光電裝置輔助設計建模工具之基於類神經網路發展自適應擬合參數最佳化

Abstract

在1980年代，高效有機發光二極管（OLED）的發現，引起了對有機半導體的廣泛關注。然而，精準的物理性質很難明確定義，例如載流子遷移率（carrier mobility）或密度狀態（DOS）。有許多需要被確定的參數，這些參數在測量中通常存在著爭議。對於參數有多種可能性的組合來說，研究人員很難快速且精準的確認出合適的值。在本論文中，我們基於AI技術開發了自動擬合程序。我們將使用卷積神經網絡（CNN）來擬合數據。通過CNN的梯度下降行為，可以減少擬合過程的時間並優化參數且匹配實驗數據。在CNN模型的幫助下，與暴力搜索方法相比，擬合時間顯著減少。此外，CNN還可以用於優化LED效率。根據該結果，對於有漸變成分組成的EBL（GEBL）UVLED，據由調整GEBL中的Al濃度，效率下降率從7.83％降至6.35％且啟動電壓由4.96V降至4.86V。通過適當配置CNN模型，我們可以最大限度地利用計算資源來協助工程師進行裝置設計。

The discovery of highly efficient organic light-emitting-diodes (OLEDs) in the 1980s has attracted extensive attentions on organic semiconductors and devices. However, the accurate physical properties is difficult to be defined clearly. For example, carrier mobility or density of states for organic materials are difficult to be identified correctly. There exist many parameters to be configured out and those parameters are usually controversial in the measurement. With so many possible configurations of parameters, it is very hard for researchers to quickly sort out the accurate configuration. In this thesis, we developed an automatic fitting programs based on AI techniqies. We used convolutional neural networks (CNN) for data fitting. By the gradient descent behavior of CNN, it can reduce the time for fitting process and optimize parameters to match the experimental data. With the assistance from CNN model, the fittng time is significant reduced compared to brute-force search method. The parameters obtained from CNN model shows a good agreement to the experimental result. In addition, CNN can also be used to optimize LED efficiency. According to the results, it is reduced the efficiency droop from 7.83\% to 6.35\% and reduced the turn on voltage from 4.96V to 4.86V for a UV LED with gradual composition EBL (GEBL) by adjusting the aluminum composition in GEBL. With a proper configuration of CNN model, we can maximize the use of the computing resources to assist engineers for device designs.