Reflection high-energy electron diffraction (RHEED) is widely used because it allows in-situ observations of the surface behavior of a sample during molecular beam epitaxy (MBE) growth. In particular, RHEED patterns have been used for the calibration of growth conditions because they change considerably depending on the sample temperature, amount of material supplied, and supply ratio. However, RHEED pattern analysis depends on the expertise of the operator and there is also a time limit; therefore, it cannot be readily applied to real-time feedback control. In our previous study, we reported a model that acquires RHEED images of MBE GaAs grown on a GaAs substrate via machine learning using a convolutional neural network. However, the model that classifies the two patterns was a binary classification model and extensibility via data addition was limited. In this study, we report a classification model that can classify three or more classes. This suggests the possibility of a generalized RHEED pattern classifier.
Read full abstract