Research On Yield Diagnosis Method For Addressable WAT Based On Machine Learning

In IC industry,yield enhancement means lower production costs and higher profit rate,and therefore is the key factor for IC companies to maintain core competitiveness in the market.With the continuous scaling down of semiconductor feature size,the manufacture process of integrated circuit has become more complicated,which makes the yield loss problem more serious.How to carry out yield enhancement with the mass of production data collected in integrated circuit manufacture process becomes a key problem for IC companies.Wafer Acceptance Test(WAT),as an important test in the manufacture process of integrated circuit,is often used to extract device parameter model and monitor process fluctuation,and plays a key role in yield enhancement.This thesis focuses on the Addressable WAT scenario and makes a research about the solution of using machine learning method for yield diagnosis.Aiming at the shortcomings of the existing methods in analyzing efficiency and accuracy,a yield diagnosis method based on random forest algorithm is proposed on the basis of experiments.The main content of this thesis include:(1)Performing analysis of two main addressable WAT yield diagnosis methods:multidimensional analysis and decision tree-based yield diagnosis,and explaining their limitations.(2)Proposing a yield diagnosis method based on random forest algorithm.In this method,random forest algorithm is used to establish the classification model between test structure parameters and yield,and the class imbalance in WAT test data set is dealt with by combining resampling and cost sensitivity.(3)Proposing a random forest rule extraction method to solve the shortness of comprehensibility.In this method,sequential forward search(SFS)is used to reduce the size of the random forest while the classification performance is not influenced.Greedy iteration is used to extract key rules from the original rule set.The evaluation criteria of the rule are also be modified to improve the ability of discovering unknown samples.The proposed method has been experimented on dataset from a real addressable WAT,and the obtained rule set has better classification performance than the rule set obtained by method based on decision tree.