Peak Sequence Transformer: A Real-Time AI RHEED Monitoring Framework for Ferroelectric Nitride MBE Growth

Reflection high-energy electron diffraction (RHEED) is a cornerstone technique for in situ monitoring of surface crystallography, reconstruction transitions, thin-film roughness, and growth mode during materials growth by molecular beam epitaxy (MBE). However, current analysis workflows rely heavily on manual interpretation and often neglect the temporal continuity feature of material growth. Here, we present the Peak Sequence Transformer (PS-Transformer), a lightweight temporal modeling framework that combines a physics-inspired peak-aware encoder, which achieves a 2490:1 reduction in input dimensionality, with Transformer-based attention to capture temporal evolution. Initial frame-level labels are generated from manually assigned start and end scores with linear interpolation and subsequently refined through pseudolabel self-training, enabling robust performance with minimal manual effort. When evaluated on ferroelectric nitride RHEED data sets, PS-Transformer achieves 99.87% classification accuracy, 105.3 frames per second inference speed, and 75.51% lower power consumption compared to three-dimensional (3D) convolutional neural network (CNN) baselines, while requiring substantially fewer parameters and floating-point operations. This physics-guided framework provides a practical and generalizable solution for real-time RHEED video analysis in semiconductor manufacturing and high-throughput research.