dc.description.abstract | Yield rate, or defect-free rate, is one of many indexes to evaluate semiconductor manufacturers’ competitiveness. Manufacturers not only strive to introduce new products to the market earlier than their competitors, but also have to maintain a steady yield rate so as to ensure their competitive edge in the industry. All data related to the laser component manufacturing process is recorded in a CIM system. However, the multidimensional data is too numerous to be counted. Lots of time and manpower would be consumed if data is classified and analyzed manually or using the conventional statistic methods.
Upon receiving the laser chip-process and packaging-process data from a chosen corporation, the author started to mine data in two phases. In Phase 1, highly complicated machine parameters were obtained from the crystal particle manufacturing process for data preprocessing and thus data was converted into the data-mining format. In Phase 2, several crystal particles’ key machine parameters and packaged finished products’ yield rates were integrated into the data; the finished products were classified using the yield rate and the batch yield rate specified by the chosen corporation. After that, various classification techniques such as decision tree, neural network, and Support Vector Machine (SVM) are used to evaluate the machine parameters and to assess the results of finished products’ yield rates obtained from data-mining. Lastly, the analysis results were presented to process technicians for validation and evaluation.
The author attempted to identify the relationship between the key parameters as well as the implicit parameters related to crystal particle manufacturing process and thereby presented the basic specifications for process technicians’ reference. Most importantly, the experimental results indicate that the semiconductor laser process itself was sufficient to identify the key parameters of the manufacturing process as well as the machine parameters that affect the yield rate but they are not controlled at this moment. The results also indicate that CART, when boosted by multiple classifiers, is the most effective algorithm. | en_US |