C 4-Equivariant Flow Matching on Anisotropic Power-Diagram Graphs for Microstructure Generation

arXiv:2610.11549 · cs.LG · Submitted 2026-10-08 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: I'm Tom, and with me are Jane, Lu, senior AI researcher at Tsinghua, Meng, lead engineer at a mysterious AI startup and Lalam, the in-house Large Language Model.

Jane: Today's paper: "C 4-Equivariant Flow Matching on Anisotropic Power-Diagram Graphs for Microstructure Generation".

Tom: The gist The model introduces a generative model for synthesising realistic polycrystalline microstructures using flow matching and graph neural networks.

Jane: First, who's behind it and why it matters.

Paper summary: Tom: So wrapping up "C four-Equivariant Flow Matching on Anisotropic Power-Diagram Graphs for Microstructure Generation," the core of the work is using flow matching over anisotropic power diagrams to generate realistic microstructures, while ensuring rotational symmetry through a C four-equivariant graph architecture <ref:2610.11549#pg1,Equivariant Flow Matching on Anisotropic Power-Diagram Graphs for Microstructure Generation>.

Jane: What this means simply is that they’ve built a system where you can generate detailed polycrystalline samples at any resolution and guide them toward specific patterns without needing to retrain the underlying model for every new design goal.

Lu: The main contribution is clearly that formulation of microstructure generation as flow matching over the compact geometric parameters of an anisotropic power diagram, and introducing this C four-equivariant architecture that models dependencies between grains <ref:2610.11549#pg1>.

Meng: The implications are that researchers can use this to quickly explore the vast space of possible material structures, which is a huge time saver when you're working on material science problems.

Lalam: And for the broader culture of AI in design, it shows how these generative models can be adapted to help create complex physical realities based on user-defined objectives.

Tom: It’s about taking something usually done by expensive, slow equipment and making it accessible through a flexible generative framework that respects the underlying physics of rotation.

Conclusion: Tom: So we're wrapping up on "C four-Equivariant Flow Matching on Anisotropic Power-Diagram Graphs for Microstructure Generation." Essentially, they built a way to generate realistic microstructures using flow matching and graph neural networks, all while keeping the shapes consistent under four times nine zero degree rotations.

Jane: That’s right. They use these power diagrams to describe grains and then apply this flow matching technique to turn random noise into those specific structures. It sounds complicated, but the goal is just to create detailed patterns for materials science simulations or even manufacturing design.

Lu: What I find really interesting is the way they handle the geometry, using anisotropic power diagrams instead of just simple grids. It makes the representation much more compact and handles small pixel errors better during generation.

Meng: From an engineering standpoint, having a model that understands those geometric relationships directly, rather than just pixels, means you can generate high-fidelity samples at any resolution without losing structural integrity. That’s practical for testing things out quickly.

Lalam: And because of the C four-equivariant part, it means if you rotate your starting point or your desired output by a quarter turn, the generated structure rotates perfectly with it. It respects the underlying physics of rotation in the material itself.

Tom: Exactly. So why does this matter for us? It means we can use these tools to explore incredibly complex material landscapes much faster than traditional methods allow us to design things just by looking at them.

Jane: Right, and they showed how you can control these generations without retraining the whole model by using something called training-free guidance. You can tell the AI exactly what shape you want it to make, like a layered structure or a cast slab.

Lu: That's because they use these differentiable objectives—things like Jlayer or Jcast—to nudge the generation process toward your specific design goal, and the model parameters stay fixed. It’s super flexible.

Meng: So I get that you can use this to generate something specific, say a weld pattern, with a defined weight distribution in one area and another in another area without having to teach the AI a whole new dataset for every single weld design. That's really powerful for iterative design cycles.

Lalam: It means the culture around materials discovery could shift toward using these generative models as creative partners, letting us rapidly prototype structures based on physical constraints and desired performance metrics.

Tom: So, to recap, they gave us a flow matching system that generates microstructures on power diagrams, ensures rotational symmetry with C four equivariance through graph networks, and lets us control the output without retraining the AI using training-free guidance objectives.

Jane: And what this means is we have a much more robust and controllable way to design complex physical materials digitally. Now we’re going to look at how they tested all these different control functions in action.

