Novel concept-oriented synthetic data approach for training generative AI-Driven crystal grain analysis using diffusion model

• Concept-oriented synthetic data method is illustrated to tackle the challenge of data scarcity. • Any required set of training data can be created from a set of concepts. • Method was used to train models to extract microstructure from microscopy images. • The framework integrates edge detection algorithms with generative machine learning diffusion models . • Trained model effectively identifies grain boundaries with accuracy of 97.23%. The traditional techniques for extracting polycrystalline grain structures from microscopy images, such as transmission electron microscopy (TEM) and scanning electron microscopy (SEM), are labour-intensive, subjective, and time-consuming, limiting their scalability for high-throughput analysis. In this study, we present an automated methodology integrating edge detection with generative diffusion models to effectively identify grains, eliminate noise, and connect broken segments in alignment with predicted grain boundaries. Due to the limited availability of adequate images preventing the training of deep machine learning models, a new seven-stage methodology is employed to generate synthetic TEM images for training. This concept-oriented synthetic data approach can be extended to any field of interest where the scarcity of data is a challenge. The presented model was applied to various metals with average grain sizes down to the nanoscale, producing grain morphologies from low-resolution TEM images that are comparable to those obtained from advanced and demanding experimental techniques with an average accuracy of 97.23 %. This represents a notable improvement, surpassing the accuracy of some of the state-of-the-art software-based methods for automated microstructure extraction, including neural network with convex hull and approximate contour (UNet + CHAC) (89 %), richer convolutional features (RCF) (91 %), and richer convolutional features with generative adversarial network (RCF + GAN) (93 %).