Algorithm Study Of SPA Based On Mixed-resolution Data

Run-to-run (R2R) processes are commonly seen in semiconductor manufacturing, which contains uninterruptable cycles that are termed as runs. Due to practical issues such as parameter estimates bias, process shifts, raw materials variances, and component breaking down, modeling, parameter estimation and feedback control are required for control a R2R process to achieve high output quality.In certain R2R processes, online accurate measurements are difficult to obtain. Instead, only categorical observations can be observed online. In some situations, accurate measurements may be provided after certain amount of delay. Currently, most conventional R2R controllers cannot be applied if no continuous observations are available. Some parameter estimation methods that can handle categorical data only use historical dataset in an offline manner. Since observations are usually collected gradually in a R2R process, it becomes meaningful to develop online algorithms for parameter estimation and process control based categorical or mixed-resolution information.Taking the lapping process in semiconductor manufacturing as an example, this work has finished the following research tasks: 1) developed two different online approaches, namely an Adjusted Maximized Likelihood method and a Bayesian method, for parameters estimation using categorical observations with white noises; 2) constructed parameter updating models with delayed accurate data for processes with both white noises and AR(1) disturbance series; 3) derived control strategies to generate receipts between runs.The computational results of performance evaluation show that the proposed methods are capable of estimating unknown parameters and controlling output quality online when initial bias exists; comparisons with other approaches also show that performances of the proposed methods already reached, even exceeded current online estimation and control methods based on accurate measurements.