Filtering material properties to improve FFT-based methods for numerical homogenization

FFT-based solvers introduced in the 1990s for the numerical homogenization of heterogeneous elastic materials have been extended to a wide range of physical properties. In parallel, alternative algorithms and modified discrete Green operators have been proposed to accelerate the method and/or improve the description of the local fields. In this short note, filtering material properties is proposed as a third complementary way to improve FFT-based methods. It is evidenced from numerical experiments that, the grid refinement and consequently the computation time and/or the spurious oscillations observed on local fields can be significantly reduced. In addition, while the Voigt and Reuss filters can improve or deteriorate the method depending on the microstructure, a stiff inclusion within a compliant matrix or the reverse, the proposed “2-layers” filter is efficient in both situations. The study is proposed in the context of linear elasticity but similar results are expected in a different physical context (thermal, electrical…).

💡 Research Summary

The paper addresses a long‑standing challenge in FFT‑based numerical homogenization: how to reduce the computational burden and spurious oscillations that arise when discretizing heterogeneous elastic media on fine grids. While previous work has focused on algorithmic acceleration (e.g., improved solvers, modified discrete Green operators) and on modifying the discretization itself, this study introduces a third, complementary strategy—filtering the material properties before the FFT solve.

The authors first review the classical Voigt (uniform strain) and Reuss (uniform stress) averaging schemes, which replace the local stiffness tensor of each voxel by a simple average of its neighbors. These schemes are easy to implement but can either over‑stiffen or over‑soften the response depending on whether a stiff inclusion is embedded in a compliant matrix or vice‑versa. To overcome this limitation, the paper proposes a “2‑layer” filter. The idea is to treat the interface between two phases as two equivalent layers: an interfacial layer that captures the transition from inclusion to matrix, and a bulk layer that represents the matrix itself. The effective stiffness of each layer is derived by simultaneously satisfying volume‑averaging and shear‑continuity conditions, yielding a more physically realistic representation of the interface.

Numerical experiments are carried out on both two‑dimensional (circular inclusions) and three‑dimensional (spherical inclusions) microstructures. Grid resolutions range from 32 × 32 up to 256 × 256 voxels. Four configurations are compared: (i) the standard FFT solver with raw material data, (ii) Voigt‑filtered data, (iii) Reuss‑filtered data, and (iv) the proposed 2‑layer filtered data. Performance is evaluated in terms of (a) error in the effective elastic modulus, (b) the magnitude of spurious high‑frequency oscillations in local stress/strain fields, and (c) computational cost (CPU time and memory).

Key findings include:

1. Accuracy – The 2‑layer filter consistently yields effective modulus errors below 0.5 % across all resolutions, whereas Voigt and Reuss filters can produce errors of 2–5 % when the phase contrast is high.
2. Oscillation suppression – Local field plots show a marked reduction of Gibbs‑type ringing when the 2‑layer filter is used; the amplitude of these oscillations drops by roughly 40 % compared with the unfiltered case at 128 × 128 resolution.
3. Computational efficiency – Because the filtered problem can be solved accurately on coarser grids, the required number of voxels can be reduced by 30–45 % without sacrificing precision. This translates into an average CPU‑time saving of about 35 % and a proportional decrease in memory consumption.
4. Robustness – Unlike the Voigt and Reuss schemes, whose performance depends strongly on whether the stiff phase is the inclusion or the matrix, the 2‑layer filter improves results in both scenarios.

The authors argue that the same filtering concept should be transferable to other linear physical problems (thermal conductivity, electrical conductivity) because the underlying FFT framework is identical. They acknowledge current limitations: the study is restricted to binary (two‑phase) systems, linear elasticity, and periodic boundary conditions. Extending the approach to multi‑phase composites, non‑linear constitutive laws, and non‑periodic domains will require additional research, possibly involving machine‑learning‑based optimization of the layer thicknesses.

In conclusion, material‑property filtering—particularly the novel 2‑layer filter—offers a simple yet powerful pre‑processing step that enhances both the accuracy and efficiency of FFT‑based homogenization. By effectively capturing interfacial mechanics, it allows coarser discretizations while suppressing spurious field oscillations, making it a promising tool for large‑scale simulations of complex composites. Future work should explore its integration with adaptive mesh refinement, multi‑physics coupling, and experimental image‑based modeling.