Generating Accurate Microstructures of Anisotropic Magnetorheological Elastomers

1 Introduction

• A magnetic field during curing aligns particles into chains, producing anisotropic MREs whose effective magneto-mechanical behavior depends strongly on particle arrangement.Without a field during curing, particles remain randomly distributed and the MRE is isotropic.Source: §1 · p. 2
• Periodic RVEs support homogenization-based multiscale simulations, reducing the need to resolve samples that may contain millions of particles.An RVE should be as small as possible while remaining representative; periodic boundary conditions are favorable for representativity.Source: §1 · p. 2
• Pixel- or voxel-based generation can be costly for large cells and does not guarantee particle shape, size, or minimum-distance constraints.The two-point correlation function (TPC) describes spatial neighborhood relations and can serve as an optimization descriptor.Source: §1 · p. 2
• The proposed method generates two-dimensional statistically equivalent cells from microstructure images using a differentiable TPC formulation that reduces optimization parameters and controls particle sizes and distances.The cells are validated by comparing their effective magneto-mechanical behavior with the reference microstructure.Source: §1 · p. 2
• The model represents particles as circles in two dimensions, simplifying the three-dimensional microstructure; quantitative results are not expected to transfer directly to three dimensions.The methodology itself is expected to transfer to three-dimensional microstructures with appropriate modifications.Source: §1 · p. 2

2 Microstructure generation

• The generation procedure uses a differentiable TPC descriptor and minimum-distance constraint to optimize two-dimensional unit cells from a reference microstructure.The section also considers the spectral density function as an alternative descriptor.Source: §2 · p. 3, §2.2 · p. 4

2.1 Spectral density function and two-point correlation

• For circular inclusions, the TPC describes the probability of finding inclusions at both ends of a vector, and its Fourier coefficients yield the spectral density function.The derived expressions for the indicator function and TPC apply only to non-overlapping inclusions.Source: §2.1 · p. 3

2.2 Optimization

• The TPC loss compares the reference and generated cells on a shared spatial grid, but only inner grid points are compared, so long-range reference information is neglected.The minimum-distance condition is enforced before each descriptor optimization step.Source: §2.2 · p. 4
• The procedure first chooses particle radii and cell size, then optimizes particle centers to minimize TPC loss while enforcing minimum spacing.Radii are fixed during center optimization; monodisperse radii use the reference area-equivalent mean, while polydisperse radii are sampled from the reference.Source: §2.2 · p. 4

3 Finite strain magneto-hyperelasticity

• The RVE response is obtained by solving the coupled microscale field problem under prescribed macroscopic loads and averaging the fields into effective quantities.The computations use plane strain and two-dimensional magnetic quantities in the X1-X2 plane.Source: §3 · p. 5

3.1 Kinematics

• The deformation map defines reference and current configurations, while the deformation gradient and strain measures describe finite deformation.The formulation includes volumetric and isochoric splits of the deformation gradient and the right Cauchy–Green tensor.Source: §3.1 · p. 5

3.2 Maxwell and mechanical balance equations

• Magnetic induction and field strength obey the quasistatic Maxwell equations, and ponderomotive stress contributes to the total stress used in mechanical balance.The total stress is the sum of mechanical and ponderomotive stress.Source: §3.2 · p. 5

3.3 Constitutive equations

• The total Helmholtz free-energy density is additively decomposed into isochoric, volumetric, magnetic, and vacuum contributions, without a constituent coupling term.The passage states that coupling within each constituent is negligible because the matrix is non-magnetizable and the particles are effectively rigid.Source: §3.3 · p. 6

3.4 Scale transition and variational formulation

• Scale transition uses an RVE under periodic boundary conditions to connect microscale fields with macroscopic quantities and determine effective material behavior.Dilatation and pressure are among the primary variables used to avoid volumetric locking in quasi-incompressible materials.Source: §3.4 · p. 6

• Generated cells are compared with the reference using their simulated effective initial stiffness tensor, initial magnetic susceptibility tensor, and coupled magneto-mechanical response.Source: §4 · p. 6
• The comparison includes the initial magnetic susceptibility tensor as a distinct effective-response measure.Source: §4 · p. 6
• A coupled magneto-mechanical load case is included alongside the initial-property comparisons.Source: §4 · p. 6

