accelerated FFT-based solver for smooth high-contrast data by Martin Ladecký, Ivana Pultarová, François Bignonnet, Indre Jödicke, Jan Zeman, Lars Pastewka (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-02613.txt.

Introduction

Martin Ladecký and co-authors fix a slow solver for smooth, high contrast data in a 2025 paper. Fast Fourier solvers are strong tools for periodic cell problems in multiscale modeling. On regular grids they beat classic methods. But when material traits vary smoothly with sharp contrast, their convergence drags badly.

Their Green-Jacobi preconditioner upgrades the standard Green version while keeping near linearithmic cost. Paired with a conjugate gradient method, it cuts iteration counts a lot in elasticity tests, they report. The gain matters for fracture models, topology design, and adaptive grid solvers, all of which create smooth trait changes.

Catalog record

CREATORMartin Ladecký, Ivana Pultarová, François Bignonnet, Indre Jödicke, Jan Zeman, Lars Pastewka
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE82 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2508-02613

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Jacobi-accelerated FFT-based solver for smooth high-contrast data Martin Ladecký, Ivana Pultarová, François Bignonnet, Indre Jödicke, Jan Zeman, Lars Pastewka Abstract The computational efficiency and rapid convergence of fast Fourier transform (FFT)-based solvers render them a powerful numerical tool for periodic cell problems in multiscale modeling. On regular grids, they tend to outperform traditional numerical methods. However, we show that their convergence slows down significantly when applied to microstructures with smooth, highly-contrasted coefficients. To address this loss of performance, we introduce a Green-Jacobi preconditioner, an enhanced successor to the standard discrete Green preconditioner that preserves the quasilinear complexity, 𝒪 ⁡ ( N ​ log ⁡ N ) \mathcal{O}(N\log N) , of conventional FFT-based solvers. Through numerical experiments, we demonstrate the effectiveness of the Jacobi-accelerated FFT (J-FFT) solver within a linear elastic framework. For problems characterized by smooth data and high material contrast, J-FFT significantly reduces the iteration count of the conjugate gradient method compared to the standard Green preconditioner. These findings are particularly relevant for phase-field fracture simulations, density-based topology optimization, and solvers that use adaption of the grid, which all introduce smooth variations in the material properties that challenge conventional FFT-based solvers. 1 Introduction and Motivation The fast Fourier transform (FFT)-based solvers have matured and became a standard numerical tool for multiscale modeling of materials. Initially developed for homogenization of periodic microstructures [ 20 , 21 ] , FFT-based solvers are now used for various simulations of heterogeneous structures on regular grids; for an overview, see Refs. [ 15 , 24 , 7 ] . The term “FFT-based solver” is broad, encompassing various formulations, discretization approaches, iterative solution methods, and a discrete Green’s operator, efficiently implemented using the FFT algorithm to accelerate the computations. From the initial use of the Fourier basis, the FFT-based solvers have expanded to various discretization schemes, such as the finite differences [ 22 , 29 , 12 ] or finite