Improved guaranteed computable bounds on homogenized properties of periodic media by Fourier-Galerkin method with exact integration

TL;DR

This paper introduces an improved Fourier-Galerkin method with exact integration for computing bounds on homogenized properties of periodic media, outperforming previous FFT-based approaches in accuracy and efficiency.

Contribution

It develops a Galerkin approximation scheme with guaranteed bounds, extending FFT-based homogenization techniques with better accuracy and computational efficiency.

Findings

Ga scheme provides more accurate bounds than GaNi.

The method effectively handles high-resolution image-based materials.

Memory usage is reduced via FFT shift reformulation.

Abstract

Moulinec and Suquet introduced FFT-based homogenization in 1994, and twenty years later, their approach is still effective for evaluating the homogenized properties arising from the periodic cell problem. This paper builds on the author's (2013) variational reformulation approximated by trigonometric polynomials establishing two numerical schemes: Galerkin approximation (Ga) and a version with numerical integration (GaNi). The latter approach, fully equivalent to the original Moulinec-Suquet algorithm, was used to evaluate guaranteed upper-lower bounds on homogenized coefficients incorporating a closed-form double grid quadrature. Here, these concepts, based on the primal and the dual formulations, are employed for the Ga scheme. For the same computational effort, the Ga outperforms the GaNi with more accurate guaranteed bounds and more predictable numerical behaviors. Quadrature technique leading to block-sparse linear systems is extended here to materials defined via high-resolution images in a way which allows for effective treatment using the FFT. Memory demands are reduced by a reformulation of the double to the original grid scheme using FFT shifts. Minimization of the bounds during iterations of conjugate gradients is effective, particularly when incorporating a solution from a coarser grid. The methodology presented here for the scalar linear elliptic problem could be extended to more complex frameworks.