Collaborative semiconductor yield forecasting with a fuzzy-neural expert system

Yield forecasting is critical to a semiconductor manufacturing factory. To accomplish this, several experts (usually experienced product engineers) will gather to discuss and predict the yield of a product in a collaborative fashion. During the process, how to aggregate the subjective opinions of these experts with the forecasts from the yield forecasting model of the product is difficult. In order to tackle this problem and to enhance the precision and accuracy of semiconductor yield forecasting, this study constructs a fuzzy-neural expert system. The major parts of this system are: multiple experts, model fitting module, collaboration module, opinion modification module, forecasting module, aggregation module, and reporting module. In addition, a web-browser based interface is provided for multiple experts to construct their own fuzzy yield learning models from various viewpoints to predict the yield of a product. This allows the subjective opinions of each expert to be easily incorporated into their fuzzy yield learning model. It also means that these experts don’t need to gather at the same place, which makes collaborative semiconductor yield forecasting more convenient. Since expert opinions may not all be equally important, an online mutual assessment mechanism is used to determine the importance of the opinion of each expert. A two-step aggregation mechanism is used to aggregate these fuzzy yield forecasts. At the first step, the fuzzy intersection is applied to aggregate the fuzzy yield forecasts into a polygon-shaped fuzzy number to improve the precision of yield forecasting. After that, a back propagation network is constructed to defuzzify the polygon-shaped fuzzy number and to generate a representative/crisp value to enhance accuracy. For evaluating the effectiveness of the fuzzy-neural expert system and to make a comparison with existing approaches, all of these approaches were applied to the practical data of three products in a real world semiconductor manufacturing factory. According to experimental results, the fuzzy-neural expert system improved both the precision and the accuracy of semiconductor yield forecasting by 42% and 37%, respectively.