China Mechanical Engineering

Previous Articles     Next Articles

#br# Wafer Acceptance Test Key Parameter Identification Based on Hybrid Feature Selection Method

LYU Youlong;XU Hongwei;ZHENG Cheng;ZHANG Jie;ZHENG Peng   

  1. College of Mechanical Engineering,Donghua University,Shanghai,201620
  • Online:2020-08-25 Published:2020-09-17

基于混合式特征选择模型的晶圆允收测试关键参数识别方法

吕佑龙;许鸿伟;郑城;张洁;郑鹏   

  1. 东华大学机械工程学院,上海,201620
  • 基金资助:
    国家自然科学基金资助项目(51905092,51435009);
    上海市青年科技英才扬帆计划资助项目(18YF1400800);
    中央高校基本科研业务费专项资金资助项目(2232018D3-28);
    国家商用飞机制造工程技术研究中心创新基金资助项目(COMAC-SFGS-2018-36175)

Abstract: Wafer acceptance test was a key process of the semiconductor manufacturing systems. The identification of key testing parameters enabled accurate prediction of wafer yield and in-time detection of product defects. With the consideration of high dimensions, strong redundancy and insignificant key parameters of testing data, a hybrid feature selection method composed of filter and wrapper was proposed to minimize the number of key parameters as well as the prediction errors of wafer yield. The relevance between each testing parameter and wafer yield and the redundancy of testing parameters were measured at the filter stage based on mutual informations. Additionally, the pre-screening of testing parameters with the goal of maximum relevance and minimum redundancy was realized.The encoding and combination optimization of testing parameters were realized by using a genetic algorithm to output key testing parameters at the wrapper stage, of which the fitness of each combination case was evaluated based on prediction errors of wafer yield provided by a backpropagation neural network. Finally, the effectiveness of this method was validated through a set of reference data as well as practical data collected from a wafer fabrication shops.

Key words: wafer acceptance test, wafer yield prediction, hybrid feature selection, minimum redundancy &, maximum relevance, genetic algorithm, neural network

摘要: 晶圆允收测试是晶圆加工过程的关键环节,对其中的关键测试参数进行准确分析识别有助于准确预测晶圆良率、及时发现工艺缺陷。针对测试参数维度高、数据冗余性强、关键特征不显著的特点,以最小化晶圆允收测试参数量和晶圆良率预测误差为目标,提出了过滤式与封装式相结合的混合式特征选择方法。在过滤式预筛选中,通过互信息度量各参数与晶圆良率的相关性,以及各参数之间冗余性,并根据最大相关、最小冗余准则,缩小候选参数规模;在封装式精选中,以遗传算法实现候选参数的编码、寻优,根据神经网络的晶圆良率预测误差进行适应度函数评价,进一步精选关键特征。最后,采用标准数据集和实例数据对所提方法进行了有效性验证。

关键词: 晶圆允收测试, 良率预测, 混合式特征选择, 最小冗余最大相关, 遗传算法, 神经网络

CLC Number: