Fault detection using random forest similarity distance

To maintain the pace of development set by Moore's law, semiconductor manufactures continue to shrink and redesign transistor architectures delivering better device performance. This has led to an increase in the complexity of the manufacturing process, where new technologies typically consist of several hundred processing steps. In this context, the impact of an incorrectly processed wafer progressing through the manufacturing process from start to finish can have a serious negative impact on profitability. In this paper, we demonstrate an unsupervised method based on random forests which can identify faulty wafers from the chemical signatures observed during a plasma etching process. The method is evaluated using both a simulated example and a real industrial dataset. Results show the correct identification of faulty wafers in both studies. The paper is concluded with a summary of research findings and a discussion on future work.