Literature review

In the past few years, methods of automated spatial pattern analysis in the semiconductor manufacturing process have been widely investigated and discussed.

They proposed many approaches to classify and recognize these kinds of spatial patterns. At least four different approaches are possible, the simplest and least efficient is visual classification performed by a human operator. In view of the time needed to analyze thousands of wafers, it is not possible to have a fast feedback to correct problems. The second approach is statistical distribution analysis such Poisson or negative-binomial distributions are normally on lot-level basis for account for cluster phenomena. For example, Friedman et al., developed statistics measuring spatial dependency of defects to detect systematic clustering [11]. Although most statistical approaches are able to detect anomalies on wafer, they are generally unable to extract meaningful data from the spatial pattern since tend to incorrectly assume a stationary probability distribution. Systematic spatial patterns caused by the semiconductor fabrication process involve a complex variation of statistical parameters which are highly dependent on the process, machine, suppliers, materials etc. Thus, traditional statistical approaches are not recommended for practical application. Moreover, the monitoring statistics often bear relatively complex

11

statistical properties, adding difficulty in implementation of these methods. The third approach is neural network algorithms that can recognize spatial patterns. However, the major limitation of neural-network approaches lies in its inability to classify two or more shift variance in the spatial patterns. It is also incapable of detecting the presence and location of a cluster. Many of these methods need a large training dataset, and due to the complexity of the algorithms, most of these methods do not provide a statistically rigorous evaluation of their performances in pattern detection. The fourth approach is image processing, this method attempts to detect cluster outliers, but fall sort at detecting real manufacturing spatial pattern, which are almost always of different and imperfect geometrical shapes. However, this approach lacks flexibility because it requires the prior knowledge of all the possible shapes in different products and difficult to predict in advance [23, 24, 25]. Image processing and artificial neural network (ANN) classification approaches are notorious to fail for noisy datasets.

Generally, neural network approaches cannot identify two or more shift variant or rotational variant spatial patterns that belong to the same defect pattern type [8].

A more robust classification scheme can be achieved using data mining algorithms. However, the accuracy and reliability of the data mining approach depends on features selection and the selected features must have some unique attribute that can be used to discriminate each characteristic pattern. As exhaustive search technique, the Hough transform is a robust method which is relatively unaffected by noise for feature detection and generally been recognized as a reliable for linear and circular object detection [14]. The Hough transform employing a normal line-to-point parameterization is widely applied in digital image processing for feature detection. It is a very powerful tool for the detection of parametric curves and generally used to detect lines and rounds, which is originally proposed by Hough in 1959. The advantage of Hough transform is as follows: it isn’t sensitive to the noise of

12

image, effectively eliminate the influence of noise. The transform is convenient to parallel computing. In the field of computer vision, some issues are complex and require a great of computation, and parallel computing is used by modified Hough transform, and then calculating the start and end of the line. In this paper, we applies the Hough transform for spatial pattern recognition and a frequency count to indentify lines or sets of lines which represent line signatures. K. P. White et al. have been demonstrated how this transform can be adapted to classify signatures on semiconductor wafers in 2008 [5], but the procedure does not appear to be as useful for detecting spatial patterns in Bull eye and Blob. Thus, we future apply circular Hough transform to detect spatial pattern representing in Bull eye and Blob which is proposed to detect rounds for spatial pattern [5, 6].

However, we found that some of the previous methods analyzed in rarely dataset.

For instance, on the basis of the analysis of simulated and real wafers, Chen and Liu concluded that the ART1 network classifies can recognize the similar defect spatial patterns more easily and correctly. However, the simulated data set was made of only 35 simulated wafers, each of which containing 294 dies. As for the patterns, there are three types of rings and four types of scratches. The real wafers analyzed in the paper were only 14 wafers, still with 294 dies per wafer [2]. L.J. Wei using neural network approaches for recognizing the bin-map patterns on the wafer. In their work, the 57 actual wafers had been analyzed which included 9 Bull eye, 14 Edge, 9 Blob, 9 Ring and 16 Line spatial patterns [12]. S.F. Liu et al. proposed a feature extraction procedure based on wavelet transform to extract features that represent different defect patterns. The presented methodology is verified with real industrial data from a semiconductor company and the experimental results show the presented methodology is able to recognize defect patterns with recognition accuracy of 95%, however, the real industrial data also made of only 65 wafers and the system failed to

13

recognize Blob type of spatial pattern [25]. It appears very difficult to define conclusions on the relative performance on the basis of such limit dataset [20]. For this reason, in this paper we analyzed much more extensive sets of manufacturing data provided by engineering database in a famous semiconductor company and sufficient coverage of simulated data generated by artificiality. Table 2.1 illustrates the comparisons between previous methodology with accuracy and data set.

Table 2.1 The comparisons between presented methodologies.

Literature

Feature selection method

Pattern Data set Accuracy

F.L. Chen and S.F. Liu, A neural-network approach to recognize defect spatial pattern in

semiconductor fabrication, 2000

-- line, ring 14 real data 60%

K.P. White et al., "Classification of defect clusters on semiconductor manufacturing wafers via the Hough

transform", 2005

L.J. Wei, "Development of wafer bin map pattern recognition model - using

neural network approach", 2006

5 features

bull eye, edge, blob, ring and line

57 real data 93.75%

S.F. Liu et al.,"Wavelet transform based wafer defect map pattern recoginition system in semiconductor

J.W. Cheng et al., "Evaluating performance of different classification

Based on this automatic classification methodology, we could obtain the information on problems related to the signature maps. Finding a process root cause is

14

not an easy task. Thus, a greater accuracy of classification is needed for engineering analysis. Subsequently, the process history of each wafer is used to create a list of the process step and keep track of which equipment must be responsible the problem in analysis. This paper proposed the use of five types of features from the semiconductor CP map on the wafer and evaluates the performance of several classification algorithms. We will demonstrate that the result present in more high accuracy with sufficient coverage of data set.