| Abstract: | 由於軟顯的特殊可繞性， 理論上其生產方式可由T F T- L C D 利用 photolithography 的序列生產方式（一層又一層），改進為捲軸的連續印刷方式 (Roll to Roll Printing; R2R)。快速、大量、低價的生產方式，是軟顯造成影響的主要原因之一。然而由於功能需求，軟顯的基板素材以塑料為主，其對溫度、張力、基板厚度、壓力以及其他的因素極為敏感。換言之，在生產過程中的環境控制（非靜態的空氣中微粒數，而是更困難的動態環境變數）對生產良率極為重要。工研院於今年成立「軟電量產開發實驗室」，也是台灣首座軟電實驗室，目的在於開發具有量產技術能力的生產製程，期待能開發具有材料合成、製程研發、產品設計到試量產一貫完成的生產方式。從其製程研發來看，軟式系統整合(System Design for Roll-to-Roll)未來可能成為連續式(R2R)製程上重要的整合系統，但其主要的研究項目即在於：1)分析重要的生產參數，2)定義各生產參數的相依關係，3)監控製程的各項參數變化並即時調整製程的穩定性。如果以上的問題無法有效掌握，軟性基版在Roll-to-Roll 製程中將產生嚴重的對位問題，致使其良率大幅降低。本計畫基於在製程上仍然沒有一個標準的控制與檢驗模式，以及將面對的問題具有多變量、非線性，高度動態及快速反應的特質與要求，將利用三年期進行相關議題探討。第一年將著重利用實驗設計以及多變量建模方式，確立須監督以及控制的變數群，及其相依關係。第二年將採用類神經網路(Artificial Neural Networks)（具處理非線性與快速反應的能力)結合多變量理論(Kolmogorov Theorem)建構具快速收斂與回應的多變量模式。將以R2R中的某一製程作為研究對象。第三年將利用貝氏網路(Bayesian Networks)(具處理高度動態以及快速分類的能力)，利用其優異的分群能力，建構多模式化(multi-mode)的動態控制機制，以解決現場生產的動態變異所導致的品質問題。並將以R2R的連續生產為研究標的。此研究目的為改善R2R製程的控制系統模組並提升生產良率。此外對整合人工智慧方法於製程管控一應用上具有一定之學術貢獻，而且將有助於建立工業工程的知識在未來軟性電子與軟性顯示器產業上的貢獻。另外在實務應用的預期貢獻：本研究成果將有助於R2R製程順利提升良率並進入量產；並提供業界未來在導入新製程後，順利建立標準化之製程管控。
It is feasible to produce the flexible display (FD) in batch mode and using the existed facilities of TFT-LCD. However, it is plausible in producing the FD in Roll-to-Roll (R2R) Printing mode due to its attractive parallel processing and low cost advantages. R2R is functioning in a very fast speed rolling process and the substrate will have to be changed from glass to plastics, such as ITO. However, due to the nature of the plastic substrate, it is very sensitive to the temperature, tension, thickness of the film and the pressure operated in the process. That means the traditional conditions to the semi com and TFT-LCD are not crucial as usual. The new constraints are embedded in the new manufacturing processes. Achieving the acceptable yield rate becomes a new challenge to the new product and new processes. It is the fact that the new industrial standard R2R manufacturing process is not existed yet due to the multivariate, nonlinear, highly dynamic natures and the very fast response requirement. In this research, we are to adopt a three phases in three years term. The first year, a feature selection research will be conducted to determine the set of decisive and sufficient features. Upon those features, the data will be collected for the later part of research. The second year, due to the quick response requirement, an Artificial Neural Network model is adopted to perform as the basic model for the control mechanism. However, due to the multivariate nature, the Kolmogorov representation theorem will be investigated to design a new architecture of the neural network. In the third year, the Bayesian network approach will be conducted to resolve the dynamic problem. Via selecting the right and sufficient variables, constructing a more accurate and fast responsive control mechanism and with dynamic adjustment, the yield rate problem of R2R manufacturing process can be resolved to certain degree. Besides, the new approaches in this research will be investigated deeply and the practical contribution to the industry is foreseeable. |
