Accurate and consistent composite voxel methods for digital images in computational micromechanics

The composite voxel method improves regular grid based computational homogenization schemes by designing a laminate-based surrogate material law to voxels containing more than one material. Recent work showed that the original methods for estimating the relevant volume fractions and the interface normal may be improved in case an analytic or a level-set description of the interface is known. This work is concerned with the case where a high-resolution image of the microstructure is available instead. We introduce two novel normal estimation techniques: the Interface-Centered Volume Regression (IVR) and the Digitally Averaged Interface Normal (DAN) methods. Both approaches are developed systematically and compared to existing strategies in terms of precision and applicability across different microstructures as well as material contrasts. The IVR method offers a robust regression-based alternative to the Composite Boxel (ComBo) approach at reduced implementation cost. The DAN method is fully local, lightweight to implement, and - unlike previous approaches - provably convergent in the limit of infinitely fine subvoxel resolution. The effectiveness of both methods is demonstrated through dedicated computational experiments.