Scholarlink

The proximity effect in optical lithography causes substantial pattern distortions due to diffraction, spherical aberration, and light scattering, severely affecting resolution and pattern fidelity in advanced semiconductor fabrication. This study introduces a hybrid correction methodology that integrates machine learning, deconvolution, and diffraction-aware layout enhancement. A convolutional neural network is trained to classify and quantify the contributions of Gaussian, sinc, and Airy disk point spread functions. A secondary U-shaped convolutional neural network predicts optimal placement of serif and hammerhead features to improve line-end integrity. In parallel, spherical aberration is explicitly modeled and incorporated into the process. The process includes size reduction, followed by Richardson–Lucy deconvolution, and the strategic insertion of sub-resolution assist features (SRAFs) to enhance aerial image contrast. The final corrected pattern achieves significant improvements: peak signal-to-noise ratio (PSNR) of 18.18, structural similarity index (SSIM) of 0.9445, edge normalized cross-correlation (ENCC) of 0.1304, and a deconvolution quality factor (DQF) of 0.9840, demonstrating high-resolution pattern transfer.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Neural-Augmented Optical Correction: High-Fidelity Mask Optimization via Mixed Point Spread Function Learning and Sub-Resolution Assist Feature-Aware Richardson–Lucy Deconvolution

  • 章节

  • 发表日期：

    2026年

查看全文
  • Tanmay Wani,
  • Harshika Mhapsekar,
  • Bikash Dev Choudhury

摘要

The proximity effect in optical lithography causes substantial pattern distortions due to diffraction, spherical aberration, and light scattering, severely affecting resolution and pattern fidelity in advanced semiconductor fabrication. This study introduces a hybrid correction methodology that integrates machine learning, deconvolution, and diffraction-aware layout enhancement. A convolutional neural network is trained to classify and quantify the contributions of Gaussian, sinc, and Airy disk point spread functions. A secondary U-shaped convolutional neural network predicts optimal placement of serif and hammerhead features to improve line-end integrity. In parallel, spherical aberration is explicitly modeled and incorporated into the process. The process includes size reduction, followed by Richardson–Lucy deconvolution, and the strategic insertion of sub-resolution assist features (SRAFs) to enhance aerial image contrast. The final corrected pattern achieves significant improvements: peak signal-to-noise ratio (PSNR) of 18.18, structural similarity index (SSIM) of 0.9445, edge normalized cross-correlation (ENCC) of 0.1304, and a deconvolution quality factor (DQF) of 0.9840, demonstrating high-resolution pattern transfer.