Abstract
With the burgeoning development of semiconductor industry and more complicated wafer fabrication process to come, semiconductor industry has been confronted with high quality demand, high cost pressure and strict production environment. Therefore, a lot of semiconductor manufacturers strive for system development of yield rate enhancement and manufacturing flow improvement, hoping to achieve the objectives of quality promotion and cost reduction. In recent years, many semiconductor foundries have invested a large sum of capital in Advanced Process Control (APC) system, in which the most important part is the induction and development of Fault Detection and Classification (FDC) system. The engineers can immediately detect the unusual situations of operator machines on-line by FDC to improve the yield rate. Based on Support Vector Machines (SVM) and Decision Tree theories, the study tries to construct FDC system on heating curves of soft bake in photolithography area. To get characteristics’ memberships of certain heating curve as the input values of SVM system, and this SVM system can classify raw data to normal or abnormal category at first stage. Through data collection and processing abnormal curves to obtain characteristics, this study adopts Classification and Regression Tree (CART) to construct its rule base. By using 410 process data collected from one wafer fabrication factory in Taoyuan to test and verify FDC system can 42 abnormal curves be identified successfully. The only 6 curves misjudged out of the other 368 normal curves are possibly the signs of aberrant process. And the characteristics' memberships of these 6 curves are located at the fuzzy region, which possibly makes the system misjudge. The research finding proves that the SVM system has better detection and classification efficiency. Processing engineers are able to detect an abnormal curve immediately, identify its situation correctly via the rule base, find out