Machine Learning Hotspot Prediction Significantly Improve Capture Rate on Wafer

• Published in: 2020 International Workshop on Advanced Patterning Solutions (IWAPS) pp. 1 - 4
• Main Authors: Yuan, Wei, Lu, Yifei, Li, Ming, Pan, Bingyang, Gao, Ying, Tian, Yu, Li, Zhi-qin, Ji, Liang, Huang, Ying, Chen, Hao, Yao, Yueliang, Park, Sean
• Format: Conference Proceeding
• Language: English
• Published: IEEE 05.11.2020
• Subjects: hotspot prediction; image based; Machine learning
• DOI: 10.1109/IWAPS51164.2020.9286795

In a real mask tape-out (MTO) process, an end user would typically use simulation tools to capture hotspot candidates which are at risk of appearing on wafers. The tight turn-around-time(TAT) in a fab requires an efficient method to categorize these candidates and sampling before measurement. Traditionally, in order to capture hotspots, verification tools mainly focus on limited parameters such as contours, local image contrast and parameters extracted from the full aerial and resist information. This approach makes it difficult to quickly pinpoint high risk hotspots, especially when the hotspot count is large. In contrast, by using advanced machine learning techniques, Newron hotspot prediction is an innovative method that makes full use of whole simulated images to generate accurate prediction information for every hotspot candidate. Newron hotspot prediction is able to significantly reduce the amount of required input information and improve the hotspot capture rate.