In recent years, we have witnessed many excellent machine learning (ML) solutions targeting circuit layouts with the emerging of AI. Researchers have developed more and more ML models for hotspot detection in chip layout. These ML models have demonstrated super potential in classification and can provide very fast predictions on various design objectives. However, almost all existing ML solutions for hotspot detection have neglected the basic interpretability requirement from potential users that could cause taping-out failure. As a result, it is very difficult for users to figure out any potential accuracy degradation or abnormal behaviors of given ML models.

In this work, a new technique named APPLE1 is proposed to explain each ML prediction at the resolution level of circuit elements...[ Read more ]

In recent years, we have witnessed many excellent machine learning (ML) solutions targeting circuit layouts with the emerging of AI. Researchers have developed more and more ML models for hotspot detection in chip layout. These ML models have demonstrated super potential in classification and can provide very fast predictions on various design objectives. However, almost all existing ML solutions for hotspot detection have neglected the basic interpretability requirement from potential users that could cause taping-out failure. As a result, it is very difficult for users to figure out any potential accuracy degradation or abnormal behaviors of given ML models.

In this work, a new technique named APPLE1 is proposed to explain each ML prediction at the resolution level of circuit elements and we also employ guided backpropogation method to interpret convolutional neural network (CNN) model’s focus region by leveraging gradient-based analysis. According to current knowledge, this is the first effort to explain ML predictions on circuit layouts. This framework provides a significantly more reasonable, useful, and efficient explanation for lithography hotspot prediction, compared with the highest-cited prior solution for natural images. And this holds significant practical value such as preventing backdoor attack.