Scholarlink

<p>Deposited film thickness is a critical quality metric in Chemical Vapor Deposition (CVD) processes, essential for ensuring the performance and yield of semiconductor devices. Conventional metrology relies on sparse measurements, limiting process visibility and increasing costs. This motivates the adoption of data-driven virtual metrology (VM) techniques. While recent studies have explored deep learning models tailored for tabular data, their application in VM scenarios remains limited due to three key challenges: (1) limited labeled data; (2) lack of dynamic feature selection; and (3) localized learning that fails to capture cross-batch variations. To address these challenges, this paper proposes S2GA-VM, a self-supervised and global-aware VM framework for accurate and scalable film thickness prediction using tabular sensor data. It integrates three key innovations: (1) a self-supervised pretraining strategy that constructs a pretext task from feature–target correlations, enabling robust parameter learning from unlabeled data; (2) a stochastic gating mechanism for dynamic feature selection, enhancing generalization and reducing redundancy; and (3) a global-aware representation learning approach that employs projection space embedding to capture cross-batch process variations. Extensive validation on industrial datasets from three CVD systems (over 17,000 wafers), demonstrates the effectiveness of S2GA-VM. It consistently outperforms 25 state-of-the-art models, achieving up to 0.68 R<sup>2</sup> and superior directional accuracy in online wafer batch prediction. These results validate S2GA-VM as a viable real-time solution to enhance process control.</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

S2GA-VM: self-supervised and global-aware virtual metrology for accurate film thickness prediction in semiconductor manufacturing

  • 期刊论文

  • 发表日期：

    2025年9月30日

查看全文
  • Weiping Xie,
  • Jiangchen Wu,
  • Yuping Wang,
  • Yining Chen

摘要

Deposited film thickness is a critical quality metric in Chemical Vapor Deposition (CVD) processes, essential for ensuring the performance and yield of semiconductor devices. Conventional metrology relies on sparse measurements, limiting process visibility and increasing costs. This motivates the adoption of data-driven virtual metrology (VM) techniques. While recent studies have explored deep learning models tailored for tabular data, their application in VM scenarios remains limited due to three key challenges: (1) limited labeled data; (2) lack of dynamic feature selection; and (3) localized learning that fails to capture cross-batch variations. To address these challenges, this paper proposes S2GA-VM, a self-supervised and global-aware VM framework for accurate and scalable film thickness prediction using tabular sensor data. It integrates three key innovations: (1) a self-supervised pretraining strategy that constructs a pretext task from feature–target correlations, enabling robust parameter learning from unlabeled data; (2) a stochastic gating mechanism for dynamic feature selection, enhancing generalization and reducing redundancy; and (3) a global-aware representation learning approach that employs projection space embedding to capture cross-batch process variations. Extensive validation on industrial datasets from three CVD systems (over 17,000 wafers), demonstrates the effectiveness of S2GA-VM. It consistently outperforms 25 state-of-the-art models, achieving up to 0.68 R2 and superior directional accuracy in online wafer batch prediction. These results validate S2GA-VM as a viable real-time solution to enhance process control.