Machine Learning-Based Hotspot Detection: Fallacies, Pitfalls and Marching Orders

Extensive technology scaling has not only increased the complexity of Integrated Circuit (IC) fabrication but also multiplied the challenges in the Design For Manufacturability (DFM) space. Among these challenges, detection of design weak-points, popularly known as `Lithographic Hotspots', has attracted substantial attention. Hotspots are certain patterns which exhibit a higher probability of causing defects due to complex design-process interactions. Identifying such patterns and fixing them in the design stage itself is imperative towards ensuring high yield. In the early days of hotspot detection, Pattern Matching (PM) based methods were proposed. While effective in identifying previously known patterns, these methods failed to identify Never-Seen-Before (NSB) hotspots. To address this drawback, Machine Learning (ML) based solutions were introduced. Over the last decade, we have witnessed a plethora of ML-based hotspot detection methods being developed, each slightly outperforming its predecessors in accuracy and false-alarm rates. In this paper, we critically analyze the ML-based hotspot detection literature and we highlight common misconceptions which are found therein. We also pinpoint the underlying reasons that have led to these misconceptions by dissecting the ICCAD-2012 benchmark dataset, which has largely guided the evolution of this area, and revealing its limitations. Furthermore, we propose an enhanced version of this benchmark dataset, which we deem more appropriate for accurately assessing hotspot detection methods. Finally, we offer our suggestions to improve the effectiveness of ML-based Hotspot Detection methods and demonstrate about 5X reduction in false-alarms in comparison to the state-of-the-art.