The Effect of Additional Information on the Prediction Quality of Wafer Fabrication Operation with a Neural Network and Clustering Approach

Most of the performance assessment of semiconductor manufacturer is based on their self-appraisal or subjective judgments. The needs to measure fabrication (fab) operation performance along with its various dimensions have led to the development of a large number of quantitative performance indicators. An overall scheme to measure the performance of fab operation involving multi-input and multi-effects (output) has not been well established yet. In this study, we approach the performance assessment and prediction by combining clustering approaches with Artificial Neural Networks (ANN) approaches. We use historical data from a Taiwan semiconductor major player which comprise input/ investment data (such as headcount, salary, cost for machines, running the fab, etc.) as well as output of each fab (such as margin, waferoutput rate, stepmove, number of patents). The data comprise several years, during which some of the fabs have been ramped up. In the first phase of our approach, we studied several clustering algorithms (K-Means, X-means, Kernel K-Means, SIB, and EM) on the data. We found several clusterings that were meaningful according to human experts. One of the clustering approaches clearly divided one older fab from newer fabs, and also was able to distinguish fabs in ramping up phase from fabs that are in stable operation phase. Other approaches formed clusters according to the grade of performance (bad, medium-bad, medium-good, good) of the data sets. Second, we use the classification to let a neural network learn the status of a fab, so that for a new fab, the status can be judged by the neural net. In a third step, we let a neural network learn the relationship between the multiple inputs and outputs. As result we found a neural net structure that is able to predict changes in the inputs of a fab on the different output factors. By this we enable the fab management to obtain a prediction, which effect a planned measure (e.g. increase or decrease of headcount) has on the output of the fab, that is the performance figures.