Abstract: Wafer yield is a key indicator for measuring the quality of semiconductor products.Stable and accurate predictions can help identify defects in wafer processing,improve chip quality and control chip production costs.Due to the influence factors of wafer yield,large data volume and complex data relationship,based on the electrical test parameters in wafer processing,a wafer yield prediction method based on improved continuous deep belief network was proposed.A two-stage data preprocessing method for wafer electrical test parameters was proposed,in which the first stage performed data cleaning on missing values and outliers in the wafer electrical test parameters,and the principal component analysis was performed on the multi-collinearity relationship between wafer electrical test parameters to obtain the input variables of predictive model in the second stage.A wafer yield prediction model was designed based on deep belief network,which realized the automatic extraction of key features by improving the continuous-type restricted Boltzmann machine of the hidden layer,and the error back propagation mechanism of the output layer was used to accurately predict the wafer yield.With the laboratory actual production data,the prediction accuracy of the proposed method and the existing literature method was compared,and the effectiveness of the method was verified.

Key words: wafer yield prediction, continuous deep belief network, wafer acceptance test parameters, principal component analysis

摘要： 晶圆良率是衡量半导体产品质量的关键指标,对其进行稳定、准确的预测能够帮助发现晶圆加工工艺缺陷、提高芯片质量、控制芯片生产成本。针对晶圆良率的影响因素多、数据体量大、数据间关系复杂等特点,以晶圆加工过程中的电性测试参数为依据,提出一种基于改进的连续型深度信念网络的晶圆良率预测方法。首先提出晶圆电性测试参数的两阶段数据预处理方法,第一阶段对晶圆电性测试参数中的缺失值、异常值进行数据清洗,第二阶段对晶圆电性测试测试参数间的多重共线性关系进行主成分分析,以获取预测模型的输入变量。然后设计了基于深度信念网络的晶圆良率预测模型,通过改进隐藏层的连续型受限制玻尔兹曼机,实现了关键特征的自动提取,利用输出层的误差反向传播机制,实现了晶圆良率的准确预测。采用实例数据,对比了所提方法与现有文献方法的预测准确率,从而验证了所提方法的有效性。

关键词: 晶圆良率预测, 连续型深度信念网络, 晶圆电性测试参数, 主成分分析法