Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
TokMan:Tokenize Manhattan Mask Optimization for Inverse Lithography
Authors: Yiwen Wu, Yuyang Chen, Ye Xia, Yao Zhao, Jingya Wang, Xuming He, Hao GENG, Jingyi Yu
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | This paper presents Tok Man... Experimental results demonstrate that Tok Man not only achieves high pattern fidelity and manufacturability but also offers improved runtime efficiency compared to traditional and learning-based ILT methods. 5 Experiments 5.1 Experimental Settings 5.2 Comparison with SOTA Methods 5.3 Ablation Study |
| Researcher Affiliation | Academia | Yiwen Wu 1 Yuyang Chen 1 Ye Xia1 Yao Zhao1 Jingya Wang1 Xuming He1 Hao Geng 1 Jingyi Yu 1 1Shanghai Tech University Equal contribution Corresponding author |
| Pseudocode | No | The paper describes the segmentation algorithm and the Diffusion Transformer process in detail within Section 4.1 and 4.2 using prose and diagrams (Figure 2 and Figure 3), but it does not present any formal pseudocode or algorithm blocks. |
| Open Source Code | No | Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [Yes] Justification: We used publicly available datasets, which is mentioned in experiments. For privacy, we will consider to release our code in future. |
| Open Datasets | Yes | To train our model effectively on Manhattan-aligned layout patterns, we construct a large-scale dataset derived from the publicly available benchmark [24]. Specifically, we extract polygonal shapes from the original layouts and systematically regenerate new samples by rearranging these polygons based on realistic placement characteristics observed in actual design environments. This data generation strategy ensures that the resulting samples remain consistent with typical layout design constraints while significantly enriching the diversity of polygon combinations. In total, we generate 89,697 layout samples as our training data. We evaluate our method on the ICCAD2013 benchmark [45], which includes simple and complex cases: ICCAD13-S and ICCAD13-L. |
| Dataset Splits | Yes | In total, we generate 89,697 layout samples as our training data. We evaluate our method on the ICCAD2013 benchmark [45], which includes simple and complex cases: ICCAD13-S and ICCAD13-L. |
| Hardware Specification | Yes | Training Settings. Tok Man model uses 6 Transformer blocks with embedding dimension 512 and 16 heads of self-attention, followed by a linear projection layer for opc correction output. We implement our work wth Py Torch-Geometric toolkit and train our model on the platform that possesses 32x NVIDIA H20 Graphics Cards, requiring approximately 80 GB of memory for batch size of 40 and test on 1x A100 Graphics Card. |
| Software Dependencies | No | We implement our work wth Py Torch-Geometric toolkit and train our model on the platform that possesses 32x NVIDIA H20 Graphics Cards, requiring approximately 80 GB of memory for batch size of 40 and test on 1x A100 Graphics Card. |
| Experiment Setup | Yes | Training Settings. Tok Man model uses 6 Transformer blocks with embedding dimension 512 and 16 heads of self-attention, followed by a linear projection layer for opc correction output. We implement our work wth Py Torch-Geometric toolkit and train our model on the platform that possesses 32x NVIDIA H20 Graphics Cards, requiring approximately 80 GB of memory for batch size of 40 and test on 1x A100 Graphics Card. For lithography simulator settings, the photoresist intensity threshold for lithography settings is set at 0.225, and sigmoid steepness is 50. The lithography wavelength is 193 nm with a defocus range of 30 nm and a dose range of 3%. The resolution of the mask and corresponding wafer image are all 1nm/pixel. |