Direct RHEED Video Learning and Prediction Using a Hybrid Video Transformer–U-Net Framework for Molecular Beam Epitaxy

Abstract

Reflection high-energy electron diffraction (RHEED) is a key in situ growth monitoring technique for molecular beam epitaxy (MBE). In conventional practice, only a limited portion of the information contained in RHEED patterns is analyzed during MBE growth due to the rich features in RHEED patterns. Machine-learning (ML)-assisted analysis has been employed to address this limitation, and more useful knowledge about thin-film growth has been extracted from RHEED patterns. Most existing approaches, however, rely on extracted RHEED features rather than directly learning from a spatiotemporal RHEED video. Here, we demonstrate the future-frame prediction of RHEED using a hybrid transformer framework that allows direct RHEED video learning and prediction. The predicted RHEED patterns match the ground-truth well. This study establishes a foundation for predictive MBE monitoring and advances ML-assisted RHEED-pattern analysis. The limitations of the framework are also discussed.