Abstract
The Czochralski (CZ) crystallization process is used to produce monocrystalline silicon, which is used in solar cells and electronics. This paper considers adaptive system identification of the dynamics from the triode for alternating current (TRIAC) to the actual (measured) heating element power at the CZ process. The system identification work is based on a dataset logged at a real-life CZ process. Two approaches for adaptive system identification are considered: (i) The MATLAB command rarmax is used to tune all parameters of an ARMAX model. (ii) Only the gain of the model is tuned, while the pole, the zero, and the noise model are fixed. The adaptive gain can be computed using the recursive least squares method, which is simple to implement and has numerical advantages. These adaptive system identification approaches are compared to a non-adaptive ARMAX model. The performance criterion used is the mean squared one-step-ahead prediction error. The adaptive gain model performs much better than the non-adaptive model and the adaptive model which tunes all parameters. There are also other reasons why the adaptive gain approach is favorable.