From Contrastive to Generative Alignment: Large-Scale Hierarchical Multimodal Pretraining for Hotspot Detection

The continuous reduction in semiconductor feature sizes has made hotspot detection (HSD) a critical yet challenging task in optimizing mask designs for manufacturability. While deep learning-based methods show promise, their reliance on large labeled datasets and training from scratch for each design makes them impractical for industrial use due to costly and time-intensive HSD data labeling. To address these challenges, we are the first to investigate self-supervised large-scale multi-modal pre-training for HSD, leveraging both layout images and GDSII text data to learn robust and transferable representations. To enable large-scale pre-training, we construct AugLayout-500K, a dataset of 500K paired layout images and GDSII files generated through an automatic augmentation pipeline. Building on this, we propose a novel self-supervised multi-modal pre-training framework that aligns paired layout images and GDSII text data through hierarchical alignment. Our framework introduces two key components: a global alignment module that aligns holistic image and text features using a sigmoid-based contrastive loss, and a fine-grained alignment module that decodes polygon coordinates via image-to-text generation. This approach captures both high-level correspondences and fine-grained spatial relationships, enabling robust and efficient downstream adaptation. To ensure fair comparisons and advance future research in HSD, we curate consistent and reliable evaluation benchmarks, addressing redundancies and inconsistencies in existing datasets. Extensive experiments on multiple datasets across low-resource fine-tuning, cross-design generalization, and full fine-tuning settings show that our method significantly outperforms state-of-the-art approaches.