Recognition and Location of Mixed-type Patterns in Wafer Bin Maps

Abstract

Wafer map analysis is one of the most critical steps for monitoring wafer quality and tracking failures in the semiconductor manufacturing process. Defective dies on wafer bin maps (WBMs) usually cluster into specific spatial patterns, which contain critical information for cause identification and yield improvement. This paper proposes an effective pattern recognition and location approach for both single pattern and mixed-type pattern maps. A convolutional neural network is introduced to classify WBMs with the single pattern or mixed-type (non-overlapped) patterns after seed filling segmentation. Meanwhile, WBMs with overlapped patterns are located and classified through template matching method. Further, for assessing the impact of each pattern, the pattern sizes are calculated after pattern location. The overall classification accuracy in two real-world datasets reaches 91.2% and 84.3%, which demonstrates the proposed approach has excellent classification performance for WBMs with mixed-type patterns.