臺灣博碩士論文加值系統

在先進半導體製程上，如何準確利用光刻模型描述CMOS節點為很重要的一項課題，傳統光阻模型通常使用由大數據趨勢建構出的嚴謹光阻模型。然而，因為許多製程上的變異，會導致此傳統光阻模型變得不夠精確，例如：光學系統中的像差和光阻的厚度等等。因此，在本研究中，我們使用深度學習來建立光阻模型，旨在準確預測光阻輪廓。此外，我們提出了一個獨立於製程條件的模型。
此研究中，將光刻模型分成光學模型與光阻模型兩部分。使用MATLAB軟體撰寫光學模型，並藉由Hopkin演算法來模擬生成光源經過光罩後的光強分佈。我們研究了兩種深度學習的方法來模擬光阻模型，分別為卷積神經網路（CNN）與全卷積網路（FCN），並結合光學模型中部分同調光源的光強分佈計算，目標是能夠提供精確的光刻模型。在此架構下，光場分佈為深度學習模型的輸入資料，SEM影像經處理過後的光阻輪廓為輸出資料，在優化訓練後得到模型結果。
我們設計了一個週期為800奈米、半週期為400奈米的正方形孔洞斜向陣列，並使用Canon FPA-3000i5 +Stepper在8英寸矽晶圓上曝上這些圖案，其中數值孔徑(NA) = 0.63，波長= 365 nm（i-line），部分同調因子(partial coherence factor)σ= 0.8，利用SEM拍攝顯影後的影像，並使用Python圖像處理技術生成SEM影像的光阻輪廓。卷積神經網路（CNN）與全卷積網路（FCN）是由卷積層和完全連通的密集層(類神經網路)組成的兩種深度學習架構，我們系統性地研究了它們的收斂行為以及不同結構、不同訓練數據的影響等。計算光阻的目標輪廓和預測輪廓之間的邊緣放置誤差。特定製程之光刻模型(The process specified resist model)顯示，2層卷積層、2層池化層、3層全連接層、2層dropout層的九層卷積神經網路(CNN)，平均和最大深度學習誤差指數分別為5.2nm和23.3nm，對於具有6個卷積層、4層池化層、6層反卷積層與5層上採樣層的二十一層全卷積網路(FCN)，平均和最大深度學習誤差指數分別為5.6nm和26.9nm。製程獨立之光刻模型(The process independent resist model)結果顯示，對於具有7層卷積層、3層池化層、4層全連接層、3層dropout層的十七層卷積神經網路(CNN)，平均和最大深度學習誤差指數分別為8.6nm和54.3nm，對於4層卷積層、3層池化層、5層反卷積層與4層上採樣層的十六層全卷積網路(FCN)，平均和最大深度學習誤差指數分別為14.6nm和53.3nm。

Lithographic models that accurately describe the optical and resist behaviors of any given CMOS nodes are vital to modern semiconductor manufacturing. The conventional model commonly uses a compact resist model which is built up by enormous data trends. However, the compact resist isn’t very accurate because of the process variation, such as optical aberration and thickness of photoresist, etc. Therefore, in this study, we use deep learning to build up a resist model that aims to accurately predict resist profiles. Furthermore, we propose a model that is independent of process conditions.
In this work, lithography is coupled with the optical model and photoresist model. The optical model that generates an aerial image of the mask is simulated by the Hopkin approach algorithm using MATLAB. We investigate a deep learning algorithm based on convolutional neural networks (CNN) and fully convolutional networks(FCN) to simulate the photoresist model, and in conjunction with the aerial image calculation based on partially coherent illumination conditions, we aim to deliver an accurate lithographic model. In this framework, the aerial image is used as the input for the deep learning model undergoing optimizations, while the output is photoresist contours obtained from SEM images (training) or predictions (testing).
