Learning Metal Microstructural Heterogeneity through Spatial Mapping of Diffraction Latent Space Features

To leverage advancements in machine learning for metallic materials design and property prediction, it is crucial to develop a data-reduced representation of metal microstructures that surpasses the limitations of current physics-based discrete microstructure descriptors. This need is particularly relevant for metallic materials processed through additive manufacturing, which exhibit complex hierarchical microstructures that cannot be adequately described using the conventional metrics typically applied to wrought materials. Furthermore, capturing the spatial heterogeneity of microstructures at the different scales is necessary within such framework to accurately predict their properties. To address these challenges, we propose the physical spatial mapping of metal diffraction latent space features. This approach integrates (i) point diffraction data encoding via variational autoencoders or contrastive learning and (ii) the physical mapping of the encoded values. Together these steps offer a method offers a novel means to comprehensively describe metal microstructures. We demonstrate this approach on a wrought and additively manufactured alloy, showing that it effectively encodes microstructural information and enables direct identification of microstructural heterogeneity not directly possible by physics-based models. This data-reduced microstructure representation opens the application of machine learning models in accelerating metallic material design and accurately predicting their properties.

💡 Research Summary

The paper introduces a novel data‑driven framework for representing metallic microstructures that overcomes the limitations of conventional physics‑based descriptors, especially for complex hierarchical structures produced by additive manufacturing (AM). The core idea is to encode point‑wise electron backscatter diffraction (EBSD) Kikuchi patterns directly into a low‑dimensional latent space using either a variational autoencoder (VAE) or a VAE augmented with contrastive learning (SimCLR). The encoder‑decoder architecture consists of convolutional neural networks; the encoder compresses each 480 × 480 pixel pattern into a latent vector of 16–256 dimensions, while the decoder reconstructs the pattern to enforce fidelity. A pixel‑wise L2 reconstruction loss combined with a Kullback‑Leibler (KL) divergence regularizes the latent distribution. In the contrastive variant, three realistic augmentations (Gaussian noise, salt‑and‑pepper noise, gamma intensity alteration) are applied to the same pattern, and a contrastive loss term (weighted 20× relative to KL) forces the representations of the augmented views to be close, thereby making the latent space robust to acquisition noise.

After training on 96 000 randomly selected Kikuchi patterns from two alloys, the encoder is applied to full‑field EBSD maps (≈1 mm² area). Each spatial point receives a latent vector, which is then mapped back onto the physical grid, producing a “latent‑space map” of the microstructure. Because the latent vector retains all information embedded in the raw diffraction pattern—not just orientation or band positions—it captures subtle features such as lattice strain, dislocation density, cellular sub‑structures, and chemical fluctuations. Visualization can be performed by coloring each point according to selected latent dimensions or by applying dimensionality‑reduction techniques (t‑SNE, UMAP) to reveal clusters.

The methodology is demonstrated on wrought, fully recrystallized Inconel 718 and on an as‑built AM‑processed Inconel 718. Conventional inverse pole figure (IPF) maps show the wrought alloy as equiaxed grains with uniform orientation, while the AM alloy exhibits elongated grains, high fractions of low‑angle grain boundaries, and strong orientation gradients. Latent‑space maps of the wrought sample display homogeneous color fields, confirming microstructural uniformity. In contrast, the AM sample’s latent maps reveal pronounced spatial heterogeneity: distinct clusters correspond to regions of high dislocation density (sharpness), cellular structures, and localized lattice rotations. The contrastive VAE (SimCLR‑enhanced) achieves lower reconstruction error and higher robustness to noise than the plain VAE, especially when the latent dimensionality is set to 64–128, where the trade‑off between information retention and over‑fitting is optimal.

Key insights include: (1) Encoding the full diffraction pattern preserves far more information than extracting a handful of physical descriptors; (2) Spatial mapping of latent features provides a direct, high‑resolution picture of microstructural heterogeneity across multiple length scales; (3) The latent vectors can serve as compact, physics‑agnostic descriptors for downstream machine‑learning models that predict mechanical properties such as yield strength, fatigue life, or creep resistance. The authors argue that integrating these latent‑space descriptors into regression or graph‑network models will enable accurate property prediction for alloys with complex, multiscale microstructures, accelerating materials design cycles.

Overall, the work establishes a powerful pipeline—diffraction pattern → latent space → physical map—that bridges experimental microscopy and data‑driven materials science, offering unprecedented sensitivity to subtle microstructural features and a scalable route to incorporate full‑field diffraction data into predictive modeling frameworks. Future directions suggested include extending the approach to other alloy systems, coupling with other characterization modalities (e.g., X‑ray tomography), and implementing real‑time latent‑space mapping for in‑situ process monitoring in additive manufacturing.