Polycrystal Informatics 3DFM

• Polycrystal Informatics 3DFM is a comprehensive suite that integrates experimental, computational, and data-driven tools for 3D analysis of polycrystalline materials.
• It employs advanced methodologies—from voxelized microstructures and balanced power diagrams to virtual diffraction—to achieve high fidelity grain mapping and simulation.
• The framework supports uncertainty quantification, reduced-order modeling, and automated structural classification for actionable insights in materials design.

Polycrystal Informatics 3DFM refers to the comprehensive, multiscale computational and experimental toolkit for extracting, processing, and interpreting three-dimensional microstructural, crystallographic, and mechanical information from polycrystalline materials. “3DFM” encompasses both the theoretical foundation (crystallography, micromechanics, diffraction physics) and a suite of pipelines for measurement, simulation, surrogate modeling, and data-centric inference—spanning atomic, grain, and continuum scales. The goal is to connect experimental data, structure–property relations, and physics-aware reduced representations for applications ranging from experiment design to high-throughput analysis and uncertainty quantification.

1. Geometric and Crystallographic Model Foundations

At the core of polycrystal informatics 3DFM are precise, multi-level data structures encoding grain topology, orientations, and associated field variables. Grain maps can be represented using voxelized microstructures, generalized balanced power diagrams, or parameterized tessellations. In the voxel paradigm, individual elements are explicitly assigned grain IDs, orientation fields (typically as unit quaternions, rotation matrices, or Rodrigues vectors), and possibly phase/defect masks. An alternative, highly compact representation is provided by generalized balanced power diagrams (GBPD), where each grain is defined by its center-of-mass $p_i$, exact volume $v_i$, and, for non-equiaxed morphologies, a second-order inertia tensor $M_i$ extracted via PCA from experimental data. Cells are specified as regions of $\mathbb{R}^3$ minimizing an ellipsoidal norm-based "power distance," leading to a unique tessellation that exactly matches prescribed geometric statistics via solution of a sparse, totally unimodular linear program. This representation is particularly efficient for integration with simulation, graph-based ML, and mesh-based numerical solvers (Alpers et al., 2014).

Crystallographic states are encoded in orientation fields, with choices of parametrization optimized for group operations (e.g., quaternion, Rodrigues, or Euler angles), and texture distributions compactly described through discrete orientation distribution functions (ODFs) sampled over the fundamental zone of the point group. These structural descriptions are critical inputs for simulation, experimental inversion, and statistical analyses (Wei et al., 7 Dec 2025).

2. Measurement and Reconstruction Methodologies

Experimental access to 3D grain-resolved information spans methodologies from atomic probe tomography (APT), high-energy diffraction microscopy (HEDM), time-of-flight 3D neutron diffraction (ToF 3DND), and X-ray-based full-field imaging. Techniques such as ToF 3DND leverage the deep sample penetration of cold/thermal neutrons to resolve centimeter-scale polycrystalline samples with 100 μm spatial, 1–2° angular precision. Grain shapes are reconstructed by transmission extinction mapping and inverse Radon back-projection of rotationally acquired image sequences, with orientation determination via fitting Bragg's law and forward diffraction models in orientation–reciprocal-space (Cereser et al., 2017). The process is validated by cross-modal techniques such as EBSD, ensuring that 3DFM data pipelines maintain quantitative congruence to reference microstructure datasets.

At the atomic scale, APT crystallographic orientation is extracted via direct-space statistical analysis (e.g., DF-Fit), which performs pole-finding based on local distribution function deviations from the Poisson expectation, enabling robust detection without artifacts associated with Fourier-based approaches (Haley et al., 2019). Downstream segmentation, visualization, and export are performed in hierarchical formats suitable for integration into larger informatics systems.

3. Simulation and Forward Modeling: Virtual Diffraction and Mechanics

Synthetic data generation and virtual experimentation are critical both for experiment planning and for testing structure–property-function hypotheses. High-fidelity virtual diffractometers, as specified for 3DFM, transform loaded finite element meshes with per-element orientation and strain fields into synthetic detector images via sequential application of coordinate transformations, Laue/Bragg constraint evaluation, beam–detector raytracing, and detector point-spread modeling. Each quadrature point yields projected spot coordinates and attenuated intensity, with full-field intensity maps constructed via Gaussian convolution and PSF-aware pixel binning (Dawson et al., 2023).

Mechanical response at the mesoscale is captured by large-scale voxel-based finite element solvers such as ELAS3D-Xtal (Jeong et al., 7 Feb 2026), which can efficiently solve the anisotropic elastic equilibrium equations with per-grain (or per-voxel) orientation-resolved stiffness tensors. OpenMP-parallel preconditioned conjugate gradient solvers achieve scalability to domains exceeding $800^3$ voxels, supporting direct import of XCT/EBSD-derived microstructures, explicit defect insertion, and control over both statistical and deterministic grain texture. Output fields include full stress tensors, von Mises stress, and microstructure metadata in HDF5, facilitating immediate downstream informatics.

