A computational homogenization framework with enhanced localization criterion for macroscopic cohesive failure in heterogeneous materials

Ke Lü, F.p. van der MeerJournal of Theoretical Computational and Applied Mechanics 2022 · DOI

This paper, published in the Journal of Theoretical, Computational and Applied Mechanics, presents a significant advancement in how we simulate material failure across different scales. Researchers Lu Ke and Frans P. van der Meer from Delft University of Technology have developed a robust framework that allows engineers to predict how macroscopic cracks emerge and grow based on the complex, microscopic "messiness" of heterogeneous materials like composites or concrete.

TL;DR

The authors present a new multiscale framework that successfully bridges the gap between microscopic damage and macroscopic cracks. By introducing a refined localization criterion and a geometrically grounded "scale transition parameter," they ensure that simulation results remain objective and consistent, regardless of the mesh size or the size of the microscopic model used.

What Problem Does This Paper Address?

In the world of structural engineering, we often use Computational Homogenization (CH). Instead of modeling every single fiber or void in a giant bridge, we look at a tiny "Representative Volume Element" (RVE) to determine the material's overall properties.

This works beautifully when materials are hardening (getting stronger as they deform). However, it breaks down when materials soften—i.e., when they start to crack. Once a crack forms, the assumption that the RVE is "representative" of the whole material vanishes. This leads to a major mathematical headache: the results become "mesh-dependent," meaning the predicted strength of your structure changes depending on how small you draw your simulation grid. This paper aims to fix that "loss of objectivity" so that multiscale failure modeling can be used reliably in real-world engineering.

Methods

The researchers used a "two-scale" approach (often called FE²).

  • Macroscale: They used the phantom-node method to represent cracks. This allows a crack to cut through a standard finite element without needing to align the mesh with the crack path.
  • Microscale: Within each macroscopic point, they simulate an RVE where micro-cracks are modeled by inserting cohesive elements on the fly between material grains or fibers.

The Novelty: 1. Enhanced Localization Criterion: Usually, researchers use the "acoustic tensor" to figure out when and in what direction a crack will start. However, in complex materials, this method can get "confused" by multiple possible failure paths. The authors added a robust selection condition that picks the failure mode most closely aligned with the bulk strain field. 2. The Scale Transition Parameter ($h$): To bridge the scales, you need a way to turn a macroscopic "displacement jump" (how much a crack has opened) into a microscopic "averaged strain." The authors propose a new geometric interpretation for this parameter based on periodic crack spacing ($d_c$), ensuring energy is conserved across scales.

Datasets

As this is a theoretical and computational mechanics paper, the "datasets" consist of rigorous numerical benchmarks used to validate the framework:

  • Fiber-Reinforced Microstructures: Simulations of transverse tension and longitudinal shear in materials with random fiber inclusions.
  • Uniaxial Tensile Strips: Tests with both regular and complex (random) micro-crack patterns to check if the results change with RVE size.
  • Three-Point Bending: A classic engineering test compared against a "Direct Numerical Simulation" (DNS)—a gold-standard reference where every micro-detail is modeled explicitly.
  • Mixed-Mode Failure: A test designed to produce a curved crack path to see if the framework can handle arbitrary crack growth.

Key Insights and Ideas

  • The "Aha!" moment for $h$: The authors realized that the scale transition parameter $h$ shouldn't just be a random number or the height of the RVE. It represents the amount of material that "unloads" (relaxes) when a crack forms. By setting $h$ equal to the periodic crack spacing ($d_c$), they achieved a perfect energy balance between the scales.
  • Robustness over Simplicity: The standard bifurcation criteria often fail in shear or compression. The authors' "tie-breaker" condition (aligning failure with the strain field) proved essential for getting the crack angle right in complex loading scenarios.
  • Generality: The framework is "constitutive-agnostic." It doesn't care if your micro-model uses cracks, plasticity, or damage variables; the scale-bridging logic remains the same.

Key Results

  • Objectivity Confirmed: In uniaxial tests, the force-displacement curves for RVEs of different sizes (from 1x1 to 3x3 fiber arrays) were virtually identical. This proves the method is not sensitive to the size of the microscopic model.
  • Mesh Insensitivity: In the three-point bending test, using 80, 120, or 160 elements along the crack path yielded the same result, solving the "mesh-dependency" problem.
  • High Accuracy: The multiscale results matched the expensive, "brute-force" DNS reference solution almost perfectly, but at a fraction of the potential computational cost for large structures.

Limitations and Future Work

  • Boundary Condition Locking: The authors noted that "strong periodic boundary conditions" can sometimes cause "artificial locking." This happens when a crack wants to grow in a direction that doesn't align with the RVE's periodic edges, preventing the material from failing completely.
  • On-the-fly Detection: While they proposed a geometric interpretation for $h$, automatically detecting the "localization band orientation" in real-time for any arbitrary, evolving microstructure remains a challenge for future research.
  • Computational Cost: While more efficient than DNS, FE² is still heavy. Future work might involve integrating machine learning "surrogates" to speed up the RVE calculations.

Summary updated

Abstract

Computational homogenization allows to let the macroscopic constitutive behavior of materials emerge from microscale simulations without loss of generality with respect to microstructure and microscale constitutive response. Although computationally demanding, computational homogenization works very well for the hardening response of materials where the macroscopic stress and strain fields are smooth. However, in case of softening materials, when localization of deformation takes place, special care is needed to ensure objectivity of the method. In this paper, a generic multiscale computational homogenization approach for modeling onset and propagation of cracks in heterogeneous materials that is capable of considering various microscale mechanisms is presented. The common acoustic tensor bifurcation criterion is reinforced by an additional condition to help detect the localization mode more robustly. After the onset of macroscale localization, a key scale transition parameter is needed to translate the macroscopic displacement jump to an averaged strain over the micromodel domain. Then the macroscale crack is governed by a homogenized traction-separation relation evaluated from the underlying micromodel in which micro-failure accumulates. The scale transition parameter is studied for a range of different scenarios and endowed with a geometrical interpretation. Various numerical tests have been performed to confirm the objectivity and validity of the framework. The framework is generic in the sense that no assumptions on the microscale constitutive or kinematic representation of material failure are made in the scale transition. The framework is also highly compatible with the first order computational homogenization, which minimizes the additional complexity of adding macroscopic crack growth to the computational implementation.
BibTeX
@article{lu2022computational,
  author = {Ke Lü and F.P. van der Meer},
  title = {{A computational homogenization framework with enhanced localization criterion for macroscopic cohesive failure in heterogeneous materials}},
  journal = {Journal of Theoretical Computational and Applied Mechanics},
  year = {2022},
  doi = {10.46298/jtcam.7707},
}
Figure