elements [ 25 , 11 , 16 ] . These improvements mitigated discretization errors and reduced spurious oscillations that degrade the quality of local solution fields; for an overview, see Table 1 in Ref. [ 24 ] . The requirement for a regular discretization grid, intrinsic in the FFT algorithm, can be relaxed through local grid adaptation techniques [ 30 , 31 , 1 ] or by handling composite voxels either using an effective material property [ 10 ] or an X-FEM enrichment [ 5 ] . To solve the problem discretized on a regular grid, early FFT-based solvers [ 20 , 21 , 4 ] relied on fixed-point iterative schemes. Over time, a range of linear and nonlinear iterative solvers have been introduced to improve the convergence of these matrix-free methods. For an overview, see Table 2 in Ref. [ 7 ] or Table 4 in Ref. [ 24 ] . Although the state-of-the-art FFT-based solvers and their ancestors differ in many aspects, they all use the discrete Green’s operator in their core algorithms. Whether employed as a projection operator in strain-based schemes [ 20 , 32 ] or as a preconditioner in displacement-based schemes [ 25 , 11 , 16 ] , this operator plays a crucial role in improving the conditioning of the resulting system of equations. The discrete Green’s operator ensures that the conditioning of the system of equations remains independent of the mesh size, making FFT-based solvers particularly well-suited for problems with fine discretization and a large number of degrees of freedom (DOFs). Moreover, the sparse, block-diagonal structure of the discrete Green’s operator in the Fourier space enables its application via the FFT, with a quasilinear complexity of \mathcal{O}({N_{\mathrm{N}}}\log{N_{\mathrm{N}}}) , where {N_{\mathrm{N}}} denotes the total number of nodal points. During their 30-year history, FFT-based solvers have been applied to advanced phase-field models for crystal plasticity with fatigue cracking [ 14 , 3 , 18 ] , or topology optimization [ 19 ] . However, for such problems typical of smooth data with high-phase contrast, the Green’s operator preconditioned FFT-based solvers may exhibit slow and suboptimal convergence, as can be observed from the results presented in [ 14 , 3 , 18 , 19 ] . This limitation motivates our current research, and is demonstrated in the following simple example. Motivating example. Figure 1: Number of iterations of the preconditioned conjugate gradient (PCG) method required to solve mechanical equilibrium on a regular grid of 256^{2} nodal points as a function of the number of Gaussian filter applications, i . The green dashed line indicates the results for Green PCG, while black solid line indicates results for Green-Jacobi PCG. Figure (a.1) shows the initial two-phase material density \rho_{\mathrm{0}} , while \rho_{\mathrm{I}} in (a.2) is the density for which the number of iterations of Green PCG attains its maximum. The last density \rho_{\mathrm{II}} in (a.3) has the smallest total phase contrast \chi_{\mathrm{II}}=10^{2} . Panel (b) shows cross sections of material densities \rho_{0} , \rho_{\mathrm{I}} , and \rho_{\mathrm{II}} , at the middle row of nodal points shown by the dotted, dashed-dotted, and dashed lines in panels (a.1) to (a.3) , respectively. Let us consider a compliant circular inclusion in a stiff matrix, discretized on a grid of 256^{2} nodal points. The material is linear elastic and described with a stiffness tensor \mathchoice{\displaystyle\mathbf{C}}{\textstyle\mathbf{C}}{\scriptstyle\mathbf{C}}{\scriptscriptstyle\mathbf{C}}(\mathchoice{\displaystyle\boldsymbol{x}}{\textstyle\boldsymbol{x}}{\scriptstyle\boldsymbol{x}}{\scriptscriptstyle\boldsymbol{x}})=\rho_{0}(\mathchoice{\displaystyle\boldsymbol{x}}{\textstyle\boldsymbol{x}}{\scriptstyle\boldsymbol{x}}{\scriptscriptstyle\boldsymbol{x}})\mathchoice{\displaystyle\mathbf{C}}{\textstyle\mathbf{C}}{\scriptstyle\mathbf{C}}{\scriptscriptstyle\mathbf{C}}^{0}, which depends on a density function \rho_{0}(\mathchoice{\displaystyle\boldsymbol{x}}{\textstyle\boldsymbol{x}}{\scriptstyle\boldsymbol{x}}{\scriptscriptstyle\boldsymbol{x}}) . The density function is \rho_{0}^{\textrm{soft}}=10^{-4} in the soft material phase and \rho_{0}^{\textrm{hard}}=1 in the stiff material phase, see Figure 1 (a.1) . Therefore, the initial material contrast is \chi_{0}=\rho_{0}^{\textrm{hard}}/\rho_{0}^{\textrm{soft}}=10^{4} . Now, to construct smoother density fields \rho_{i} , we repeatedly apply a Gaussian filter to the initial function \rho_{0} . The filtering process involves discrete convolution of the density field with the kernel G=1/16\,\begin{bmatrix}1&2&1\end{bmatrix}^{{}^{\mathrm{T}}}\begin{bmatrix}1&2&1\end{bmatrix} , such that \rho_{i+1}=\rho_{i}\star G . The index i=\mathrm{0},\ldots,\mathrm{I},\ldots,\mathrm{II} indicates the number of successive filtering steps. We show two-dimensional plots of two samples of densities \rho_{\mathrm{I}} in Figure 1 (a.2) , and \rho_{\mathrm{II}} in Figure 1 (a.3) . Due to the filtration, the total phase contrast decreases, as we see in Figure 1 (b) . Here we show the densities \rho_{0} , \rho_{\mathrm{I}} , and \rho_{\mathrm{II}} in the middle row of nodal points. For each density \rho_{i} , we solve the micromechanical boundary value problem defined by the following system of equations: \displaystyle-\nabla\,\cdot\, \displaystyle\mathchoice{\displaystyle\boldsymbol{\sigma}}{\textstyle\boldsymbol{\sigma}}{\scriptstyle\boldsymbol{\sigma}}{\scriptscriptstyle\boldsymbol{\sigma}}(\mathchoice{\displaystyle\boldsymbol{x}}{\textstyle\boldsymbol{x}}{\scriptstyle\boldsymbol{x}}{\scriptscriptstyle\boldsymbol{x}})=\mathchoice{\displaystyle\boldsymbol{0}}{\textstyle\boldsymbol{0}}{\scriptstyle\boldsymbol{0}}{\scriptscriptstyle\boldsymbol{0}}, \displaystyle\mathchoice{\displaystyle\boldsymbol{\sigma}}{\textstyle\boldsymbol{\sigma}}{\scriptstyle\boldsymbol{\sigma}}{\scriptscriptstyle\boldsymbol{\sigma}}(\mathchoice{\displaystyle\boldsymbol{x}}{\textstyle\boldsymbol{x}}{\scriptstyle\boldsymbol{x}}{\scriptscriptstyle\boldsymbol{x}})=\mathchoice{\displaystyle\mathbf{C}}{\textstyle\mathbf{C}}{\scriptstyle\mathbf{C}}{\scriptscriptstyle\mathbf{C}}(\mathchoice{\displaystyle\boldsymbol{x}}{\textstyle\boldsymbol{x}}{\scriptstyle\boldsymbol{x}}{\scriptscriptstyle\boldsymbol{x}}):\mathchoice{\displaystyle\boldsymbol{\varepsilon}}{\textstyle\boldsymbol{\varepsilon}}{\scriptstyle\boldsymbol{\varepsilon}}{\scriptscriptstyle\boldsymbol{\varepsilon}}(\mathchoice{\displaystyle\boldsymbol{x}}{\textstyle\boldsymbol{x}}{\scriptstyle\boldsymbol{x}}{\scriptscriptstyle\boldsymbol{x}}), (constitutive