Variational Auto Encoder for automated structure segmentation of materials

Atomic-scale structural heterogeneity plays a crucial role in determining the properties of nanomaterials and complex energy systems. High-resolution STEM imaging provides critical insights into these structures, but conventional segmentation techniques often struggle with intricate patterns, relying on linear dimensionality reduction (e.g., NMF) and manual interpretation. In this study, we introduce an unsupervised deep learning framework that integrates Gabor filtering, VAE, and k-means clustering to systematically classify structural features in atomic-resolution images, with demonstrated applicability to different material systems. To demonstrate the versatility of this approach, we apply it to two distinct material systems: (1) polycrystalline metallic nanoparticles, where it successfully segments grain orientations and surface reconstructions, and (2) LLO, where it distinguishes C2/m and R-3 m phases. By leveraging a nonlinear latent space representation, our method provides a complementary framework to traditional linear techniques (e.g., NMF) for capturing complex structural variations, reducing subjective biases and enhancing scalability for automated materials characterization across diverse crystalline systems.