Portable Microstructure Representation

• Portable microstructure representation is a set of compact encoding schemes that capture morphological, crystallographic, topological, and statistical features for multiscale analysis.
• They integrate discrete graph models, statistical correlations, and machine learning to achieve dimension reduction, symmetry preservation, and robust cross-material transfer.
• These representations empower surrogate modeling, uncertainty quantification, and inverse design by balancing computational efficiency with physical fidelity.

Portable microstructure representation refers to a class of compact, encoding schemes and data structures that capture the salient morphological, crystallographic, topological, and/or statistical features of a material microstructure for use across diverse analysis, simulation, design, and machine learning contexts. A representation is deemed "portable" if it is sufficiently concise, interpretable, and general to enable robust transfer across material systems, scales, and downstream tasks, without requiring ad hoc re-engineering or loss of critical microstructural information. Contemporary research spans graph-theoretic, statistical, topological, physical, and data-driven models—often incorporating dimensionality reduction, symmetry preservation, and hierarchical organization.

1. Discrete and Graph-Based Microstructure Encodings

Discrete microstructure representation schemes model material architecture as graphs, hypergraphs, or networks of sites and cells, suitable for both direct simulation and topological analysis. One canonical form is the Corner-Sharing Tetrahedra (CoST) abstraction, in which the microstructure is encoded as a pair $(V,T)$ with $V$ the nodes (positions of points in $\mathbb R^3$), and $T$ a set of tetrahedra such that any two tetrahedra share at most one vertex—never an edge or a face. The tetrahedral incidence matrix $B \in \{0,1\}^{n \times m}$ enforces this combinatorial constraint (Sitharam et al., 2018).

Physical constraints (e.g., strut lengths, bar-joint equilibrium) are assigned per edge, allowing a unified bar-joint or tensegrity model. Hierarchical refinement proceeds by barycentric subdivision, enabling multiscale representations—each refinement multiplies tetrahedra by 8; K refinements yield $O(8^K)$ complexity, but via hierarchical trees only the required resolution level is streamed. Algorithms support (1) incremental insertion/removal with $\mathcal{O}(1)$ hashing, (2) randomization mimicking disorder (perturbing vertex positions and local Newton solvers), and (3) conversion to continuous representations for finite element or meshless methods. CoSTs are compact and serve as a backbone for design, analysis, and manufacturing, outperforming high-resolution voxel or global phase-field models for storage, scalability, and multiscale analysis (Sitharam et al., 2018).

2. Statistical and Correlation-Based Representations

Statistical descriptors such as n-point correlation functions are foundational for microstructure quantification but traditionally incur intractable storage and computational costs. The n-point polytope functions $P_n$ generalize beyond 2-point statistics by considering the probability all $n$ vertices of a regular $n$-polytope of edge $r$ fall in a phase of interest:

$$P_n(r) = \langle \prod_{i=1}^n \mathbb{I}[x + v_i(r)] \rangle,$$

where $v_i(r)$ are polytope vertex offsets (Chen et al., 2019). For $n=2$, $P_2(r)$ recovers the classical two-point correlation; higher $n$ aggregate higher-order geometries (triangles, squares, polyhedra) and reveal hidden symmetries or emergent order. The normalized quantities $f_n(r) = [P_n(r) - \phi^n]/[\phi - \phi^n]$ encode deviations from the uncorrelated limit.

$P_n$ functions can be computed directly from multimodal imaging data (2D/3D/4D), and for $n\leq8$ deliver a portable, expressive, and interpretable microstructure representation. This hierarchical, basis-like approach dramatically compresses information: only $n$ curves need storage for up to 8-point statistics, facilitating cross-material transfer and property modeling. Compared to $S_n$ functions, $P_n$ curves are universal, as they are derived from regular geometries and can be computed on any dataset without material-specific tuning (Chen et al., 2019).

3. Machine Learning and Latent Space Approaches

Latent space representations employ unsupervised or supervised neural architectures—autoencoders, variational autoencoders (VAEs), convolutional deep belief networks (CDBNs)—to learn compact, typically high-level codes $z \in \mathbb{R}^d$ summarizing complex microstructural detail. In Convolutional Deep Belief Networks, a five-layer architecture encodes $200 \times 200$ binary micrographs into $30$-dimensional binary codes, achieving approximately $1000\times$ dimension reduction while preserving ensemble 2-point statistics and mean property values (Cang et al., 2016). The decoder supports stochastic reconstruction, yielding statistically faithful synthetic microstructures in diverse material classes.

The VAE-based framework for multiscale simulation encodes $24^3$-voxel models into $20$-dimensional latent vectors $z$, from which both microstructure reconstructions and fitted property vectors are recovered (Jones et al., 2024). Karhunen–Loève expansions enable sampling of spatially correlated latent fields for uncertainty quantification and functional gradation. These learned representations maintain scale separation, transfer across geometries, and inject seamlessly into finite element pipelines.

Similarly, the ReMiDi pipeline compresses high-resolution 3D meshes into a 16-dimensional latent code via spectral expansion and an autoencoder, facilitating end-to-end inversion against experimental observables (e.g., dMRI signals) (Khole et al., 4 Feb 2025). In microstructure evolution, latent codes from convolutional autoencoders coupled with neural operators (DeepONet) enable orders-of-magnitude acceleration of mesoscale simulation (Oommen et al., 2022).

