InfinityEBSD : Metrics-Guided Infinite-Size EBSD Map Generation With Diffusion Models

TL;DR

InfinityEBSD addresses the challenge of generating large, statistically representative EBSD maps by combining a latent diffusion model with a VAE-based latent space and metric-conditioned cross-attention. The method supports extending existing EBSD maps and generating entirely new maps from eight microstructural descriptors, using a patch-wise extension strategy to maintain grain boundary continuity. Quantitative validation shows that generated maps preserve target metric distributions and exhibit spatial coherence, while exports to CTF enable immediate post-processing in MTEX or DIGIMU. This approach enables cost-efficient, scalable synthetic microstructure data for simulations, benchmarking, and virtual materials design, with potential extensions to 3D generation from 2D data.

Abstract

Materials performance is deeply linked to their microstructures, which govern key properties such as strength, durability, and fatigue resistance. EBSD is a major technique for characterizing these microstructures, but acquiring large and statistically representative EBSD maps remains slow, costly, and often limited to small regions. In this work, we introduce InfinityEBSD, a diffusion-based method for generating monophase realistic EBSD maps of arbitrary size, conditioned on physically meaningful microstructural metrics. This approach supports two primary use cases: extending small experimental EBSD maps to arbitrary sizes, and generating entirely new maps directly from statistical descriptors, without any input map. Conditioning is achieved through eight microstructural descriptors, including grain size, grain perimeter, grain inertia ratio, coordination number and disorientation angle distribution, allowing the model to generate maps that are both visually realistic and physically interpretable. A patch-wise geometric extension strategy ensures spatial continuity across grains, enabling the model to produce large-scale EBSD maps while maintaining coherent grain boundaries and orientation transitions. The generated maps can also be exported as valid Channel Text Files (CTF) for immediate post-processing and analysis in software such as MTEX or simulation environments like DIGIMU. We quantitatively validate our results by comparing distributions of the guiding metrics before and after generation, showing that the model respects the statistical targets while introducing morphological diversity. InfinityEBSD demonstrates that diffusion models, guided by physical metrics, can bridge the gap between synthetic and realistic materials representation, paving the way for future developments such as 3D realistic microstructure generation from 2D data.

Figures (32)

• Figure 1: Overview of the proposed method for extending EBSD maps from small ones (512×512 cells) to large generated maps (2048×2048 cells). The generated microstructures follow statistical and physical characteristics of the input microstructures. Map (a) corresponds to Inconel 718 (In718), a nickel-based superalloy and Map (b) corresponds to an austenitic stainless steel (316L).
• Figure 2: Overview of the denoising diffusion process, the RP. Previously the FP gradually corrupts a clean EBSD map into noise by sequentially adding Gaussian perturbations until it becomes fully noisy. Then, as the figure shows, in the RP we iteratively recover the original structure through denoising steps.
• Figure 3: Illustration of the EBSD data processing pipeline. (a) Cropping of the original EBSD map into patches of $512\times512$ cells. (b) Example of a raw EBSD map. (c) The same map from (b) after pre-processing, including twin removal and the application of disorientation angle and grain size thresholds. (d) Examples of data augmentation applied to the patches, showing 90°, 180°, and 270° rotations.
• Figure 4: Overview of the VAE workflow. The input EBSD map is compressed by the Encoder to a latent space $Z$, and then reconstructed by the Decoder.
• Figure 5: Latent encoding of the inputs used in the diffusion model. a) The full EBSD map $\mathbf{X}$ is encoded by the VAE encoder and perturbed with Gaussian noise to produce the noisy latent $\mathbf{Z_{\text{noisy}}}$. b) The partial (masked) EBSD map $\mathbf{X_m}$, corresponding to the known visible region, is also encoded into its own latent representation $\mathbf{Z_m}$. c) The binary mask $\mathbf{M}$ is downsampled to match the latent resolution.