4.1 Determination of the reference microstructure

• The reference cell is obtained by photographing a particle-containing matrix and then segmenting and parameterizing the individual particles.Source: §4.1 · p. 6

4.1.1 Experimental setup

• Nickel particles are dispersed in PDMS and exposed to a 216 mT field in the X2 direction; chains form within one minute before crosslinking.The sample is degassed and cooled before the field is applied.Source: §4.1.1 · p. 7, Figure 1 · p. 7

4.1.2 Particle segmentation

• Image processing enhances contrast, thresholds the image, and uses watershed segmentation with seed-distance filtering; particle regions may be merged, split, or omitted by area and shape criteria.Fitted circles preserve particle centers of mass and areas; boundary particles are removed and the cell is mirrored to enforce periodicity, yielding 1716 particles at area fraction 0.1188.Source: §4.1.2 · p. 7

• The generated cells use monodisperse particles, with square geometry and a resolution ratio of 4; higher resolution or polydispersity did not systematically or significantly lower effective-property error.Source: §4.2 · p. 7
• Each of 20 initializations is optimized for 1500 iterations, and the five cells with the lowest final loss are retained for each particle count.The loss ceases to decrease early for smaller cells but continues to decrease gradually for cells with more particles.Source: §4.2 · p. 8, Figure 3 · p. 8
• Generating a 400-particle cell takes approximately 3 hours on one CPU core, compared with about 5 minutes for a 64-particle cell.Source: §4.2 · p. 8
• Generated cells with 400, 256, 64, 36, 16, 9, and 4 particles have two-point correlations similar to the reference, though visible differences remain.The selected generated microstructures contain chains several particles thick, alongside some unaggregated particles.Source: §4.2 · p. 8
• Alternative tested optimizers and the combined descriptor-distance loss approach did not consistently or significantly reach lower loss.The tests included gradient and Barzilai–Borwein solvers, plus SLSQP, BFGS, L-BFGS-B, and CG.Source: §4.2 · p. 8

4.3 Homogenization results

• The evaluation compares initial elasticity and susceptibility tensors and response under coupled loading, using plane-strain kinematics and in-plane magnetic fields.The study’s generated-cell examples include 400, 256, 64, 36, 16, 9, and 4 particles.Source: §4.3 · pp. 8–9, Figure 4 · p. 9

4.3.1 Elasticity tensor

• The reference’s deviatoric stiffness is anisotropic, peaking along particle chains; stiffness error generally falls as generated-cell particle count increases.Near-incompressibility makes total directional stiffness much larger than deviatoric stiffness and nearly isotropic.Source: §4.3.1 · p. 10, Figure 6 · p. 10

4.3.2 Magnetic susceptibility tensor

• Susceptibility peaks along particle chains and declines with angular deviation; its error decreases through 16 particles, then remains relatively constant.Figure 7 reports means and standard deviations over the five lowest-loss cells for each particle count.Source: Figure 7 · p. 11, §4.3.2 · p. 11

4.3.3 Magneto-mechanically coupled behavior

• Across particle counts, stress error decreases as count increases until 16 particles, then remains 2%–3%; magnetization error similarly decreases from above 1% to around 0.4%–0.7%.These errors are averaged over the five generated cells with the lowest loss.Source: §4.3.3 · p. 11

4.3.4 Evaluation of the results

• Cells with 64 particles capture the considered effective response with errors about an order of magnitude below an isotropic 400-particle cell; 16-particle cells remain rougher proxies.The 64-particle cell area is more than 6 times smaller than the non-mirrored original; 16 particles reduce area by more than 26 times.Source: §4.3.4 · p. 11