4. Symmetry, Topology, and Physical-Invariance in Representation

Modern microstructure representation strategies increasingly enforce physical, geometric, and symmetry constraints for both efficiency and transferability. The spectral embedding approach for crystallographic data transforms SO(3)/H-valued orientation fields (accounting for point symmetry H) into low-dimensional, symmetry-respecting coordinates via Wigner D-matrix averaging and Gram–Schmidt orthonormalization, delivering generically injective, continuous, and SO(3)-equivariant descriptors (Pothagoni et al., 15 Oct 2025). This embedding achieves near-perfect CNN microstructure classification accuracy across data regimes and exhibits strong generalization to small volumes and varying textures.

Topological and geometric invariance is addressed through persistent homology and persistence images (Szemer et al., 16 Aug 2025). By tracking the appearance and disappearance of k-dimensional holes in a microstructure under filtration, one obtains persistence diagrams, which are subsequently transformed via smooth kernels into fixed-size images (persistence images). These can be vectorized and used as fully translation- and rotation-invariant descriptors, outperforming handcrafted morphological pipelines for property regression and classification.

Hybrid explicit–implicit neural representations, such as the Holoplane (Xue et al., 1 Feb 2025), encode geometry by storing a symmetry-reduced 2D tensor and reconstruct geometric and physical fields via an MLP. Physics-awareness is enforced by jointly encoding signed distance fields and homogenized displacement fields, aligning the representation with the material's mechanical response. Latent diffusion models trained in this compact space yield microstructures with guaranteed periodicity, symmetry, and property accuracy.

5. Algorithmic Frameworks, Hierarchy, and Scalability

Portable representations increasingly exploit hierarchical or multiscale organization. Stochastic tiling methods using Wang tiles (Doškář et al., 2020) replace periodic unit cells by a relatively small library of tiles encoding local microstructure, stochastically combined to generate large, aperiodic material domains. Offline computation of characteristic fluctuation fields (under first-/second-order gradients) for each tile enables re-use under arbitrary geometry or loading in a Generalized FEM ansatz. This structure ensures the reduced basis adapts to macro-boundary changes without repeated microstructural solves, offering scale-separation and transferability.

Hierarchical feature systems also underpin image-based descriptors for ultrahigh carbon steel (DeCost et al., 2017), where local keypoint features (SIFT), convolutional neural network features (VGG16), and orderless aggregation (VLAD, BoW) yield descriptors invariant to translation, robust to scale and contrast, and efficiently classifying microconstituents while generalizing across processing schedules and magnifications.

6. Applications, Performance, and Limitations

Portable microstructure representations act as foundational data structures across (a) surrogate modeling, (b) property prediction, (c) multiscale simulation, and (d) inverse design. They enable uncertainty quantification, generative synthesis, and accelerated experiment–design optimization loops.

Performance metrics must be matched to context: for statistical descriptors, faithfulness is measured via correlation function overlay and mean property preservation (Cang et al., 2016, Chen et al., 2019); for latent ML methods, dimension reduction ratios, mean-absolute-percentage errors (MAPE) on test sets, and transfer performance are the gold standards (Jones et al., 2024, Whitman et al., 28 Jan 2025). In all settings, information-theoretic efficiency, modularity, and interpretable alignment with material physics are critical.

Limitations, however, persist. Some latent representations may lack interpretability or physical guarantee outside trained manifolds. Fixed-size decoders can encounter scalability issues, and fine details may be sacrificed for dimensionality reduction. Ensuring robust performance on unseen microstructure types, rare phases, or in high-anisotropy scenarios often necessitates hybrid or physics-informed architectures, and statistical descriptors may miss features outside the chosen basis.

7. Comparative Summary of Selected Portable Microstructure Representations

Representation	Form/Encoding	Portability Properties
CoSTs (Sitharam et al., 2018)	$(V,T)$ corner-sharing tetrahedra	Discrete, multiscale, FEM-compatible
Polytope $P_n$ (Chen et al., 2019)	$n$-point regular polytope functions	Universal/statistical, interpretable basis
CDBN (Cang et al., 2016)	30D latent code via deep belief network	Task-agnostic, 1000$\times$ compression
pVAE (Jones et al., 2024)	20D latent + property mapping	FE drop-in, UQ, functional gradation
ReMiDi (Khole et al., 4 Feb 2025)	Spectral + 16D autoencoder code	Generalizable mesh, dMRI inversion
Persistence Image (Szemer et al., 16 Aug 2025)	Topological (birth–death) grid	Translation/rotation invariance
Holoplane (Xue et al., 1 Feb 2025)	64$\times$64$\times$32 tensor + MLP	Physics-aware, symmetry/periodicity
Spectral SO(3)/H (Pothagoni et al., 15 Oct 2025)	4D spectral embedding for orientation	Crystal symmetry, low data, injective

These frameworks collectively define the evolving toolbox for quantitative, explainable, transferable microstructure representation in computational materials and related disciplines.