Hotspot Detection using Machine Learning

Kareem Madkour1, Sarah Mohamed1, Dina Tantawy2, Mohab Anis3
1Mentor Graphics, 2Cairo Univeristy, 3American University in Cairo


As technology nodes continues shrinking, lithography hotspot detection has become a challenging task in the design flow. In this work we present a hybrid technique using pattern matching and machine learning engines for hotspot detection. In the training phase, we propose sampling techniques to correct for the hotspot/non-hotspot imbalance to improve the accuracy of the trained Support Vector Machine (SVM) system. In the detection phase, we have combined topological clustering and a novel pattern encoding technique based on pattern regularity to enhance the predictability of the system. Using the ICCAD 2012 benchmark data, our approach shows an accuracy of 88% in detecting hotspots with hit-to-extra ratio of 0.12 which are better results compared to other published techniques using the same benchmark data.