Modeling MBE RHEED signals using PCA and neural networks

This paper introduces a novel technique for constructing an empirical model which relates RHEED intensity patterns to the physical characteristics of MBE grown thin films. A fractional factorial experiment is used to systematically characterize the growth of a five-layer, undoped AlGaAs-InGaAs single quantum well structure on a GaAs substrate as a function of time and temperature for oxide removal, substrate temperatures for AlGaAs and InGaAs layer growth, beam equivalent pressure of the As source and quantum well interrupt time. MBE growth takes place in a Varian Gen-II MBE system using substrate rotation, and RHEED signals are monitored for each experimental trial. RHEED pattern variation is used as an indicator of defect density, X-ray diffraction, and photoluminescence of the grown films. Principal component analysis is used to reduce the dimensionality of the RHEED data set, while maintaining the integrity of the information contained within. The reduced RHEED data set is used to train back-propagation neural networks to model the process responses. These models are quite accurate (about 3% RMSE on training data and less than 10% RMSE for test data), implying that the principal components are a reliable source of input data. Continued development will lead to models which provide a platform upon which to build an automated process control system for MBE growth