Wafer Shape Diagnosis And Prediction Using Big Data Strategy

Silicon wafer is crucial substrate material for integrated circuits.Data analysis was seriously focused.With the high-speed development of IC line width,the quality request for silicon wafer became more and more rigorous,thus the traditional experiential model and engineering understandings went incompetent as the practical production environment was multi-factor and complicated.In the quality analysis and improvement area,statistical data analysis technology earned more respect and had been widely applied in production.The application of computer science and calculation technology interact with traditionary subject create some new research branches.In this research,the inspection raw-data of semiconductor silicon wafer was studied,there were two parts:the first part was to achieve an automated diagnosis machine to judge the root cause of wire-saw,lapping and acid etching processes;the second part was to forecast the etched wafer shape before the actual process based on the modeling of acid etching,an experiential model was created and solved in order to estimate the removal distribution under different factor settings,especially,when the factor settings were fixed,the spatial removal distribution model was created in order to forecast the post-etching wafer shape,as to predict the successful yield on the post-lapping rejected wafers.In the diagnosis section,32 statistics were calculated from the inspection raw-data,then 9 features were extracted,the features were verified to achieve the discriminate between various failure root causes and also showed sufficient similarity within group,which approved their effectiveness for the training and verifying.The minimum error output was from the random forest classifier(0.3%);In the prediction section,the experiential model was created by engineering understandings and solved,the mean sum error was 0.061um for average,0.139 for maximum which proved the accurate when forecasting the removal distribution with various factor settings;Then the spatial distribution was estimated,the mea sum error was 0.028952um which can be used in precisely shape prediction for the remedy decision of upper-stream rejected production.