Scholarlink

<p>Accurate classification of wafer defect maps is essential for improving semiconductor yield. Yet, conventional methods often underutilize domain-specific knowledge and consequently achieve limited performance. To address this gap, we propose a knowledge-data dual-driven framework that embeds wafer-specific expertise into deep learning. Our method comprises two main components: the Knowledge-Aware Feature Enhancement (KAFE) module and the Knowledge-Data Dual-Driven Network (KDDNet). KAFE transforms raw wafer maps into continuous defect-density heatmaps, selectively enhancing regions with concentrated defects while suppressing background noise. Then, KDDNet builds on these enhanced representations through a dual-branch design. A multi-scale convolutional subnetwork extracts explicit spatial defect patterns, while a variational feature-pyramid subnetwork learns latent statistical distributions. These complementary features are fused for final classification. On the MixedWM38 benchmark, our method achieves 98.64% accuracy, 98.65% precision, 98.64% recall, and 98.64% F1-score with only 100 training iterations-delivering nearly 40<InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="10845_2025_2713_Article_IEq1.gif" Format="GIF" Height="13" Rendition="HTML" Resolution="72" Type="Linedraw" Width="19" /> </InlineMediaObject> <EquationSource Format="TEX">\(\times \)</EquationSource> </InlineEquation> faster convergence than prior methods. By embedding process knowledge directly into the learning pipeline, this approach provides a fast, interpretable, and highly accurate solution for wafer defect analysis and process diagnosis.</p>

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

Density-aware multi-modal feature fusion for wafer map defect classification in semiconductor manufacturing

  • 期刊论文

  • 发表日期：

    2025年10月30日

查看全文
  • Wei Zhao,
  • Yongkang Liu,
  • Yuandong Gu,
  • Xiupeng Shi

摘要

Accurate classification of wafer defect maps is essential for improving semiconductor yield. Yet, conventional methods often underutilize domain-specific knowledge and consequently achieve limited performance. To address this gap, we propose a knowledge-data dual-driven framework that embeds wafer-specific expertise into deep learning. Our method comprises two main components: the Knowledge-Aware Feature Enhancement (KAFE) module and the Knowledge-Data Dual-Driven Network (KDDNet). KAFE transforms raw wafer maps into continuous defect-density heatmaps, selectively enhancing regions with concentrated defects while suppressing background noise. Then, KDDNet builds on these enhanced representations through a dual-branch design. A multi-scale convolutional subnetwork extracts explicit spatial defect patterns, while a variational feature-pyramid subnetwork learns latent statistical distributions. These complementary features are fused for final classification. On the MixedWM38 benchmark, our method achieves 98.64% accuracy, 98.65% precision, 98.64% recall, and 98.64% F1-score with only 100 training iterations-delivering nearly 40 \(\times \) faster convergence than prior methods. By embedding process knowledge directly into the learning pipeline, this approach provides a fast, interpretable, and highly accurate solution for wafer defect analysis and process diagnosis.