We designed a square-via array with a half pitch CDs 400 nm and expose those patterns on an eight-inch silicon wafer using a Canon FPA-3000i5+ Stepper where NA= 0.63, wavelength= 365 nm(i-line), and partial coherence factor σ=0.8. The resist profile of the SEM image is generated by applying image processing techniques using Python. Two deep learning architectures, namely, CNN and FCN, composed of convolutional layers and fully-connected dense layers are investigated, where we systematically study their convergence behavior, the impact of layer structures, training data, etc. An edge placement error is also calculated based on the difference between the target and predicted photoresist profiles. The process specified resist model shows that for CNN with 2 convolutional, 2 max pooling, 3 dense layers and 2 dropout layers, the average and maximum deep learning error figure are 5.2nm and 23.3nm, respectively, and for FCN with 6 convolutional layers, 4 max pooling, 6 de-convolutional, 5 up sampling layer, the average, and maximum deep learning error figure are 5.6nm and 26.9nm. The process independent photoresist model result shows that for CNN with 7 convolutional, 3 max pooling, 4 dense layers and 3 dropout layers, the average and maximum deep learning error figure are 8.6nm and 54.3nm, respectively, and for FCN with 4 convolutional layers, 3 max pooling, 5 de-convolutional layer and 4 up sampling, the average, and maximum deep learning error figure are 14.6nm and 53.3nm.

摘要 I
ABSTRACT III
誌謝 V
目錄 VII
圖目錄 IX
表目錄 XII
第 一 章 、緒論 1
1.1 光學微影發展 1
1.2人工智慧發展 3
1.2.1人工智慧 4
1.2.2機器學習 4
1.2.3深度學習 5
1.3研究動機 5
第 二 章 、研究原理 7
2.1光學微影 7
2.1.1光學微影系統 7
2.1.2光源 8
2.2.3光罩 10
2.1.4光阻 10
2.2微影系統之成像原理 11
2.2.1 Hopkin’s approach 12
2.2.2 Sum of Coherent System(SOCS) 12
2.3深度學習 14
2.3.1類神經網路 14
2.3.2卷積神經網路 16
2.3.3全卷積網路 16
2.4實驗流程與設計 17
第 三 章 、研究方法 19
3.1微影製程與量測分析 19
3.1.1 光罩設計 19
3.1.2 黃光微影製程 21
3.1.3 掃描式電子顯微鏡(SEM) 24
3.1.4影像分析技術 27
3.2微影系統成像之模擬 28
3.3深度學習之建立 29
3.3.1資料之預處理 30
3.3.2深度學習之模型優化 31
第 四 章 、研究結果 32
4.1 特定製程之光刻模型建立 32
4.1.1製程變異 32
4.1.2模型訓練結果 34
4.2 製程獨立之光刻模型建立 40
4.2.1製程變異 40
4.2.2模型訓練結果 42
第 五 章 、總結與未來展望 47
第 六 章 、參考文獻 49

[1] http://www.icinsights.com/news/bulletins/Semiconductor-Unit-Shipments-Exceeded-1-Trillion-Devices-In-2018/
[2] Lin, Burn Jeng, and Burn Jeng Lin. "Optical lithography: here is why." Bellingham: SPIE, 2010.
[3] https://en.wikipedia.org/wiki/CMOS
[4] Chou, Stephen Y., Peter R. Krauss, and Preston J. Renstrom. "Nanoimprint lithography." Journal of Vacuum Science & Technology B: Microelectronics and Nanometer Structures Processing, Measurement, and Phenomena 14.6 (1996): 4129-4133.
[5] Altissimo, Matteo. "E-beam lithography for micro-/nanofabrication." Biomicrofluidics 4.2 (2010): 026503.
[6] https://en.wikipedia.org/wiki/Photolithography
[7] Mack, Chris. Fundamental principles of optical lithography: the science of microfabrication. John Wiley & Sons, 2008.
[8] Official website of NVIDIA: https://blogs.nvidia.com/blog/2016/07/29/whats-difference-artificial-intelligence-machine-learning-deep-learning-ai/
[9] AlaphaGo: https://deepmind.com/research/alphago/
[10] Gibney, Elizabeth. "Google AI algorithm masters ancient game of Go." Nature News 529.7587 (2016): 445.
[11] Yang, Haoyu, et al. "GAN-OPC: Mask optimization with lithography-guided generative adversarial nets." 2018 55th ACM/ESDA/IEEE Design Automation Conference (DAC). IEEE, 2018.
[12] Lin, Yibo, et al. "Machine learning for yield learning and optimization." 2018 IEEE International Test Conference (ITC). IEEE, 2018.
