Using neural networks to construct models of the molecular beamepitaxy process

摘要

This paper presents the systematic characterization of thenmolecular beam epitaxy (MBE) process to quantitatively model the effectsnof process conditions on film qualities. A five-layer, undoped AlGaAsnand InGaAs single quantum well structure grown on a GaAs substrate isndesigned and fabricated. Six input factors (time and temperature fornoxide removal, substrate temperatures for AlGaAs and InGaAs layerngrowth, beam equivalent pressure of the As source and quantum wellninterrupt time) are examined by means of a fractional factorialnexperiment. Defect density, X-ray diffraction, and photoluminescence arencharacterized by a static response model developed by trainingnback-propagation neural networks. In addition, two novel approaches forncharacterized reflection high-energy electron diffraction (RHEED)nsignals used in the real-time monitoring of MBE are developed. In thenfirst technique, principal component analysis is used to reduce thendimensionality of the RHEED data set, and the reduced RHEED data set isnused to train neural nets to model the process responses. A secondntechnique uses neural nets to model RHEED intensity signals as timenseries, and matches specific RHEED patterns to ambient processnconditions. In each case, the neural process models exhibit goodnagreement with experimental results