Generative reconstruction of 2D and 3D polycrystalline microstructures using symmetrized hyperspherical harmonics

This paper presents an open-source, differentiable framework implemented in MCRpy that utilizes symmetrized hyperspherical harmonics and advanced spatial correlation descriptors to efficiently generate high-fidelity 2D and 3D polycrystalline microstructures from limited 2D orientation data, thereby enabling robust structure-property linkage studies for materials design.

The Gist

Imagine you are a master chef trying to recreate a complex, multi-layered cake. You have a photo of the finished cake (the 2D data), but you need to build the entire 3D cake from scratch. The problem is, you don't have the recipe, and you can't see the inside layers just by looking at the photo. You have to guess the ingredients, the texture, and how the layers stack up, all while making sure the final cake tastes and looks exactly like the one in the photo.

This paper is about a new, high-tech "recipe generator" for materials scientists. Instead of cake, they are rebuilding polycrystalline materials (like metals) which are made of millions of tiny, interlocking crystal grains.

Here is the breakdown of their invention, using simple analogies:

1. The Problem: The "Flat Photo" vs. The "3D Reality"

Materials scientists often have a flat, 2D picture of a metal's internal structure (taken with a special microscope called EBSD). They want to use this to simulate how the metal will behave in the real world, which requires a full 3D model.

• The Old Way: Previous methods were like trying to guess the 3D shape of a cloud by looking at a single shadow. They often used "Euler angles" (a way to describe rotation) which are like trying to navigate a city using a map that has a giant hole in the middle. When you get near that hole, the directions get confused and break (mathematical "singularities").
• The New Way: The authors built a new system called MCRpy that uses a different mathematical language called Symmetrized Hyperspherical Harmonics (SHSH).
  • Analogy: Imagine describing a spinning top. Instead of using three confusing numbers that break when the top spins upside down, they use a smooth, continuous "sphere" of numbers. No matter how the top spins, the numbers flow smoothly without ever hitting a "dead end" or a glitch. This makes the computer much better at figuring out the correct 3D shape.

2. The Recipe: Three Special Ingredients (Descriptors)

To build the 3D metal from the 2D photo, the computer needs to know what to look for. The authors created a "checklist" of three specific features to ensure the new 3D model matches the real one:

• Ingredient A: The "Neighbor Check" (Two-Point Correlation):
  This asks, "If I pick a grain here, what kind of grain is usually found a few steps away?" It ensures the grains are the right size and shape (e.g., long and thin, or round).
• Ingredient B: The "Curvature Check" (Hybrid Three-Point Variogram):
  This is a new, fancy tool. It doesn't just look at neighbors; it looks at how the grains bend and curve relative to each other.
  • Analogy: If Ingredient A tells you the bricks are the right size, Ingredient B tells you if the wall is straight or if it has a nice, smooth curve. It helps the computer draw sharp, realistic boundaries between the grains instead of blurry, fuzzy ones.
• Ingredient C: The "Smoothness Check" (Mean Variation):
  This acts like a gentle hand smoothing out the clay. It stops the computer from creating weird, noisy static (like TV snow) while making sure it doesn't smooth too much and erase important details.

3. The Cooking Process: Gradient-Based Optimization

How does the computer actually build the model?

• The Old Way: It was like a blindfolded person throwing darts at a board, hoping to hit the bullseye. They would guess a shape, check if it was close, and if not, guess again. This took forever.
• The New Way: The authors use Gradient-Based Optimization.
  • Analogy: Imagine you are standing on a foggy mountain and want to get to the lowest valley (the perfect 3D model). Instead of throwing darts, you feel the ground under your feet. You can feel exactly which way is "downhill." The computer takes a step in that direction, feels the ground again, and takes another step. It keeps sliding down the hill until it reaches the bottom. This is incredibly fast and efficient.

4. The Results: From 2D to 3D

The team tested this on an aluminum alloy that had been processed with heat and pressure.

• The Test: They gave the computer a 2D slice of the metal and asked it to generate the full 3D block.
• The Outcome: The computer successfully "grew" a 3D block that looked and behaved statistically like the real metal. It captured the shape of the grains and their crystal directions perfectly.
• The Catch: The system works great when the metal looks the same everywhere (homogeneous). However, if the metal has a "gradient" (like being very coarse on one side and very fine on the other), the system tends to average it out. It's like trying to recreate a sunset that fades from orange to purple; the system might just make the whole sky a uniform pinkish-orange because it's looking for the "average" color.

Summary

This paper introduces a powerful new tool that allows scientists to turn a flat, 2D photo of a metal's microscopic structure into a full, 3D digital twin. By using a smooth, glitch-free mathematical language (SHSH) and a "sliding downhill" optimization method, they can generate these 3D models much faster and more accurately than before. This helps engineers design better materials by simulating how they will behave in the real world without needing to build expensive, complex 3D scans every time.

Technical Summary

Technical Summary: Generative Reconstruction of 2D and 3D Polycrystalline Microstructures Using Symmetrized Hyperspherical Harmonics

Problem Statement
Establishing robust structure-property linkages in polycrystalline materials requires representative two- (2D) and three-dimensional (3D) microstructural inputs for full-field simulations, such as phase-field or crystal plasticity models. While experimental 3D characterization is possible, it is often cost-prohibitive and complex for high-throughput Integrated Computational Materials Engineering (ICME) workflows. Consequently, the field relies on Microstructure Characterization and Reconstruction (MCR) to generate statistically equivalent synthetic microstructures from limited 2D data.

