Statistical comparison of fault detection models for semiconductor manufacturing processes

A variety of statistical and data-mining techniques have been developed for the fault detection (FD) modeling of semiconductor manufacturing processes over the past three decades. However, few studies have analyzed which models are adequate for different types of fault data. In this paper, we define a FD model as an algorithm combining feature extraction, feature selection, and classification. We prepare six process data scenarios and collect data by simulating an etching tool. In total, 117 possible algorithm combinations are tested as FD models for the six datasets. With these test results, we conduct statistical analyses from two perspectives: 1) the algorithm perspective and 2) FD model perspective. From the algorithm perspective, we compare the performance of competing algorithms in the three model-building steps using multiple comparison methods and discuss the advantages and disadvantages of individual algorithms. From the model perspective, we determine which algorithm combinations are recommended for FD models of the semiconductor process and explain why some combinations do not exhibit the expected performance. In both analyses, we interpret some results using 3-D plots.