Model Transferability from ImageNet to Lithography Hotspot Detection

Detecting lithographic hotspots is significant for VLSI fabrication before transferring the designed circuit layout pattern to silicon. However, the cost of detecting hotspot patterns by simulation is high; moreover, the cost of obtaining positive samples is also high. This paper explores the transferability of models pre-trained by the natural image dataset ImageNet for lithographic hotspot detection, especially to reduce the need for positive sample size. Various state-of-the-art (SOTA) CNN architectures, class weight ratios, and freezing layers are experimented with, all of which are fine-tuned for unbalanced sample datasets. The experimental results show that the migrated model can achieve high hotspot detection accuracy and low false alarm even with limited samples. Moreover, this paper verifies that there is inconsistency in the distribution of the original training set and the test set by merging and resampling the original training set and the test set into a new training set that will be used to train the model with better results, proving that the sample partitioning method is critical for hotspot detection.