Dawid Lipinski, Jixiang Qing, Henry Moss

School of Mathematical Sciences, Lancaster University

cs.LG

Submitted: 2026-10-08

Updated: 2026-10-08

Comments: Accepted at the NeurIPS 2026 workshops on Representations for the Physical Sciences and Geometric Distributional Deep Learning. 12 pages, 2 figures, 1 table

Code: https://github.com/DawidLipin/Flow_APDs

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 78/100

The gist: The gist The model introduces a generative model for synthesising realistic polycrystalline microstructures using flow matching and graph neural networks.

Key concepts

Anisotropic Power Diagrams (APDs): Microstructures are represented as APDs, which are compact geometric descriptions that capture the grain positions in a 2D space and the anisotropy of each grain using an ellipse parametrization. This representation is resolution-independent and helps regularize small segmentation errors.
C4-Equivariant Flow Matching: This method trains a continuous flow to transform simple noise into target microstructures. The architecture is C4-equivariant, meaning the generated output rotates consistently when the input rotates by multiples of 90 degrees, ensuring geometrically sound results during generation.
Training-Free Guidance Framework: This technique allows users to control the generated microstructure without retraining the model. By modifying the velocity field based on a differentiable objective function (like Jlayer or Jcast), samples are biased toward desired features in real-time.

Terminology

Flow Matching — 1× in this paper · explained in 15 episodes

Summary

The gist The model introduces a generative model for synthesising realistic polycrystalline microstructures using flow matching and graph neural networks.

Microstructure Representation

We represent microstructures as anisotropic power diagrams (APDs) to learn a compact geometric parametrisation that is independent of raster resolution and regularises small pixel-level segmentation artefacts. Each grain is represented by its position in [0, 1]2 and an ellipse-based parametrisation of its anisotropy.

C4-Equivariant Flow Matching

Flow matching is used to train a continuous normalising flow to transport samples from a simple source distribution p0, typically Gaussian noise, to a target data distribution p1. Samples are generated by numerically integrating the learned vector field from t = 0 to t = 1. The model architecture is a graph-based, C4-equivariant architecture that models dependencies between grains and ensures consistent generative behaviour under rotations by multiples of 90◦. The invariant edge features supplied to the message-passing network are defined as phiij = dij2, wi, wj, ai2, aj2, a top i aj, ai x aj, a top i rho(dij), a top j rho(dij), b(xi), b(xj). The output head predicts C4-invariant scalar coefficients multiplying geometric bases that transform equivariantly under 90◦ rotations. For any 90◦ rotation R, we have vθ(Rz, t) = R vθ(z, t), so rotating the initial state produces the correspondingly rotated generated APD.

Training-Free Guidance Framework

We introduce a training-free guidance framework for controlling generated microstructures through user-defined differentiable objectives. Following gradient-based guidance methods for diffusion and flow models, we bias generation toward a differentiable objective J by instead sampling pJ1(z1) ∝ p1(z1) exp[-βJ (z1)]. At time t, we estimate the terminal state by bz1 = zt + (t − t)vθ(zt, t) and modify the velocity as veθ(zt, t) = vθ(zt, t) − s(t)β∇ztJ (bz1), where s(t) is a time-dependent schedule. The model parameters remain fixed, enabling training-free control through arbitrary differentiable APD objectives.

Guidance Objectives for Specific Microstructures

We demonstrate how training-free guidance can be used to generate complex microstructures, based on user defined objective function. The four columns in Figure 2 are independent samples from the same pretrained model, using various guidance functions. Specific objectives include Jlayer for layered structures, J cast for cast slabs, Jweld for welds, and Jhl for heterogeneous lamella titanium. For the heterogeneous-lamella objective, we use a zero-mean two-level relative-weight target where C contains the NC largest centred weights and F = 1,..., N. The cross-entropy places the current coarse population in the bands, SMSE maintains separated coarse and fine weight populations without fixing grain labels, and the shape term weakly aligns coarse grains with the local band tangent while driving fine grains towards equiaxed shapes.

Anisotropy Matrix Parametrisation

We construct an unconstrained, reversible parametrisation in which any network output corresponds to a valid anisotropy matrix. Since Ai is SPD, its matrix logarithm Si = log Ai = Qi log(Λi)Q T i is real and symmetric. We additionally separate anisotropic shape from isotropic scale by constraining det Ai = 1. This constraint implies tr(Si) = 0, resulting in Si being a symmetric trace-free 2 × 2 matrix written using only two unconstrained scalars ai,1 and ai,2. The inverse transformation is simply Ai = exp(Si).