law) \displaystyle\mathchoice{\displaystyle\boldsymbol{\varepsilon}}{\textstyle\boldsymbol{\varepsilon}}{\scriptstyle\boldsymbol{\varepsilon}}{\scriptscriptstyle\boldsymbol{\varepsilon}}(\mathchoice{\displaystyle\boldsymbol{x}}{\textstyle\boldsymbol{x}}{\scriptstyle\boldsymbol{x}}{\scriptscriptstyle\boldsymbol{x}})=\bar{\mathchoice{\displaystyle\boldsymbol{\varepsilon}}{\textstyle\boldsymbol{\varepsilon}}{\scriptstyle\boldsymbol{\varepsilon}}{\scriptscriptstyle\boldsymbol{\varepsilon}}}+\nabla_{\text{s}}\mathchoice{\displaystyle\boldsymbol{\tilde{u}}}{\textstyle\boldsymbol{\tilde{u}}}{\scriptstyle\boldsymbol{\tilde{u}}}{\scriptscriptstyle\boldsymbol{\tilde{u}}}(\mathchoice{\displaystyle\boldsymbol{x}}{\textstyle\boldsymbol{x}}{\scriptstyle\boldsymbol{x}}{\scriptscriptstyle\boldsymbol{x}}), where the Cauchy stress tensor \mathchoice{\displaystyle\boldsymbol{\sigma}}{\textstyle\boldsymbol{\sigma}}{\scriptstyle\boldsymbol{\sigma}}{\scriptscriptstyle\boldsymbol{\sigma}}(\mathchoice{\displaystyle\boldsymbol{x}}{\textstyle\boldsymbol{x}}{\scriptstyle\boldsymbol{x}}{\scriptscriptstyle\boldsymbol{x}}) is a function of the spatially varying elastic stiffness tensors \mathchoice{\displaystyle\mathbf{C}}{\textstyle\mathbf{C}}{\scriptstyle\mathbf{C}}{\scriptscriptstyle\mathbf{C}}(\mathchoice{\displaystyle\boldsymbol{x}}{\textstyle\boldsymbol{x}}{\scriptstyle\boldsymbol{x}}{\scriptscriptstyle\boldsymbol{x}}) and the small strain tensor \mathchoice{\displaystyle\boldsymbol{\varepsilon}}{\textstyle\boldsymbol{\varepsilon}}{\scriptstyle\boldsymbol{\varepsilon}}{\scriptscriptstyle\boldsymbol{\varepsilon}}(\mathchoice{\displaystyle\boldsymbol{x}}{\textstyle\boldsymbol{x}}{\scriptstyle\boldsymbol{x}}{\scriptscriptstyle\boldsymbol{x}}) . The small strain tensor \mathchoice{\displaystyle\boldsymbol{\varepsilon}}{\textstyle\boldsymbol{\varepsilon}}{\scriptstyle\boldsymbol{\varepsilon}}{\scriptscriptstyle\boldsymbol{\varepsilon}}(\mathchoice{\displaystyle\boldsymbol{x}}{\textstyle\boldsymbol{x}}{\scriptstyle\boldsymbol{x}}{\scriptscriptstyle\boldsymbol{x}}) is the sum of two parts: a constant macroscopic strain tensor \bar{\mathchoice{\displaystyle\boldsymbol{\varepsilon}}{\textstyle\boldsymbol{\varepsilon}}{\scriptstyle\boldsymbol{\varepsilon}}{\scriptscriptstyle\boldsymbol{\varepsilon}}} and the symmetrized gradient \nabla_{\text{s}}\mathchoice{\displaystyle\boldsymbol{\tilde{u}}}{\textstyle\boldsymbol{\tilde{u}}}{\scriptstyle\boldsymbol{\tilde{u}}}{\scriptscriptstyle\boldsymbol{\tilde{u}}}(\mathchoice{\displaystyle\boldsymbol{x}}{\textstyle\boldsymbol{x}}{\scriptstyle\boldsymbol{x}}{\scriptscriptstyle\boldsymbol{x}}) of the displacement fluctuation field \mathchoice{\displaystyle\boldsymbol{\tilde{u}}}{\textstyle\boldsymbol{\tilde{u}}}{\scriptstyle\boldsymbol{\tilde{u}}}{\scriptscriptstyle\boldsymbol{\tilde{u}}}(\mathchoice{\displaystyle\boldsymbol{x}}{\textstyle\boldsymbol{x}}{\scriptstyle\boldsymbol{x}}{\scriptscriptstyle\boldsymbol{x}}) , subject to periodic boundary conditions. We use a standard continuous and piecewise linear finite element (FE) discretization on a regular grid and the preconditioned conjugate gradient (PCG) method to solve the resulting linear system. We compare two different preconditioners: Green, which is the standard choice in FFT-accelerated solvers, and Green-Jacobi, which is introduced and studied in the remainder of this paper. In Figure 1 , we see the number of iterations of the PCG with respect to the number of the Gauss filter applications i . For the initial density \rho_{0} with sharp interfaces the Green PCG needs a substantially smaller number of iterations. However, for smoother densities, \rho_{i}

Read the complete text