本研究主要針對半導體批次製程模式,運用倒傳遞類神經網路發展一製程即時控制系統。藉由倒傳遞類神經網路非線性的函數對應能力,針對製程產生不同程度之雜訊 (noise)、偏移 (shift) 與雜訊及飄移 (drift) 與雜訊等干擾進行控制,並且與傳統EWMA運算式為基之控制器進行控制效益之比較。此外,研究中更針對控制器間的控制速度進行探討,進而驗證倒傳遞網路控制器在製程控制上的執行效益。最後則以MISO (multiple input, single output) 系統加以探討此控制器之執行能力,以驗證本研究在實際半導體批次製程控制上的可行性與實用性。 本研究經驗證後可歸納下列三點:一、本研究所提出之BPNN控制器在各種不同之製程干擾下,其控制效益均較EWMA運算式為基之控制器為佳。二、於控制速度方面,BPNN控制器於前段批量已可達到即時控制,於後段批量亦可維持高度之平穩性。三、多個製程參數的線外訓練,將有效提高製程的估測能力,並對於整體之控制效益均可有效提昇。
The purpose of this research is to develop a Back-propagation neural network-based (BPNN) run-to-run real-time process controller for the common process disturbances in semiconductor industries. We consider using BPNN for controlling the process disturbances, including the process noises, shifts and drifts. The performances of control result were compared with the EWMA controllers. On the other hand, we compared the control speed between the controllers and considered the MISO (multiple input single output) system for the follow-up research. After the study, we conclude that the BPNN controller outperform the EWMA controllers in the different process disturbances. The BPNN controller has higher stability and real-time control ability. The MISO system promotes the process estimation and prediction efficiently, and the demonstration of the BPNN controller performance can arise the application and feasibility in the practical process.