Physics-constrained EBSD image inpainting via adversarial graph learning: bridging crystallographic rules and multimodal deep learning

Electron Backscatter Diffraction (EBSD) is a crucial characterisation method in materials engineering. The reliability of EBSD data is essential in the aerospace, nuclear, and automotive industries, as material performance greatly affects operational safety. While industrial practice makes perfect EBSD data difficult, with sample preparation errors, beam drift, and instrumental noise corrupting up to one-third of datasets. Automated crystallographic fidelity restoration solutions are needed because corrupted data force engineers to abandon valuable experiments or manually restore datasets at risk of errors. Current image inpainting techniques fail to maintain crystallographic constraints, resulting in restorations that violate the basic rules for crystalline materials. A novel physics-constrained framework is proposed to fill this gap. It integrates adversarial learning with graph neural networks (GNNs) for crystallographically consistent EBSD image inpainting.The proposed GTRG method consists of three elements: (Formula presented) a generative adversarial network (GAN) for reconstructing grain boundaries; (Formula presented) a crystallography-guided graph transformer (T) that converts pixel data into orientation-boundary graphs; and (Formula presented) a regression graph convolutional network (RGCN) that links grain, orientation and boundaries to predict missing crystal orientations. The framework mandates a single orientation per grain and preserves grain boundary structure through structured graph representations. A strategy for creating automated EBSD datasets that incorporates realistic corruption patterns supports effective model training and evaluation. Experimental validation shows better performance than current methods, with a 3.5 % improvement in SSIM (0.950 vs. 0.918) and a 63.0 % reduction in FID (16.55 vs. 44.70) compared to AOT-GAN. The study on aerospace niobium alloys further validates practical utility, showing statistically consistent grain orientation and size distributions (Kolmogorov-Smirnov (Formula presented), (Formula presented) ). This work introduces two key advancements: 1) the first integration of graph neural networks with adversarial learning for topology-aware image inpainting, and 2) a physics-informed framework bridging computer vision and materials science, enabling effective restoration of corrupted EBSD data for subsequent engineering applications.