Deep Learning Assisted Prediction of Reflection High‐Energy Electron Diffraction Patterns During Heteroepitaxy

Reflection high‐energy electron diffraction (RHEED) is an in situ probe for monitoring surface evolution during molecular beam epitaxy (MBE), yet coupling future‐pattern prediction with downstream recognition remains limited. Here, we present PredRHEED, a cascaded deep‐learning framework that combines RHEED‐pattern classification with future‐frame prediction during scandium aluminum nitride (ScAlN) growth in MBE. The framework combines a hybrid classifier composed of convolutional neural network (CNN) and Transformer, in which CNN extracts local diffraction features and Transformer encoder captures their global arrangement, to identify three RHEED‐pattern classes, streaky, transition, and spotty, while the prediction module employs a memory‐enhanced self‐attention convolutional long short‐term memory (MSAM‐ConvLSTM) predictor to forecast future RHEED frames. During inference, the predictor uses the preceding 15 s of RHEED data to predict the evolution of the RHEED pattern over the next 5 s. The predicted frames are then classified to determine the anticipated RHEED‐pattern class. Using 22,110 RHEED images acquired during a single continuous ScAlN growth experiment, the classification branch achieved 93.68% accuracy, whereas the prediction branch generated future frames with high structural fidelity. Across 50 classifier‐adaptation epochs, the maximum observed test‐set accuracy was 84.7% at epoch 11, while accuracy at the final epoch was 82.7%. The epoch‐wise test‐set trajectory is exploratory.