Data-driven Prediction Method Of Multi-performance Index For Semiconductor Wafer Fabrication System

As the basis of information industry, semiconductor wafer fabrication system (SWFS) which has ever-changing market requirement and fierce industrial competition, knowing well about information of production line is an important method to improve product quality, ensure production efficiency and control cost budget. Performance indicators in line contains a large number of processing information, can be predicted to achieve production performance improvement and provide parameter support for subsequent optimization scheduling objectives. Therefore, this article focuses on the study of yield, throughput rate and the level of work in process(WIP). Analyze the characteristics of each index, and use the data-driven method to establish the forecasting model corresponding to each index.(1) To solve the problem of existing methods need large amount of data and rarely consider prediction range, a method of yield prediction by multi-agent fuzzy collaborative is explored, the multi-agent is used to replace the expert opinion, and solve fuzzy yield model parameters combined with linear programming. The results obtained by multiple yield models are aggregated using fuzzy rules, then the support vector regression (SVR)defuzzifys the fuzzy result, to realize the synchronous prediction of minimum range and the exact value.(2) According to the analysis of throughput rate prediction in SWFS, a prediction method which combines back propagation neural network(BPN)and the mathematical programming is proposed. the analysis of the relevant performance of the semiconductor throughput rate, Multiple linear regression model is constructed with key performance index which are selected by the principal component analysis(PCA) .The regression model is substituted into the mathematical programming method to adjust the parameters of BPN to improve the precision of traditional throughput rate models.(3) In view of analyzing dynamic time series of WIP in the semiconductor production line, a time series prediction method based on the echo state network(ESN) is studied. Simulation result shows the leaked integral echo state network(Leaky-ESN) predict the time series of WIP in SWFS is available, improve the prediction speed and accuracy.The simulation experiments show that these three kinds of prediction methods can realize prediction performance better, prediction accuracy is higher. It can provide parameter support for subsequent optimization scheduling objectives.