TL;DR
This paper introduces a novel, lightweight deep learning model for detecting lithography hotspots in semiconductor manufacturing. By cleverly integrating Squeeze-and-Excitation (SE) and Efficient Channel Attention (ECA) mechanisms into a simplified Convolutional Neural Network (CNN) architecture, the model achieves high accuracy and efficiency, outperforming established models like ResNet50 and Swin Transformer.
What Problem Does This Paper Address?
As semiconductor devices shrink, the precision required for lithography becomes increasingly critical. The "Optical Proximity Effect" (OPE) causes distortions between the mask pattern and the actual wafer pattern, leading to manufacturing defects known as lithography hotspots. Detecting these hotspots before manufacturing is crucial to reduce defects and improve yield. While methods like lithography simulation and pattern matching exist, they have limitations. Simulation is computationally expensive, and pattern matching struggles with novel or unknown hotspot patterns. Deep learning offers promise but often results in complex, resource-intensive models that are difficult to deploy in real-world manufacturing environments. This paper aims to bridge this gap by developing a deep learning model that is both highly accurate and computationally lightweight.
Methods
The core of this work is a custom-designed lightweight CNN architecture. The authors integrate two key attention mechanisms:
- Squeeze-and-Excitation (SE) Module: This mechanism adaptively recalibrates channel-wise feature responses by explicitly modeling inter-dependencies between channels. It learns to emphasize informative features and suppress less useful ones without significant computational overhead.
- Efficient Channel Attention (ECA) Module: Building on SE, ECA further enhances feature representation by capturing cross-channel interactions more effectively. Crucially, it avoids dimensionality reduction, which can lead to information loss, and uses a simple 1D convolution to achieve efficient channel attention.
The proposed model simplifies the CNN structure by using seven convolutional layers and four pooling layers, followed by three fully connected layers. This streamlined design reduces the number of parameters and computational complexity compared to standard architectures.
Datasets
The authors utilized two datasets for their experiments:
- Collected Layout Dataset: A custom dataset of 16,251 polysilicon gate-layer layouts from a wafer fabrication facility, specifically for 55 nm polysilicon gate-layer masks. This dataset was categorized into six classes: BRIDGE_NOM, BRIDGE_PW, PINCH_NOM, PINCH_PW, LINE_END, and non-hotspots.
- ICCAD 2012 Mask Layout Dataset: A publicly available dataset commonly used for evaluating lithography hotspot detection methods. This dataset includes five benchmarks with varying layout sizes and densities, representing 32 nm and 28 nm technologies. The authors applied data augmentation (flipping and mirroring) to address data imbalance issues within this dataset.
Key Insights and Ideas
- Attention for Efficiency: The SE and ECA mechanisms are not just about improving accuracy; they are key to achieving a lightweight model. By adaptively focusing on relevant features, they allow for a simpler overall network architecture without sacrificing performance.
- Cross-Channel Interaction is Key: The ECA module's ability to efficiently capture interactions between channels, without dimensionality reduction, is a significant advantage. This allows the model to learn more nuanced relationships within the layout data.
- Simplicity Can Be Powerful: The paper demonstrates that a carefully designed, simplified CNN structure can outperform more complex, established architectures when augmented with effective attention mechanisms. This challenges the notion that state-of-the-art performance always requires massive models.
- Generalization Across Datasets: The model's strong performance on both the custom dataset and the challenging ICCAD 2012 dataset highlights its robustness and ability to generalize to different types of lithography layouts and defect patterns.
Key Results
- Lightweight Design: The proposed LHD model has the fewest parameters among the compared models (22,132,548), significantly less than ConvNeXt (196,155,270), Swin Transformer (28,265,032), and ResNet50 (23,520,326).
- High Accuracy: On the collected dataset, the LHD model achieved an accuracy of 96.83%, outperforming ResNet50 (96.64%), Swin Transformer (96.49%), and ConvNeXt (94.39%).
- Superior Runtime: The LHD model achieved a runtime of 4.95 h/mm², which is notably faster than ResNet50 (6.10 h/mm²), Swin Transformer (7.75 h/mm²), and ConvNeXt (7.67 h/mm²).
- ICCAD 2012 Performance: The LHD model consistently outperformed other classic models across most benchmarks in the ICCAD 2012 dataset, particularly showing superior mean precision and recall while maintaining lower detection times. For instance, in Benchmark 3, it achieved 96.30% accuracy, 86.14% precision, and 95.33% recall with a runtime of 21.76 h/mm².
Limitations and Future Work
- Data Availability: The authors state that the data underlying the results are not publicly available at this time but can be obtained from the authors upon reasonable request. This limits immediate reproducibility for the broader community.
- Specific Hotspot Types: While the model performs well, the paper notes that challenges remain with high similarity in certain hotspot categories (e.g., PINCH_NOM), though the model still achieved a respectable accuracy of 79.9% in this category.
- Future Work: The paper implicitly suggests that further exploration of attention mechanisms and network architectures could lead to even greater improvements in accuracy and efficiency for lithography hotspot detection. The successful generalization across diverse datasets also opens avenues for applying this lightweight approach to other complex pattern recognition tasks in semiconductor manufacturing.
Summary updated