Control-relevant modeling of the Czockralski single crystal growth process using neural networks

Abstract

This paper present the first known attempt to empirically model and experimentally verify the growth of ilmenite single crystals using the Czockralski process which is an industrial crystal pulling process extensively used for silicon and germanium single crystal growth. The experimental apparatus for ilmenite growth process has been significantly improved allowing the performance of repeatable ilmenite growth experiments, and the acquisition of noise-free experimental data for empirical modeling. A feedforward multilayer perceptron is used to develop a single-step predictor, modeling the thermal response of the crystal growth process. The training of the neural network is performed using adaptive backpropagation. Testing of the predictor is performed using a series of additional single crystal ilmenite growth experiments. The comparison of the predictor response with experimental measurements demonstrates the good generalization capability of the neural network model.