Anke Müller

Advisors
Philipp Schlieper (M.Sc.), Robert Richer (M.Sc.), Prof. Dr. Björn Eskofier

Duration
03/2019 – 09/2019

Abstract
After ‘Big Data’ has been a big buzz for a couple of years now, only a small impact in the semiconductor manufacturing industry can be noticed yet. Although the manufacturing industry generates about a third of all data today and one of the biggest contributors in that is the semiconductor manufacturing industry, the data volume is not yet effectively used [1]. As a small step towards using the data intelligently and improving the Overall Equipment Effectiveness (OEE) [4], this work aims at analyzing the chemical vapor deposition (CVD) tool in the process of wafer fabrication. During CVD chemical reactions of gaseous precursors in a heated environment result in the formation of a thin solid coating on a substrate [2]. The CVD tool can be regulated via three adjustable parameters: deposition time, flow of doping gas and temperature. While the first two parameters have a linear dependence to the resulting measurable coating thickness and the resistance, respectively, the temperature has no detected obvious primary linear dependence. However, it has a direct impact on the internal mechanical tensions on the wafer, called stress, which can be measured by means of scanning infrared depolarization (SIRD). Stress on a wafer increases the breakage probability during the integrated circuit (IC) manufacturing process [3] and hence has a direct impact on the quality, the operability and therefore on the OEE. Michael Splinter, Senior Vice President and General Manager of Intel’s technology and manufacturing group estimated a lost revenue of $100,000 for a downtime of a critical unit in the IC manufacturing process of an hour [1]. Although the manual adjustment of the tool cannot be categorized as equipment downtime, it also cannot be categorized as manufacturing time [4] and therefore eventually results in revenue loss. However, to this day the temperature is still set manually by experience-based knowledge. To optimize this production step machine learning is used to automate the tool parameter adjustments for the temperature.

Accordingly, the first step of this master’s thesis is to define a design of experiment to cover the design space for the stress measurement of the CVD tool optimally. Using this knowledge, a further step is to classify existing SIRD images by means of a machine learning algorithm, possibly a convolutional neural net or a recurrent neural network [8,9,10]. Afterwards, another machine learning algorithm, possibly a neural net or a recurrent neural network [8,9,10], is used for predicting the parameter settings for the front, side and back temperature of the CVD tool. The algorithms will be implemented in Python using Tensorflow [5] and Keras [6,7].

References:

1. Munirathinam, S., Ramadoss, B., Big Data Predictive Analytics for Proactive Semiconductor Equipment Maintenance, Proceedings – 2014 IEEE International Conference on Big Data, IEEE Big Data 2014
2. Choy, K.L., Chemical Vapor Deposition of Coatings, Progress in Materials Science, 2003
3. PVA TePla America, <https://www.pvateplaamerica.com/semi_stressimaging.php>, 2018 (29.01.2019
4. De Ron, A J, Rooda, J.E., OEE and equipment effectiveness E: an evaluation, International Journal of Production Research, 2005
5. Tensorflow, <https://www.tensorflow.org/guide/keras>, 2018 (29.01.2019)
6. Keras, <https://keras.io/>, 2018 (29.01.2019)
7. Gulli, A., Pal, S., Deep Learning with Keras, Packt Publishing Ltd., ISBN: 978-1- 78712-842-2, 2017
8. Bishop, C., Pattern Recognition and Machine Learning, Springer Science + Business Media, LLC, 2006
9. Goodfellow, I., Bengio, Y., Courville, A., Deep Learning, <www.deeplearning.org>, MIT Press, 2016
10. Nielsen, M., Neural Networks and Deep Learning, <neuralnetworksanddeeplearning.com>, 2015