C4-Equivariant Model Updates

We construct a velocity field satisfying vθ(Rz, t) = R vθ(z, t) for every rotation R ∈ C4. We separate the update into two components: the message-passing network operates on C4-invariant scalar features and predicts invariant scalar coefficients. Second, these coefficients multiply fixed geometric basis vectors with known transformation properties. The contribution of edge e to the positional velocity is mx e = γe (αe,1de + αe,2xs + αe,3Ssde + αe,4Srde). The node velocities are obtained by aggregating all incoming edge contributions. Summation preserves the transformation properties of the individual messages.

Experimental Validation and Future Work

The model is currently trained only on synthetic APDs, so fidelity to higher-order experimental statistics remains unvalidated. Future work will aim to evaluate and fine-tune the model on EBSD data, incorporate crystallographic orientation, and extend the representation to three-dimensional microstructures. Exploring alternative guidance objectives, particularly those based on physics-informed loss functions, is also a promising direction for future research. The model can render generated samples at arbitrary pixel resolution. The code can be found at https://github.com/DawidLipin/Flow APDs. All samples in the figure use N = 300 generators, k = 16 nearest neighbours, and K = 60 Euler steps with tn = n/K. The training set contained 12,000 synthetic APDs with between 100 and 500 grains, generated using PyAPD [2].

Improvements for AI systems

textbf Learned Microstructure Generation from Noise Flow Matching (FM) over APDs and C4-Equivariant Graphs: Improved Sample Fidelity and Resolution Independence. The system can generate realistic polycrystalline microstructures like copper weld, cast metal slab, 3D-printed stainless steel and heterogeneous lamella titanium by formulating microstructure generation as flow matching over the compact geometric parameters of an anisotropic power diagram. This allows for resolution independence because the representation is independent of raster resolution and regularises small pixel-level segmentation artefacts.
textbf Training-Free Control over Global Microstructure Morphology: Enables Complex Design Specifications. Using a training-free guidance framework, the system can impose complex, spatially varying global structure based on user-defined objectives like those for a copper weld or 3D-printed stainless steel, without retraining the model. This means users can dictate specific features, such as grain alignment or spatial patterns extending across the entire microstructure.
textbf C4-Equivariant Geometric Reasoning: Ensures Physically Consistent Rotational Symmetry in Generation. The architecture enforces rotational symmetry by ensuring that rotations of the input noise produce corresponding rotations of the generated microstructure. This is achieved by having a C4-equivariant architecture where nodes and edges are defined using features like b(u) = (u21 u22, u21 + u22, u41 + u42) encodes position relative to the square boundary while remaining C4-invariant.
textbf Robust Anisotropy Representation: Preserves Mathematical Validity of Grain Geometry. The model uses an unconstrained, reversible parametrisation for anisotropy matrices where any network output corresponds to a valid anisotropy matrix, and it imposes the constraint that det Ai = 1 by utilizing the identity that allows the network to predict parameters while ensuring validity is preserved by construction.
textbf Controllable Feature Manipulation via Multi-Objective Guidance: Fine-Tuning Specific Microstructure Statistics. The system can achieve precise control over specific aspects of the microstructure by using multiple, weighted objectives, such as controlling grain density or relative weight profiles like those defined in the Cast slab or Heterogeneous-lamella structure objectives. This allows for targeting features like grain-density target assigns fractions fch = 0.34, fcol = 0.48, and fcore = 0.18 to the three zones.

Abstract

Acquiring realistic microstructure data through Electron Backscatter Diffraction (EBSD) is costly and time consuming, often relying on specialised equipment. As microstructures strongly influence material properties, generating realistic samples is essential for modelling the behaviour of polycrystalline materials. We introduce a generative model for synthesising realistic polycrystalline microstructures using flow matching and graph neural networks. By representing microstructures as anisotropic power diagrams, our model learns a compact geometric parametrisation and can render generated samples at arbitrary pixel resolution. A C 4-equivariant architecture incorporates rotational symmetry directly into the model, ensuring that rotations of the input noise produce corresponding rotations of the generated microstructure. We also demonstrate how training-free guidance can be used to generate complex microstructures, based on user defined objective function. In particular, we generate microstructures resembling a copper weld, cast metal slab, 3D-printed stainless steel and heterogeneous lamella titanium.