Geometric-Robust Contrastive Learning Framework With Layout Topology Representation and Self-Calibration Voting for Hotspot Detection

Recently, machine learning-based techniques have been increasingly employed for layout hotspot detection. However, existing supervised-based methods often struggle to learn a reliable decision boundary across the entire dataset and ignore the geometric properties and topology of the polygons. Moreover, robustness to geometric transformations of layout patterns has received limited attention in hotspot detection, potentially affecting model generalization. To address these challenges, this article introduces GRC-HD, a novel contrastive learning framework that aligns image features with layout topology representations while integrating geometric-robust feature learning. The framework improves the ability to distinguish between hotspots and non-hotspots by similarity computations instead of a single decision boundary. To effectively incorporate geometric information into the model training process, we propose Layout2Seq, which encodes polygon shapes as vector sequences and feeds them into GRC-HD. Furthermore, to better represent topological information, we develop an absolute position embedding, replacing the standard position encoders used in Transformer architectures. To improve the robustness of hotspot detector, we design a configurable geometric transformation library that generates multiple manufacturing-relevant views. The library is leveraged through an auxiliary Siamese network during training for geometric-robust representation learning, and through a self-calibration voting strategy at inference, where transformed instances act as independent voters to enhance decision reliability. Extensive experiments on various benchmarks demonstrate that GRC-HD outperforms current state-of-the-art methods.