Predication Of Filter Flow Rate Of JK750Machine Based On RBF Neural Network

No matter early250nm made technology and now of28nm technology, and wet cleaning technology plays a very important role in the process of semiconductor manufacturing, and the control of chemical flow rate and concentration is one of core technology of wet cleaning. Filter change control can’t be ignored during the flow rate control of chemical. This paper research is related to the filter flow rate prediction and filter replacement optimization within one filter lifetime of JK750machine chemical tank in the backend of line (BEOL) processing of semiconductor. It mainly solves the waste of premature filter replacement and eliminates the yield impact while EP (Excursion Prevention) alarms during product running and that contribute the processing cost savings. The main research and work of the thesis is illustrated as follows:(1) List the background of the filter flow rate prediction application of JK750machine and the value of its, introduce the methods and measures of filter replacement in the major semiconductor manufacturers at home and abroad and simply analyze the all kinds of typical neural network prediction model of nonlinear system structure, its advantages, its disadvantages and other applications.(2) Study the filter working principle of JK750machine and the flow rate impact of filter replacement, and present the technical route and technical measures how to establish the filter flow rate predicative model of JK750machine.(3) The autocorrelation coefficient between output data and previous input data, the gradient descent algorithm and predication result of different neurons quantity are discussed based on RBF network filter flow rate prediction model of JK750machine. Theory analysis and simulation research of the prediction model are carried out, which is provided theory basis for filter flow predictive modeling of JK750machine further.(4) Improve the structure of the predication model of RBF network in order to enhance the predicative precision and reduce the error of prediction model and at the same time the optimal clustering number is established with fuzzy clustering algorithm research and application, the quantity of neurons. After the simulation calculation, the various parameters&value of the model are received. Finally, the results of simulation demonstrate this prediction model is more effective. (5) The improved prediction model is applied effectively and successfully in JK750machines with real-time flow rate output and the software and hardware are developed in the practice platform respectively. An improved RBF network prediction model with fuzzy clustering algorithm is used in this practice platform. The results of simulation demonstrate it can effectively and accurately forecast filter flow rate in the next moment, which provides precise schedule of filter replacement.In recent years, the competition of large-scale semiconductor manufacturing enterprises became fiercer in the market. In the face of the shrinkage of the degrees of profits, they have to be committed to focus cost saving and yield improvement in order to enhance the competitiveness of the enterprise themselves. The application of the model is to optimize filter replacement of JK750machine, to eliminate yield lost and to reduce the cost of the step processing. It is a valuable practice in the semiconductor manufacturing.