Generic lithography modeling with dual-band optics-inspired neural networks

Haoyu Yang, Zongyi Li, Kumara Sastry, Saumyadip Mukhopadhyay, Mark Kilgard, Anima Anandkumar, Brucek Khailany, Vivek Singh, Haoxing Ren

Abstract

Lithography simulation is a critical bottleneck in semiconductor manufacturing, particularly as feature sizes shrink to the 7nm node and beyond. While traditional physics-based simulators are computationally expensive, existing machine learning alternatives often struggle to balance model size with the long-range optical interactions required for high-fidelity results. This paper introduces DOINN (Dual-band Optics-Inspired Neural Network), a novel architecture that integrates a Fourier-based global perception path with a convolutional local perception path. DOINN achieves an 85x speedup over traditional simulators and is 20x smaller than previous state-of-the-art machine learning models, while maintaining high accuracy at a 1nm²/pixel resolution.

Introduction

In modern semiconductor fabrication, Design for Manufacturability (DFM) is essential for ensuring that mask patterns translate accurately onto silicon wafers. This process relies heavily on lithography simulation to predict and correct errors before manufacturing. Historically, these simulations have relied on the Hopkins diffraction model, which is physically rigorous but demands immense computational resources.

Recent attempts to accelerate this process using Convolutional Neural Networks (CNNs) have faced a significant "receptive field" challenge. High-fidelity lithography requires capturing long-range optical interactions—the "optical diameter"—where patterns far from a specific point still influence its final shape. To capture these global interactions at a 1nm²/pixel resolution, standard CNNs often become prohibitively large and slow, limiting their practical application in full-chip layouts.

Methodology: The DOINN Architecture

DOINN addresses the limitations of standard CNNs by mimicking the underlying physics of light through a "dual-band" architecture. The model splits the simulation task into two parallel paths that handle different frequency components of the optical signal.

1. Fourier Global Perception (GP) Path

Inspired by Fourier Neural Operators (FNOs), the GP path handles low-frequency, global information. Since lithography is fundamentally a low-pass filtering process (diffraction), DOINN uses an optimized Fourier Unit to perform operations in the frequency domain. This allows the model to efficiently capture the "big picture" of light intensity across the layout without the need for excessively deep convolutional layers.

2. Convolutional Local Perception (LP) Path

While the Fourier path captures the global intensity, it can lose high-frequency details due to truncation. The LP path uses VGG-style convolutional blocks to capture fine-grained mask details and sharp edges, ensuring the final contour map remains precise.

3. Image Reconstruction (IR) Path

The features from the global and local paths are concatenated and processed through an IR path. Using transposed convolutions, the model upscales the data into a high-resolution (1nm²/pixel) resist contour map.

4. Large Tile Simulation Scheme

To process real-world full-chip layouts, the researchers implemented a "sliding window" approach using half-overlapping tiles. This scheme ensures that the optical diameter is fully accounted for at every point, preventing boundary errors that typically plague tile-based simulation.

Experimental Results

The DOINN framework was evaluated against three major benchmarks: ISPD-2019 (via layers), ICCAD-2013 (metal layers), and N14 (a proprietary 14nm node with high-density vias). Simulations were conducted at both 1000x1000 and 2000x2000 resolutions to test scalability.

Performance and Efficiency

  • Computational Speed: DOINN demonstrated an 85x speedup compared to traditional physical simulators like Lithosim and Calibre, with only a marginal (~1%) loss in accuracy.
  • Model Compactness: By leveraging the efficiency of the Fourier domain, DOINN requires only 1.3 million parameters. This is a 20-fold reduction compared to previous state-of-the-art models like DAMO-DLS, which utilize approximately 18 million parameters.
  • Accuracy: On the advanced N14 node, DOINN outperformed standard UNet architectures by 4% in Mean Pixel Accuracy (mPA) and 5% in Mean Intersection Over Union (mIOU).

Discussion and Key Insights

The success of DOINN highlights the importance of inductive bias in neural network design. Rather than treating lithography as a "black box" image-to-image task, the authors built a network that mirrors the physical equations of optics (Fourier transforms followed by non-linearities). This physics-inspired design allows the model to converge faster and operate more efficiently than generic architectures.

The synergy between the dual paths is vital: the Fourier path provides the "intensity" (where the light is distributed), while the CNN path provides the "sharpness" (where the edges are defined). This is the first work to demonstrate via and metal layer contour simulation at 1nm²/pixel resolution for arbitrary tile sizes using such a compact model.

Conclusion and Future Work

DOINN represents a significant advancement in high-speed, high-resolution lithography simulation. However, several avenues for future research remain. Current models utilize a "constant threshold" resist model; incorporating complex chemical and physical resist effects could further improve fidelity. Additionally, because DOINN is fast and differentiable, it holds great potential for Inverse Lithography Technology (ILT), where the model could be used to optimize mask designs directly. Finally, extending the architecture to account for 3D mask effects will be necessary for the most advanced manufacturing nodes at 3nm and below.

Summary updated

BibTeX
@inproceedings{yang2022generic,
  author = {Haoyu Yang and Zongyi Li and Kumara Sastry and Saumyadip Mukhopadhyay and Mark Kilgard and Anima Anandkumar and Brucek Khailany and Vivek Singh and Haoxing Ren},
  title = {{Generic lithography modeling with dual-band optics-inspired neural networks}},
  booktitle = {DAC},
  year = {2022},
  doi = {10.1145/3489517.3530580},
}
Figure