Research On Wafer Map Detection Based On Deep Learning

Wafer map detection is an important application of object detection and has received extensive attention.For semiconductor industry,wafer map detection has become a major defect detection problem.Semiconductor manufacturing process involves dozens of complex steps that can lead to defects on the surface for numerous reasons.Visualization and defect pattern recognition are crucial to prevent defects.Defect pattern recognition provides a reference for engineers to deal with manufacturing-related problems by identifying wafer surface defects.At present,with the decrease of wafer size and the increase of production process complexity,the number of mixed complex defects combined with multiple basic defects is increasing.When mixed defects are generated,defect detection becomes more complex,especially when tens of millions of wafer maps are detected in industrial production,the detection precision and speed are highly required.Therefore,a detection method with high precision and fast efficiency is undoubtedly required at present.The main research contents are as follows:1)In view of the problem that there is no open,complete and practical datasets,establishment of wafer map datasets is crucial for mixed defects.The mixed defects dataset is the basis of related experiments.Based on the existing sample library,the wafer map datasets were screened and made.The methods of data enhancement and normalization are also introduced.This method performs online data enhancement before training,and the number of samples increases with the increase of training times,which effectively solves the over-fitting problem caused by too little data in the network.2)Research on feature extraction algorithm based on wafer map detection.Feature extraction is the key to wafer map recognition.The wafer maps are feature-rich and various,and the traditional algorithm is difficult to cope with these features.Therefore,it is necessary to optimize the traditional algorithm to make it suitable for the feature extraction of wafer maps.In the basis of Res Net,a structure that integrates multi-scale feature pyramid network,and adds a cardinality on each residual block to improve the accuracy.The algorithm can reduce the number of hyperparameters by stacking the feature extraction network of the same module.Comparative experiments are conducted on the validation set and test set,and the quantitative analysis of the results shows that compared with the traditional algorithm,the algorithm can fully extract features at a shallower level,which improves the accuracy to reach 98.5%.3)Research on region extraction algorithm based on wafer map detection.In order to effectively eliminate redundant anchors and solve the problem of object overlap,a post-processing algorithm is designed to update the scores of anchors by optimizing the penalty term distribution model of traditional post-processing methods,and then improve the detection efficiency in the inference stage.The algorithm is able to cope with the localization problem of mixed wafer maps and solves the problem that the occluded object is difficult to detect because of the overlap of the extracted object.On different single-stage and two-stage detectors,the post-processing algorithm is replaced by the improved algorithm to verify the effectiveness of the algorithm from the detection efficiency and detection precision.In inference,the efficiency of single-stage and two-stage detectors is increased on average by 9.63% and 21.72%,respectively.4)Design and implement wafer maps detection algorithm.Systematic experiments were carried out.The constructed detection network was used to train and test the wafer maps.The structure of each part was adjusted according to the training and test results,and the network performance was evaluated.The test set is compared with the traditional algorithm to determine the performance of different methods.The algorithm successfully detects six types of wafer maps,and the object precision can reach 0.604,which is 1.9% better than before the improvement,and the detection efficiency in the inference stage is 25.8% better.