Premium
Run‐to‐run process control of a plasma etch process with neural network modelling
Author(s) -
Card Jill P.,
Naimo Mark,
Ziminsky William
Publication year - 1998
Publication title -
quality and reliability engineering international
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.913
H-Index - 62
eISSN - 1099-1638
pISSN - 0748-8017
DOI - 10.1002/(sici)1099-1638(199807/08)14:4<247::aid-qre188>3.0.co;2-v
Subject(s) - wafer , artificial neural network , plasma etching , advanced process control , semiconductor device fabrication , control theory (sociology) , process control , metrology , process (computing) , multivariable calculus , heuristic , etching (microfabrication) , computer science , materials science , engineering , mathematics , control engineering , optoelectronics , statistics , nanotechnology , control (management) , artificial intelligence , layer (electronics) , operating system
Run‐to‐run control of a plasma etch process for 8 inch diameter silicon wafers at Digital Semiconductor is determined by maintenance of targeted values of post‐etch metrology variables. The post‐etch quality variables are extremely sensitive to variation in the etch chamber conditions due to fluctuation in temperature, pressure, gas flow rates, etc., the aging of process tool parts and time since last chamber cleans. The target parameter drifts due to part wear and impurity build‐up are long‐term and distinct in character from the more immediate action of chamber parameter disturbance. We have modelled the multivariable and multimodal action of output etch parameter variation using a neural network prediction model coupled with an heuristic non‐linear feedback loop to determine optimal corrective action to be performed between runs to ensure on‐going etch tool control. The neural networks exhibited overall accuracy to within 20% of the observed values of 0.986, 0.938 and 0.87 for the post‐etch rate, standard deviation and selectivity variables respectively. We present the neural model design with specially adapted input parameter transforms and results of training, test and validation sample fits for 148 monitor wafer records. The control unit was able to detect accurately the need for part replacements and wet clean operations in 34 of 40 operations and offered suggested process setting changes which were consistent with engineers' understanding of adjustments which would yield output parameters closer to the targeted values. Our current work extends previous modelling of the 6 inch wafer plasma etch process to 8 inch wafers with the addition of genetic algorithms for input variable selection and optimization algorithm improvements. In addition, we have installed a real‐time shop floor data collection utility with extensive data quality assurance, enabling the controller to be implemented on a run‐to‐run basis with neural network updates between monitor etch runs. In this way the tool operator gets feedback as to necessary process and part changes immediately following quality wafer checks and before production lots are processed. This speed of implementation is the first of its kind in the semiconductor fabrication process arena. © 1998 John Wiley & Sons, Ltd.