Machine-learning-based identification of defect patterns in semiconductor wafer maps: An overview and proposal

Wafers are formed from very thin layers of a semiconductor material, hence, they are highly susceptible to various kinds of defects. The defects are most likely to occur during the lengthy and complex fabrication process, which can include hundreds of steps. Wafer defects are generally caused by machine inaccuracy, chemical stains, physical damages, human mistakes, and atmospheric conditions. The defective chips tend to have several unique spatial patterns across the wafer, namely ring, spot, repetitive and cluster patterns. To locate such defect patterns, wafer maps are used to visualize and ultimately lead to better understanding of what happened during the process failure. To identify the unique patterns of defects and to find the point of manufacturing process that causes such defects accurately, nature-inspired model-free machine-learning techniques have been well accepted. This paper thus reviews the theoretical and experimental literature of such models with a focus on model learnability and efficiency-related issues involving data reduction and transformation techniques, which could be seen as the key model properties to deal with big data applications.