TL;DR
The paper introduces a deep learning framework consisting of two complementary networks, LithoNet and OPCNet, to drastically accelerate semiconductor manufacturing simulation and mask optimization. By predicting non-linear circuit deformations and generating self-supervised layout corrections, the approach cuts down processing times from days to milliseconds while eliminating the need for expensive, proprietary ground-truth mask data.
What Problem Does This Paper Address?
When manufacturing integrated circuits (ICs), nanometer-scale photolithography and etching inevitably introduce non-linear shape distortions, such as line rounding or thinning. To counteract this, electronic design automation (EDA) engineers use Optical Proximity Correction (OPC) to modify the photomask layout so that the final printed circuit on the silicon wafer matches the intended design.
However, traditional physics-based or rule-based simulation tools are painfully slow—often taking days to run a full chip layout simulation. Furthermore, training machine learning models to automate and speed up this process typically requires a massive dataset of ground-truth corrected masks, which are highly proprietary, expensive, and difficult to collect. This paper tackles both challenges: providing an ultra-fast fabrication simulator and a "ground-truth-free" mask optimizer.
Methods
The authors propose a data-driven framework featuring two specialized convolutional neural networks (CNNs):
- LithoNet (Fabrication Simulator): Instead of simulating complex optical physics, LithoNet learns directly from data using a two-step approach. First, a CycleGAN translates noisy scanning electron microscope (SEM) die photos into clean, binarized "layout-styled" ground truths. Second, a U-Net-based generator predicts a dense 2D deformation map, which is used by a Spatial Transformer Network (STN) warping module to deform the input layout. Crucially, it accepts physical fabrication parameters (like scanner energy) as latent vectors to predict variations under different manufacturing conditions.
- OPCNet (Mask Corrector): To solve the lack of ground-truth mask data, OPCNet uses a clever self-supervised strategy. It is cascaded with a pre-trained, frozen LithoNet. OPCNet takes a target layout and generates a corrected mask. This mask is immediately fed into LithoNet, and the entire pipeline minimizes an Input-Output Consistency Loss. This forces the final simulated fabrication to match the original input layout, meaning the two networks effectively learn to act as mathematical inverse functions ($h \circ g = 1$).
Datasets
The models were trained and evaluated on two real-world industrial datasets provided by United Microelectronics Corporation (UMC), containing pairs of IC layout patches and corresponding wafer SEM images:
- UMC Dataset #1: Wafers fabricated under identical process conditions. It includes a 928-pair training set and a 114-pair blind testing set.
- UMC Dataset #2: Wafers fabricated across 7 different normalized parameter settings (ranging from $-0.9$ to $+0.9$) to test parameter-dependent modeling. It includes 7,399 image pairs for training and 84 pairs for blind testing.
Key Insights and Ideas
- The "Ground-Truth Free" Loop: The most elegant insight is using a forward simulator (LithoNet) to train an inverse model (OPCNet). By optimizing the loop to behave as an identity transform, the network learns to correct masks without ever seeing a human-designed or rule-based OPC mask.
- Deformation Maps Over Direct Generation: Instead of directly generating a synthetic SEM image (which often introduces blurry artifacts), LithoNet predicts a pixel-to-pixel displacement map and warps the original crisp layout. This preserves sharp boundaries and provides interpretable engineering insights into where distortions occur.
- CycleGAN as a Preprocessor: SEM images suffer from scan-line noise and shading. Using CycleGAN to map them to a clean, binary intermediate domain bridges the gap between heterogeneous data types (perfect vector layouts vs. noisy physical photos) without losing contour shapes.
Key Results
- Massive Speedups: Traditional contour simulation software takes around 34 seconds for a tiny $4 \times 1.7 \mu m^2$ patch and days for a full layout design. LithoNet and OPCNet perform their respective tasks in 0.0156 seconds and 0.0150 seconds per patch on an NVIDIA 2080Ti GPU—rendering full-scale layout screening practically instantaneous.
- High Fidelity: LithoNet successfully captures complex physical phenomena like "necking" (dangerous wire thinning caused by adjacent lines) and corner rounding.
- Parameter Sensitivity: Thanks to a dedicated parameter regression loss, LithoNet accurately predicts how changing scanner energy affects wire width, allowing fabrication plants to find optimal process windows virtually.
Limitations and Future Work
- Global vs. Local Metrics: Standard computer vision metrics like Intersection over Union (IoU) and SSIM measure global similarity well but can miss tiny, localized defects. In semiconductor manufacturing, a single sub-nanometer broken wire can completely ruin a chip's functionality even if the global IoU score is 99%.
- Future Work: The authors note the open challenge of developing better metrics that simultaneously characterize local and global manufacturability defects. They plan to extend this work into a comprehensive virtual metrology scoring system for automated layout quality assessment.
Summary updated