4. Statistical and Reduced-Order Modeling for Structure–Property Linkages

Data-driven reduced-order modeling constitutes a crucial layer of polycrystal informatics workflows, particularly when predicting mechanical response or damage evolution under realistic loading paths. A paradigm illustrated by (Zhang et al., 2023) extracts low-order moments (mean, variance, skewness, kurtosis) from massive 3D field simulations and deploys sparse model identification (via causation entropy and physically constrained regression) to uncover parsimonious ODE systems for moment evolution. These surrogates are validated both for in-distribution and predictive, extrapolative forecasting, with maximal entropy PDF reconstruction delivering probabilistic tail-risk metrics at negligible computational cost. This enables the translation of teravoxel full-field data into compact, uncertainty-quantified structure–property relations essential for materials-by-design and ICME contexts.

Similarly, the foundation model approach (Wei et al., 7 Dec 2025) leverages large-scale self-supervised pretraining on 3D orientation fields (quaternion-patch ViT) to learn a latent space that is transferable across downstream tasks, including homogenized elastic property regression and parameterization of nonlinear surrogate models (ODMN). Pretrained models display faster convergence, superior generalization in limited-data settings, and can be extended to any crystal class or multiphase system with appropriate training data. This suggests a scalable route for high-throughput property prediction and integration of experiment-derived microstructures.

5. Automated Structural Classification and Quality Metrics

For atomic or molecular dynamics datasets, robust structural identification algorithms such as Polyhedral Template Matching (PTM) (Larsen et al., 2016) underpin orientation, phase, and local strain extraction pipelines. PTM applies a suite of topological and geometric filters—neighbor shell ordering, Voronoi cell interface area, convex hull and planar graph enumeration, closed-form RMSD Procrustes alignment—to assign lattice type, orientation (as rotation/quaternion), and local elastic strain tensor to each atom. Complexity is held low ($O(1)$ per atom), and throughput reaches $10^5$ atoms/s per CPU core, enabling practical application to bulk datasets. The approach is highly robust to thermal fluctuations and avoids ambiguity found in purely geometric classification methods.

Integration of such methods with other 3DFM data streams provides a comprehensive informatics pipeline from atomic- to grain- to macroscopic-scales, supporting grain mapping, boundary network extraction, and per-grain or per-region quality/confidence mapping for downstream ML or knowledge discovery.

6. Informatics Workflow Architectures and Downstream Applications

Polycrystal informatics 3DFM emphasizes highly modular, pipelineable workflows. From microstructure generation (directly from stochastic/statistical models or experiment), through field simulation or virtual diffraction, to reduced-order representation and knowledge extraction, each module produces standardized, scalable outputs (typically HDF5 or VTK). Visualization and analysis software, such as Fourigui for reciprocal/real-space texture PDFs and pole figures (Harouna-Mayer et al., 2022), further support interpretation and real-time comparison.

Applications span experiment planning (via synthetic image generation), sensitivity analyses, real-time data-driven model updating, uncertainty quantification, grain-bounded mechanical property prediction, and integration into machine learning tasks (grain graph construction, supervised learning of structure–property relations). Quantitative agreement with experimental metrics, explicit uncertainty propagation, and fast re-optimization under new inputs are central to practical deployment.

A plausible implication is that ongoing developments in scalable, unsupervised representation learning and surrogate modeling frameworks will further streamline the translation from raw experimental/synthetic 3D datasets to actionable materials insight.

7. Performance, Validation, and Future Directions

3DFM workflows are validated through benchmark comparisons to analytical solutions (e.g., Eshelby inclusion (Jeong et al., 7 Feb 2026)), cross-modal congruence checks (EBSD vs. ToF 3DND (Cereser et al., 2017)), and quantitative metrics—center-of-mass error, voxel labelling accuracy, full-width-half-maximum and intensity errors in synthetic images, as well as information-theoretic discrepancy (KL divergence) between predicted and observed PDF tails (Zhang et al., 2023, Alpers et al., 2014).

Performance is continuously increasing: ELAS3D-Xtal achieves $10\times$ (CG) to $60\times$ (PCG) parallel speed-ups and routine solution of gigavoxel systems on single workstations (Jeong et al., 7 Feb 2026). Virtual diffractometers efficiently process hundreds of thousands of mesh elements and reflections (Dawson et al., 2023). Data-driven surrogates compress $O(10^5)$–$O(10^9)$ field samples to $O(4)$–$O(100)$ predictive parameters without sacrificing accuracy.

Further developments are anticipated in closed-loop integration (active learning, experimental feedback, online UQ), multi-modal data fusion (combining APT, diffraction, mechanical fields), and end-to-end differentiable design pipelines that can directly optimize for performance, reliability, and manufacturability given 3D microstructural data. The systematic use of polycrystal informatics 3DFM methods thus underpins contemporary efforts in quantitative, scalable, and uncertainty-aware materials science.