[13] Ding, Duo, et al. "Machine learning based lithographic hotspot detection with critical-feature extraction and classification." 2009 IEEE International Conference on IC Design and Technology. IEEE, 2009.
[14] 林大貴. TensorFlow+ Keras 深度學習人工智慧實務應用. Vol. 21710. 博碩文化股份有限公司, 2017.
[15] https://www.inside.com.tw/feature/ai/9854-ai-history
[16] Alpaydin, Ethem. Introduction to machine learning. MIT press, 2009.
[17] Bottou, Léon, Frank E. Curtis, and Jorge Nocedal. "Optimization methods for large-scale machine learning." Siam Review 60.2 (2018): 223-311.
[18] LeCun, Yann, Yoshua Bengio, and Geoffrey Hinton. "Deep learning." nature 521.7553 (2015): 436.
[19] Watanabe, Yuki, et al. "Accurate lithography simulation model based on convolutional neural networks." Optical Microlithography XXX. Vol. 10147. International Society for Optics and Photonics, 2017.
[20] Ma, Xu, et al. "Fast lithography aerial image calculation method based on machine learning." Applied optics 56.23 (2017): 6485-6495.
[21] Watanabe, Yuki, et al. "Accurate lithography simulation model based on convolutional neural networks." Optical Microlithography XXX. Vol. 10147. International Society for Optics and Photonics, 2017.
[22] Basting, D., N. Djeu, and K. Jain. "Historical review of excimer laser development." Excimer Laser Technology. Springer, Berlin, Heidelberg, 2005. 8-21.
[23] Schaller, Robert R. "Moore's law: past, present and future." IEEE spectrum 34.6 (1997): 52-59.
[24] https://en.wikipedia.org/wiki/Moore%27s_law
[25] Moore, Gordon E. "Cramming more components onto integrated circuits." (1965): 114-117.
[26] Moore, Gordon E. "Progress in digital integrated electronics." Electron Devices Meeting. Vol. 21. 1975.
[27] https://en.wikipedia.org/wiki/Photomask
[28] https://en.wikipedia.org/wiki/Photoresist
[29] Van Steenwinckel, David, et al. "Resist effects at small pitches." Journal of Vacuum Science & Technology B: Microelectronics and Nanometer Structures Processing, Measurement, and Phenomena 24.1 (2006): 316-320.
[30] He, D., and F. Cerrina. "Process dependence of roughness in a positive-tone chemically amplified resist." Journal of Vacuum Science & Technology B: Microelectronics and Nanometer Structures Processing, Measurement, and Phenomena 16.6 (1998): 3748-3751.
[31] Wong, Alfred Kwok-Kit. "Optical imaging in projection microlithography." Bellingham, WA: SPIE, 2005.
[32] J.-C. Yu, Inverse Mask and Source Synthesis for Resolution Enhancement in Optical Lithography, NCTU, 2012.
[33] Shanmuganathan, Subana. "Artificial neural network modelling: An introduction." Artificial Neural Network Modelling. Springer, Cham, 2016. 1-14.
[34] http://alexlenail.me/NN-SVG/index.html
[35] https://isaacchanghau.github.io/post/activation_functions/
[36] Sarle, Warren S. "Neural networks and statistical models." (1994).
[37] http://alexlenail.me/NN-SVG/LeNet.html
[38] Simonyan, Karen, and Andrew Zisserman. "Very deep convolutional networks for large-scale image recognition." arXiv preprint arXiv:1409.1556 (2014).
[39] Farabet, Clement, et al. "Learning hierarchical features for scene labeling." IEEE transactions on pattern analysis and machine intelligence 35.8 (2012): 1915-1929.
[40] Krizhevsky, Alex, Ilya Sutskever, and Geoffrey E. Hinton. "Imagenet classification with deep convolutional neural networks." Advances in neural information processing systems. 2012.
[41] https://blog.csdn.net/Next_FSE/article/details/85225008
[42] Michael Quirk and Julian Serda, 2nd, Semiconductor Manufacturing Technology,2000, ch.13.
[43] Official website of Taiwan Semiconductor Research Institute: https://www.tsri.org.tw/tech/equipment/equipment_main.jsp?equiId=7
[44] Michael Quirk and Julian Serda, 2nd, Semiconductor Manufacturing Technology,2000, ch.14.