Invited Paper: Unitho: A Unified Multi-Task Framework for Computational Lithography

Qian Jin, Yumeng Liu, Yuqi Jiang, Qi Sun, Cheng ZhuoICCAD 2025 · DOI

TL;DR

This paper introduces Unitho, a unified multi-task framework that significantly accelerates computational lithography by integrating mask generation, lithography simulation, and violation detection into a single, efficient model. By leveraging a Transformer-based architecture and novel training strategies, Unitho achieves an 85x speedup over traditional methods, enabling faster design iterations and improved manufacturability for advanced semiconductor nodes.

What Problem Does This Paper Address?

Modern semiconductor manufacturing relies on computational lithography to ensure that chip designs can be accurately printed onto silicon wafers. Key steps include mask generation, lithography simulation (predicting how light interacts with the mask and resist), and verification (checking for design rule violations like DRC, MRC, and LRC). Traditional methods for these tasks are often computationally intensive, time-consuming, and handled by separate, specialized tools. This fragmentation leads to long design cycles, especially when analyzing process window variations (how sensitive the print quality is to changes in focus and exposure dose). Existing machine learning approaches have focused on individual tasks, lacking a unified approach that can handle multiple aspects of lithography simultaneously and adapt to diverse process conditions. This limits their effectiveness for early-stage design-technology co-optimization (DTCO) and overall manufacturability.

Methods

Unitho is built upon a multi-task Transformer architecture, designed to perform mask/contour generation and violation detection in a unified framework.

  • Unified Lithography Simulator (Generator): This component takes a layout design and process conditions (dose, focus, etc.) as input and generates both the mask and the predicted resist contour. It uses a Swin Transformer for layout encoding and a process condition encoder to capture contextual information. A crucial element is the cross-attention fusion module, which links layout structures with process-aware features, allowing the model to understand how different process conditions affect specific layout regions. A contrastive learning strategy is employed to enhance the model's ability to distinguish between different process conditions and layouts, improving sensitivity and generalization.
  • Fast Hotspots Detector: This component is designed for efficient and accurate detection of DRC, MRC, and LRC violations. It utilizes a dual-branch backbone to process both layout and contour information for LRC detection, and a single-layer Transformer encoder for global context modeling. To address class imbalance and improve localization, it incorporates a Varifocal Loss (VFL) and a False Positive Penalty Loss (FPPL). For bounding box regression, it combines L1 loss with Generalized IoU (GIoU) loss to handle irregular shapes and sparse violations.
  • Joint Generation-Detection Framework: The key innovation is the integration of these two components. The generative model's output (predicted contour) is fed directly into the detection model. This allows for rapid process-aware LRC hotspot prediction without the need for time-consuming physical simulations. The framework is trained in three phases: pre-training the generator, pre-training the detector, and finally, fine-tuning the detector on the generated outputs.

Datasets

The authors created a large-scale lithography dataset at the 55 nm technology node using commercial tools. This dataset includes:

  • 624,129 layout-mask-contour triplets across 36 process combinations (varying source type, resist threshold, focus, and exposure dose).
  • Labeled hotspots for DRC (63,558 images), MRC (55,690 images), and LRC (116,881 images), along with 100,000 no-hotspot triplets.

Key Insights and Ideas

  • Unified Approach is Key: The fragmentation of lithography tasks into separate tools is a major bottleneck. Unifying mask generation, simulation, and detection into a single framework offers significant efficiency gains.
  • Process Conditions Matter: Explicitly encoding and fusing process conditions with layout information is critical for accurate and adaptable lithography modeling, especially for process window analysis.
  • Cross-Attention for Context: Guiding the model to focus on layout regions most sensitive to process variations via cross-attention is a powerful technique for improving robustness.
  • Contrastive Learning for Sensitivity: Differentiating between similar-but-distinct outputs (e.g., same layout under different conditions) is crucial for capturing subtle lithographic nuances.
  • End-to-End for Speed: Directly generating contours and then detecting violations within the same pipeline dramatically reduces inference time compared to sequential, physics-based methods.

Key Results

  • 85x Speedup: Unitho achieves an 85x speedup in runtime compared to a commercial tool for OPC, simulation, and verification on 6000x6000 nm layout clips.
  • High Accuracy in Generation: Unitho outperforms existing methods like CGAN and LithoNet in mask and contour generation, achieving high mPA, mIoU, and F1 scores.
  • Accurate LRC Violation Prediction: Unitho's generated contours result in LRC violation counts that closely match those of a standard simulation tool, with deviations consistently below 4%.
  • Superior Detection Performance: Unitho outperforms state-of-the-art methods in DRC, MRC, and LRC detection tasks, achieving higher F1 scores, significantly reducing false alarms (e.g., 45% reduction in DRC), and achieving much faster inference times (e.g., 1.47 ms for LRC detection vs. 28.55 ms for a baseline).
  • Qualitative Improvements: Visual comparisons show Unitho's detection results with fewer missed detections (blue boxes) and false positives (purple boxes) compared to baselines.

Limitations and Future Work

  • Dataset Size: While large, the dataset is specific to the 55 nm technology node. Generalizing to significantly different technology nodes might require further training or adaptation.
  • Future Work: The authors suggest that Unitho's capabilities can be extended to other EDA tasks, such as design rule checking (DRC) and mask optimization, further streamlining the design process. The framework's ability to provide rapid, process-aware feedback also opens avenues for more advanced Design-Technology Co-Optimization (DTCO) strategies.

Summary updated

BibTeX
@inproceedings{jin2025invited,
  author = {Qian Jin and Yumeng Liu and Yuqi Jiang and Qi Sun and Cheng Zhuo},
  title = {{Invited Paper: Unitho: A Unified Multi-Task Framework for Computational Lithography}},
  booktitle = {ICCAD},
  year = {2025},
  doi = {10.1109/iccad66269.2025.11240871},
}
Figure