Hotspot detection using squish-net

Abstract

Design-process weakpoints also known as hotspots cause systematic yield loss in semiconductor manufacturing. One of the main goals of DFM is to detect such hotspots. For the application of AI in hotspot detection, a variety of machine learning-based techniques have been proposed as an alternative to time expensive process simulations. Related research works range from finding efficient layout representations and features and developing reliable machine learning models. Main stream layout representations include density-based feature, pixel-based feature, frequency domain feature, concentric circle sampling (CCS) and squish pattern. However most of them are either suffering from information loss (e.g. density-based feature, and CCS), or not storage efficient (e.g. images). To address these problems, we propose a convolutional neural network called Squish-Net where the input pattern representation is in an adaptive squish form. Here, the squish pattern representation is modified to handle variations in the topological complexity across a pattern catalog, which still allows no information loss and high data compression. We show that different labeling strategies and pattern radius contribute to the trade-offs between prediction accuracy and model precision. Two imbalance-aware training strategies are also discussed with supporting experiments.