5 Summary and outlook

• The procedure derives the two-point correlation function for circular inclusions and optimizes cells subject to a minimum inter-particle distance constraint.Particles are identified from the reference image, and cells of varying size are generated and validated against the reference.Source: §5 · p. 11
• 64 particles yield good agreement with the reference, while 16 already provide a reasonable approximation across effective directional stiffness, susceptibility, and magneto-mechanical response.The response comparison includes mechanical stress and magnetization under a specified load case; Table 2 reports homogenization errors for stress, magnetization, elasticity, and susceptibility.Source: §5 · p. 12, Figure 8 · p. 12, Table 2 · p. 12
• The study is restricted to two-dimensional geometries; three-dimensional extension would substantially increase computational cost, and the present optimization scheme may no longer suffice.The authors identify X-ray microcomputed tomography as a route to acquiring a three-dimensional microstructure.Source: §5 · p. 12
• Future work could use generated cells to produce training data for macroscale material models, while alternative descriptors may improve optimization behavior.The loss decreases by less than one order of magnitude despite many iterations; ellipsoidal inclusions could better approximate the true geometry but require iterative distance evaluations.Source: §5 · p. 12

CRediT authorship contribution statement

• Roth led investigation, visualization, software, validation, and original-draft writing, alongside conceptualization, analysis, methodology, and review.The statement also lists Jabs for conceptualization, analysis, methodology, software, and review.Source: CRediT authorship contribution statement · p. 12
• Schneider contributed resources, supervision, funding acquisition, conceptualization, analysis, software, methodology, and review.Gerstenberger contributed investigation, software, original-draft writing, and review.Source: CRediT authorship contribution statement · p. 12
• Auernhammer contributed resources, supervision, funding acquisition, and review; Kalina contributed conceptualization, methodology, supervision, and review.Kästner contributed resources, supervision, funding acquisition, and review.Source: CRediT authorship contribution statement · p. 12

A Microstructure generation with Spectral Density Function

• The SDF is investigated as an alternative because it contains the same information as the TPC in Fourier space rather than real space, potentially aiding optimization.The reference SDF is reduced to the generated cell's physical grid size before loss computation.Source: A Microstructure generation with… · p. 13
• The loss excludes the zero-frequency component because the reduced reference and generated cells generally do not match at zero frequency.The derivative with respect to particle centers uses the SDF derivative, and monodisperse cells are generated analogously to the TPC procedure.Source: A Microstructure generation with… · p. 13
• The figure shows three best generated microstructures with 64 particles each.The section's SDF comparison averages over five generated cells with 64 particles and lowest final descriptor loss.Source: Figure 9 · p. 13, A Microstructure generation with… · p. 13
• The five lowest-loss SDF-generated cells with 64 particles have average errors of 5.68% for elasticity, 2.04% for susceptibility, 2.76% for stress, and 0.60% for magnetization.These errors are similar to those obtained with TPC, with no systematic advantage for either descriptor.Source: A Microstructure generation with… · p. 13

B Microstructure with higher area fraction

• 7.82% for elasticity, 2.09% for susceptibility, 2.03% for stress, and 0.45% for magnetization are the averaged errors for the five lowest-loss generated cells.The reference has area fraction 0.1657; the authors report that the error range resembles the lower-area-fraction case, except that stiffness error is significantly higher.Source: B Microstructure with higher area… · p. 13
• The reference image yields 2,760 particles after mirroring; 20 monodisperse 64-particle cells are generated with minimum distance 0.2 max(R) ≈ 2.8.The three cells with the lowest final loss are shown.Source: B Microstructure with higher area… · p. 13
• The reported error range is similar to that for the lower-area-fraction case, but the stiffness error is significantly higher.The comparison concerns the averaged homogenization errors of the generated cells.Source: B Microstructure with higher area… · p. 13

C Generation of polydisperse microstructures

• 8.20% (C), 2.68% (χ), 3.48% (σ), and 0.71% (m) were the averaged effective-property errors for the best five polydisperse cells, worse than the monodisperse case.Source: C Generation of polydisperse… · p. 14
• 7.68 · 10−2 was the lowest polydisperse TPC loss, slightly below the best monodisperse cell's 7.75 · 10−2.Source: C Generation of polydisperse… · p. 14
• The study generated 20 cells with 64 polydisperse particles using varying initializations and radii sampled according to Algorithm 1.Source: C Generation of polydisperse… · p. 14