Scholarlink

<p>The research of Virtual Metrology (VM) for Chemical Mechanical Planarization (CMP) process occupies an essential role in semiconductor manufacturing. This paper proposes a novel VM model to predict average Material Removal Rate (MRR) which is the crucial characteristic to indicate the productivity of CMP process. The proposed model is developed based on the integration of Convolutional Neural Network (CNN) and a reference-based feature extraction strategy, which is capable to combine the temporal knowledge, and the high complexity involved during CMP process. The proposed Reference-CNN model extracts temporal features by examining the sensor recordings to detect the changing trends, from which the references are created for further prediction. Afterwards, CNN models are constructed to interpret the comprehensive relations between data. The public datasets of 2016 Prognostics and Health Management (PHM) Data Challenge are used to evaluate model performance compared with the benchmarks selected from literature. Experimental results show that the proposed model is capable to output superior prediction accuracy.</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

A reference-based virtual metrology method for material removal rate prediction using convolutional neural network

  • 期刊论文

  • 发表日期：

    2026年2月5日

查看全文
  • Haoshu Cai

摘要

The research of Virtual Metrology (VM) for Chemical Mechanical Planarization (CMP) process occupies an essential role in semiconductor manufacturing. This paper proposes a novel VM model to predict average Material Removal Rate (MRR) which is the crucial characteristic to indicate the productivity of CMP process. The proposed model is developed based on the integration of Convolutional Neural Network (CNN) and a reference-based feature extraction strategy, which is capable to combine the temporal knowledge, and the high complexity involved during CMP process. The proposed Reference-CNN model extracts temporal features by examining the sensor recordings to detect the changing trends, from which the references are created for further prediction. Afterwards, CNN models are constructed to interpret the comprehensive relations between data. The public datasets of 2016 Prognostics and Health Management (PHM) Data Challenge are used to evaluate model performance compared with the benchmarks selected from literature. Experimental results show that the proposed model is capable to output superior prediction accuracy.