Data-Driven Oxide Problem Characterization and Optimization in Float-Zone Silicon Crystal Growth Production

It is undisputed that silicon wafers have become crucial to our modern lives and world’s commercial and military applications. As the key process for the fabrication of silicon wafers, the single-crystal growth process has been driven to increase good-for-order single-crystal silicon yield while keeping costs low. To this end, this dissertation, conducted as part of DIGIMAN4.0 funded by Horizon 2020, the EU Framework Programme for Research and Innovation, collaborating
with Topsil Global Wafers, delves into the process optimization in Float-Zone (FZ) crystal growth production, with the aim of increasing yields and reducing costs. The core of the project introduces a unique challenge, an oxide contamination problem happening in the FZ process which is strongly associated with crystal yields, and a systematic investigation of the oxide problem around three topics: what the problem is (know-what), why it happened (Know-Why) and how to
address it (Know-How) is conducted.

Firstly, an investigation was carried out regarding the nature of the surface anomaly by material characterization and visual characterization. Specifically, material characterization was achieved by the characterization of the polysilicon surface using Focus Ion Beam Scanning Electron Microscopy (FIB-SEM) and Energy Dispersive Spectroscopy (EDS), which clearly defined that the surface anomaly is associated with enrichment of oxygen and oxygen loss by evaporation. Visual characterization was performed on the FZ images captured from the FZ vision system, which showed that the surface typically appears at the beginning of the cone phase, and the surface anomaly can present in three categories, including the spot, shadow, and ghost curtain and their characteristics and potential impacts were discussed.

Secondly, to enable an efficient recognition of the oxide layer (surface anomaly) for Know-What, an oxide identification based on Deep Learning was developed, which has been shown to be effective in capturing the occurrence of oxide without involving any human being, thus laying a foundation for automatic responses of oxide. In addition, in order to build the trust on the developed oxide identification, Grad-CAM was employed to increase the transparency on the model and to explain why the model makes such a prediction.

Targeted at Know-Why, a thorough investigation of the relationship between oxide and other data sources was carried out by means of Association Rule Mining. The results demonstrated that the oxide is strongly associated with a high moisture level in the FZ chamber during the FZ process.

In order to uncover the source of the high moisture level, a Deep Learning-based multi-modal moisture predictor was established, and a model explainability analysis was conducted on the prediction model to examine its decision-making process. This can provide us with insights into corrective measures.

Finally, an automatic response conceptual framework for mitigating oxide formation was proposed for Know-How to provide constant monitoring and dynamic adaption. The framework integrates the findings from Know-What and Know-Why in the diagnostic strategy for responding to the oxide problem. In addition, the framework considered prognostic strategy, which relies on the capability of Know-When in forecasting the occurrence of the oxide problem, thus allowing the decision maker to take preventive measures to decrease the probability of the oxide occurrence. In conclusion, with the motivation of increasing crystal yields and reducing costs in the FZ process, this dissertation focuses on process optimization by undertaking a systematic investigation regarding an oxide problem that can affect crystal yields. The investigation offers a omprehensive understanding of the oxide problem from three perspectives: Know-What, Know-Why and Know-How, which sets the stage for future advancement for the field.