Data-driven Wafer Yield Prediction Method

As the core and bottleneck of high-end manufacturing,semiconductor chips are increasingly important in a series of major strategic issues such as national development and sustainable economic development.In the semiconductor wafer manufacturing process,as the size of integrated circuits continues to shrink and the processing technology becomes more complex,the number of parameters that need to be tested for wafer quality inspection is gradually increasing,and the corresponding time and cost are also increasing.In view of the large amount of data between the Wafer Acceptance Test(WAT)parameters in the semiconductor wafer manufacturing process,the data redundancy,the relationship and the complex mapping,how to propose an effective method for predicting wafer yield,so as to reduce the cumbersome test process and diagnostic range while ensuring the prediction accuracy of wafer yield,predicting the trend of yield change,reducing the subsequent test time and testing cost,It is also important to quickly discover the cause of low yield and improve the level of wafer fabrication during the wafer manufacturing process.In view of the above requirements,this paper conducts a systematic study on wafer critical WAT parameter selection and wafer yield prediction methods.The main research work is as follows:1)WAT parameters feature selection: For the characteristics of high acceptance of wafer acceptance test parameters,strong redundancy between data,and insignificant key parameters,aiming at minimizing the prediction error value and minimum feature parameters of wafer yield,A hybrid feature selection method that combines filtering and wrappered feature selection method to identify critical wafer acceptance test parameters during wafer fabrication.Firstly,the maximum correlation minimum redundant filtering parameter pre-screening method based on mutual information is proposed.The correlation between each parameter and the wafer yield value is calculated by mutual information.At the same time,mutual information is used to measure the redundancy between parameters,and the maximum correlation minimum redundancy pre-screening of single parameters is realized,and the feature size of further search is reduced.Secondly,a packaged key parameter identification model based on genetic algorithm and neural network is designed.The coding and optimization of candidate input parameters are realized by genetic algorithm.Further more the wafer yield prediction error value of the neural network and the weight information of the selected features are solve d as the fitness function to realize the selection process of the combined parameters.Finally,the validity of the proposed method is verified by the standard data set and the example data.2)Wafer yield prediction: There are many factors affecting wafer yield,such as large data volume and complex data relati onships.Based on the key WAT parameters,a wafer yield prediction method based on improved continuous deep belief network is proposed.The designed wafer yield prediction model based on deep belief network mainly includes two parts,namely,automatic extrac tion of key features by improving the hidden layer of continuous restricted Boltzmann machine.Using the error back propagation mechanism of the output layer,the prediction error of the wafer yield is fine-tuned.Using the example data,the prediction accuracy of the proposed method and the existing literature method is compared,and the effectiveness of the method is verified.Based on the above research,based on the wafer quality control scene of a 300 mm wafer production line in Shanghai,the requirements of wafer quality control system were analyzed,and the wafer yield predection system was designed and developed.Verification of the main research content provides an effective tool and platform for the change of key WAT parameters and the acquisition trend of wafer yield during the wafer manufacturing process.