From IC Layout to Die Photo: A CNN-Based Data-Driven Approach

Hao-Chiang Shao, Chao-Yi Peng, Jun-Rei Wu, Chia-Wen Lin, Shao-Yun Fang, Pin-Yen Tsai, Yan-Hsiu Liu

Abstract

Semiconductor manufacturing relies on Optical Proximity Correction (OPC) to mitigate nonlinear distortions during photolithography. Traditional physics-based simulations are computationally intensive and often fail to generalize to novel patterns. This paper introduces a deep learning framework comprising LithoNet, a fast lithography simulator, and OPCNet, a self-supervised mask corrector. By treating these tasks as image-to-image translations and utilizing spatial transformer networks for deformation mapping, the authors achieve simulation speeds over 2,000 times faster than industry standards while maintaining high accuracy across various process windows.

Introduction

In Integrated Circuit (IC) fabrication, the physical realization of a layout often deviates from its ideal design due to distortions like corner rounding and wire necking. To ensure fidelity, engineers use OPC to pre-distort masks so that the final fabricated result matches the intended design. However, the computational cost of traditional simulations is a major bottleneck; a single full-chip simulation can take days. This paper proposes replacing slow, rigid physics-based models with a data-driven neural network approach that is both fast and flexible.

Methods

The proposed framework bridges the gap between layout design and fabrication through two interconnected networks.

LithoNet: The Simulator

LithoNet predicts the fabricated appearance of a layout. To handle the inherent noise in raw Scanning Electron Microscope (SEM) images, the authors employ a CycleGAN for domain transfer, translating SEM data into a clean, "layout-styled" format. The core architecture is a U-Net-like generator integrated with a Spatial Transformer Network (STN). Instead of generating pixels from scratch, the STN learns a "deformation map" to warp the original layout, ensuring the output remains physically grounded.

Detailed architecture of LithoNet, highlighting the domain transfer pipeline and the deformation-based generator.

OPCNet: The Corrector

OPCNet generates mask modifications to compensate for expected distortions. Its training utilizes a novel self-supervised approach: rather than relying on a database of "perfect" masks, OPCNet is trained using a pre-trained LithoNet as its supervisor. By minimizing the Input-Output Consistency Loss, the framework ensures that the simulated result of the corrected mask matches the original intended layout, effectively treating the two networks as inverse functions ($f = h \circ g = 1$).

Conceptual framework of LithoNet and the joint training of OPCNet using Input-Output Consistency Loss.

Datasets

The study utilized two datasets provided by United Microelectronics Corporation (UMC):

  • Dataset #1: 928 training and 114 testing pairs of layouts and SEM images under fixed fabrication conditions.
  • Dataset #2: 7,399 image pairs across seven normalized fabrication parameter settings to test the model's ability to simulate different "process windows."

Key Results

  • Speedup: The framework achieved a speedup of over 2,000x, processing layout patches in approximately 0.015 seconds compared to 34 seconds for traditional industry software.
  • Fidelity: The model accurately captured complex phenomena such as "necking" (where wires nearly break) and "tip-to-line" rounding, outperforming standard image-to-image models like Pix2Pix in maintaining shape integrity.
  • Process Control: By adjusting latent vectors representing fabrication parameters (like etching intensity), the model successfully reflected changes in metal line widths, demonstrating its utility for process window analysis.

Conclusion

This CNN-based approach significantly reduces the computational overhead of lithography simulation and mask correction. While challenges remain in developing objective metrics for SEM contour quality and handling extreme scale variations, this work paves the way for real-time virtual metrology and automated layout scoring in semiconductor manufacturing.

Summary updated

BibTeX
@article{shao2021from,
  author = {Hao-Chiang Shao and Chao-Yi Peng and Jun-Rei Wu and Chia-Wen Lin and Shao-Yun Fang and Pin-Yen Tsai and Yan-Hsiu Liu},
  title = {{From IC Layout to Die Photo: A CNN-Based Data-Driven Approach}},
  journal = {IEEE Trans. Comput. Aided Des. Integr. Circuits Syst.},
  year = {2021},
  doi = {10.1109/tcad.2020.3015469},
}
Figure