Deep Neural Network Application to Real-Time Growth Prediction of Epitaxial GaN by Molecular Beam Epitaxy

In this work, we explore the application of machine learning(ML) methods, specifically recurrent neural networks (RNNs), for the analysis and prediction of GaN growth via molecular beam epitaxy (MBE). In particular, we show the possibility to control in real-time and predict the growth evolution based on the actual values of selected parameters. For this purpose, we tailored a specific GaN growth sequence, derived from droplet epitaxy and metal modulated growth; each deposition is characterized using the surface morphology information contained in the ereflection high-energy electron diffraction (RHEED) images together with relevant MBE experimental parameters, such as nitrogen flux, power, and cell temperature. The classification and prediction nets work collecting data on an input window of about 10 s and provides a good prediction accuracy over a similar time horizon, with a small sized, fast RNN. The simplicity of this net allows for real-time control of the MBE growth, while at the same time providing an early alarm system for unwanted secondary phases through the growth predictions.