Semi-Supervised Vision Transformer Framework for AI-Based RHEED Image Classification of Ferroelectric Nitride MBE Growth

We report a semi-supervised Vision Transformer (ViT) framework for automated reflection high energy electron diffraction (RHEED) image classification of ferroelectric nitride (ScAlN) materials grown by molecular beam epitaxy (MBE). By incorporating pseudo-labeling, the new framework reduces the need for extensive manual annotations while maintaining robust performance across multiple substrate angles. The effects of three key hyperparameters: labeled-data proportion, the number of Transformer heads, and model depth on classification outcomes are explored. Our findings show that, although parameter tuning can yield incremental accuracy gains, simpler configurations (e.g., two heads and two layers) provide an optimal balance between accuracy and computational overhead. Adjusting embedding dimensions further refines the model without incurring excessive computational costs. Compared with fully supervised approaches, the proposed framework delivers equal or better accuracy using far fewer labeled samples and effortlessly adapts to diverse RHEED angles. These findings underscore the potential of semi-supervised ViT-based solutions to facilitate AI-driven standardization and optimization in semiconductor manufacturing.