Existing reconstruction methods face significant limitations. Geometric tessellation approaches (e.g., Voronoi) are constrained by inherent convexity and often fail to capture complex intragranular substructures. Deep learning methods, while powerful, often lack physical interpretability and transparency. Furthermore, traditional descriptor-based stochastic reconstruction algorithms suffer from slow convergence due to reliance on discrete pixel-swapping and simulated annealing. A critical gap remains in extending differentiable reconstruction frameworks to directly handle crystallographic orientation data. Standard orientation representations, such as Euler angles, introduce numerical singularities (gimbal lock) and discontinuities that destabilize gradient-based optimization, making them unsuitable for high-fidelity, differentiable microstructure synthesis.

Methodology
This work introduces an orientation-based differentiable microstructure characterization and reconstruction (DMCR) framework implemented within the open-source software MCRpy. The methodology addresses the challenges of orientation representation and optimization through the following components:

1. Symmetry-Invariant Orientation Representation:
   To overcome the singularities of Euler angles, the framework utilizes unit quaternions to parameterize the orientation space. Building on this, it employs Symmetrized Hyperspherical Harmonics (SHSH) to derive a continuous, symmetry-invariant representation of crystallographic orientations. SHSHs serve as an orthonormal basis for functions defined on the 4D unit hypersphere ($S^3$), effectively handling the two-to-one homomorphism between quaternions and the rotation group $SO(3)$. This approach eliminates numerical discontinuities and allows for a spectral decomposition of the Orientation Distribution Function (ODF) that respects crystal symmetry.

2. Differentiable Descriptor Set:
   The reconstruction is driven by a loss function minimizing the statistical divergence between a candidate microstructure and a target reference. The descriptor set ($D$) combines:

   • Two-point spatial correlations ($S$): Captures global texture and cross-correlations between different SHSH modes.
   • Hybrid three-point variogram ($\gamma_3$): A novel descriptor that quantifies higher-order spatial dependencies and curvature information in orientation gradient fields, improving the recovery of local interfacial topology.
   • Mean variation ($V$): Used as a regularizer to enforce smoothness and control interface density without eliminating physically relevant intragranular features.

3. Optimization Strategy:
   The reconstruction is formulated as a differentiable inverse optimization problem. The microstructure is initialized as a random noise field and iteratively updated using gradient-based optimization. The framework utilizes the L-BFGS-B algorithm to navigate the complex loss landscape, computing gradients via automatic differentiation (TensorFlow) with respect to voxel-wise orientation components. For 3D reconstruction from 2D data, the loss function aggregates discrepancies across orthogonal planar 2D sections to enforce statistical consistency throughout the volume.

Key Results
The framework was validated through numerical studies and the reconstruction of experimental aluminum alloy microstructures processed via thermo-mechanical routes (Friction Extrusion).

• Descriptor Efficacy: Numerical benchmarks demonstrated that no single descriptor could independently recover the complexity of polycrystalline states. While the mean variation ($V$) regulated gradient density and two-point correlations ($S$) enforced global anisotropy, the hybrid three-point variogram ($\gamma_3$) was critical for defining sharp grain boundaries and local interfacial details. A combined approach yielded the highest fidelity.
• 2D Reconstruction: The framework successfully reconstructed statistically homogeneous 2D orientation maps, accurately reproducing equiaxed grain morphologies and crystallographic textures (verified via pole figures). The loss function exhibited stable, monotonic convergence, reducing residuals by four orders of magnitude.
• 3D Generative Synthesis: The method successfully generated 3D Representative Volume Elements (RVEs) from single 2D EBSD slices. The resulting 3D realizations displayed spatially coherent, equiaxed grain structures that preserved the 2D morphological characteristics and crystallographic texture, despite the lack of direct 3D experimental input.
• Inhomogeneous Fields: When applied to statistically inhomogeneous microstructures with spatial gradients, the framework successfully recovered global texture features. However, as expected with stationary statistical descriptors, the reconstruction homogenized local spatial gradients, replacing specific spatial evolutions with average correlation lengths.

Significance and Claims
The authors claim that this work provides a versatile, open-source extension to the MCRpy framework, enabling the direct, orientation-based reconstruction of polycrystalline materials. Key contributions include:

• Singularities-Free Optimization: By utilizing SHSH, the framework avoids the numerical instabilities associated with traditional Euler-based methods, making gradient-based optimization feasible for crystallographic data.
• Novel Descriptors: The integration of the hybrid three-point variogram and mean variation regularizer allows for the simultaneous capture of global texture and local interfacial topology, addressing limitations in previous descriptor-based approaches.
• Digital Synthesis: The methodology facilitates the rapid digital synthesis of polycrystalline RVEs, bridging the gap between 2D experimental data and 3D computational inputs required for full-field simulations.

The paper acknowledges current limitations, including the non-convex nature of the optimization problem which leads to sensitivity to initialization (requiring ensemble strategies for optimal results) and the inherent inability of stationary descriptors to capture non-stationary, spatially varying microstructures. Nevertheless, the work establishes a foundational methodology for the generative reconstruction of microstructures that reflect target statistical ensembles using limited 2D data.