Research On Statistical Modeling And Run-to-Run Control Of Semiconductor/TFT-LCD Manufacturing Processes

With the development of modern industry, the semiconductor/TFT-LCD manufacturing processes are becoming larger and more complex. At the same time, the manufactured flows have also getting more and more precise which can make the equipments become more expensive. So it's difficult to improve the yield of products. Aiming at the challenges of the semiconductor/TFT-LCD processes, some modeling and controller designed algorithms including statistical theory, data mining method and Run-to-Run control etc., are proposed in this paper. Through the proposed methods, the characteristics of process can be understood and the quality of products can be estimated and improved timely.The contributions in this paper as follows:1. Based on the data collected from the advanced process control system, a virtual metrology by using the dynamic canonical correlation analysis (DCCA) is presented. Then, some representative aggregative indicators which maximize the correlation between the process variables and quality variables are selected, and the high dimensionality in the input data can be settled by projecting the original process variables onto a space defined by orthogonal principal components (PCs). To overcome the nonlinear of manufacturing processes and uncertainty caused by preventive maintenance, the recursive forms for mean and covariance matrix are derived and virtual metrology model based on DCCA is presented. Comparing with the traditional principal component analysis and partial least squares in the numerical simulation, the proposed model shows several good characteristics. Superiority of the proposed model is also presented when it applied to an industrial sputtering process.2. In the modern manufacturing industry, the same products or the products with same recipe will be produced on different tools, and many different products with different recipes will also be produced on the same tool. This production mode is named as high-mix manufacturing process and is very common in semiconductor/TFT-LCD industries. To estimate the product quality of the high-mix manufacturing process, a virtual metrology based on MANCOVA and IMA(1,1) is proposed, which combines the advantage of both standard statistical methods and time serial analysis. Based on standard stepwise regression, the variables selected into the VM model are termed "'key variables" and may contain important information regarding the source of variation of the current process condition. Then, the statistical MANCOVA technique is adopted to build conjecture model for products with different specifications considering the product effects. On the other hand, by introducing the concept of time serial, the IMA (1,1) would reflect the current unmeasured information of the process, and thus the accuracy of the VM model would be substantially improved. Applications on the wet etching process demonstrate that the VM model built by the proposed approach offers good predictions of product quality of different products, which are highly desirable for semiconductor/TFL-LCD industrial applications.3. A z-EWMA algorithm is investigated under the flexible process conditions and zone target. For the constraints adjustment of some process, the acceptability of flexible inequality constraint condition that affects the system dynamics are described quantificationally. Then, an adjustment index which depends upon the prior information is employed in this paper. As a result of control, the consumption of manufacturing process will be reduced. The advantages of proposed control scheme are demonstrated by benchmark simulation and reversed engineered industrial applications.4. Tools or chambers at a single step are designed to perform the same processing in semiconductor industry. In practice, tools or chambers differ and do not process lots identically. To diagnose any reasonable difference between golden and inferior chambers, a performance matching algorithm based on statistical methods is proposed in this paper. Firstly, several SVID approaches are provided to reduce the high dimensionality of process variables. Then, F-test is used to diagnose the golden and inferior chambers. Thirdly, a synopsis of analysis results in the chart of R2 statistics vs. p-value is presented. This framework provides a systematic method of drawing inference from the available evidence without interrupting the normal process operation. The proposed concept is illustrated by an electroplating process in a local fabrication unit...