近年來人工智慧議題發展迅速，而其中機器學習領域中神經網路模型的部分，能在經由蒐集大量資料經由神經網路學習訓練後，能達到分析辨識資料的目的，而現在硬體設備比往年提升了不少，不只原始資料能更大量的產生，且能使神經網路訓練的時間更快速，在有了大量的資料幫助下，能使的神經網路的模型達到更好的效果。
　　在半導體元件模擬中，我們NEGF模擬的單位如果越精細則模擬的效果越好，而網格尺度等級為奈米等級，為了精準的模擬去接近真實情況，當我們將網格尺度單位切分越精細的情況下，原始程式在電腦中的運算時間會越來越久，且爾後如果在添加各種不同更複雜的新條件去運算的話，可想而知需要花掉更大量的時間。
　　我們想利用神經網路模型輔助NEGF模擬，將我們花了大量時間得到的一部分模擬資料當作神經網路模型的訓練資料，當神經網路模型訓練完成，便能夠快速方便的得到跟原始資料範圍內誤差不大的結果，目前大部分參數所得到的相對誤差在百分之五之內，部分超出原始資料的新資料當作輸入時，也能得到不錯的結果，由此我們便不需要不斷花大量的時間去更動或運算半導體元件模擬的程式，而能直接經由神經網路模型得到我們所需的模擬結果，更甚者我們目前能經由Autoencoder-ANN結構的神經網路模型，由我們所想要的最佳I-V圖中得到的輸出參數去反推出元件參數，讓我們簡單且有效率的得到有好表現的元件的設計參數。

In recent years, the topics relevant to artificial intelligence have attracted much attention rapidly, and the neural network model in the machine learning field can achieve the purpose of analyzing and identifying data after learning and training through the neural network by collecting a large amount of data. Now, the hardware equipment is better than in the previous years. Since it has improved significantly, not only the original input data can be produced in a larger amount, but also the neural network training time is faster. With a large amount of data, the neural network model can be utilized to a wider range of applications, such as semiconductor electronics.
In research of semiconductor devices, the finer mesh in simulations, the better description of the device itself and physics behind, and the mesh scale may be in the sub-nanometer regime. If we desire to mimic the real situation in simulations accurately, we will need to employ the finer mesh and the computing time will be longer.
We use the neural network model to assist the quantum transport simulations in this work. Based on the part of device simulation data as training input for the neural network model, it can quickly and easily get the output, and the error compared with the original data is fairly small. The result shows that most of the parameters can have less than 5% of the relative error. And we use some data that out of the original data range as input and we still can obtain good results. So, we do not need to spend additional cost (e.g., computational power and time) to modify the input for the simulator or calculate the device characteristics. We can obtain the simulation results in need directly through the neural network model. Furthermore, we can use the neural network called Autoencoder, and we first define a good performance parameter from our simulation result (I-V curve) as input. Then we can get the device structure parameters as output by trained Autoencoder. It can help us to design the device structure simply and efficiently.

[1] Kahng, Dawon, “Electric Field Controlled Semiconductor Device” U. S. Patent No. 3,102,230 (Filed 31 May 31, 1960, issued August 27, 1963).
[2]Moore, Gordon E. (1965-04-19). “Cramming more devices onto integrated circuits”. Electronics.
[3] “International Technology Roadmap for Semiconductors”. Archived from the original on 2011-08-25
[4] H. Wakabayashi, et al., “Transport properties of sub-10-nm planar-bulk-CMOS devices” IEDM Technical Digest. IEEE International Electron Devices Meeting, 2004.
[5] S.M. Sze, Kwok K. Ng. “Physics of Semiconductor Devices”
[6] John R. Koza, et al. Artificial Intelligence in Design '96. Springer, Dordrecht. pp. 151–170. doi:10.1007/978-94-009-0279-4_9
[7]Yann LeCun, Yoshua Bengio & Geoffrey Hinton “Deep learning” Nature volume521, pages436–444 (28 May 2015)
[8] Carl Doersch, “Tutorial on Variational Autoencoders” arXiv:1606.05908v2 [stat.ML] 13 Aug 2016
[9] Yann LeCun1, Yoshua Bengio & Geoffrey Hinton “Deep learning” Nature,2015
[10] I.A Basheer, M Hajmeer “Artificial neural networks: fundamentals, computing, design, and application” Journal of Microbiological Methods
Volume 43, Issue 1, 1 December 2000, Pages 3-31
[11] Luke Dormehl “What is an artificial neural network? Here’s everything you need to know” 01.5.19
[12] Diederik P. Kingma, Jimmy Ba “Adam: A Method for Stochastic Optimization” Published as a conference paper at ICLR 2015
[13] Krizhevsky, A., Sutskever, I. & Hinton, G. ImageNet classification with deep convolutional neural networks. In Proc. Advances in Neural Information Processing Systems 25 1090–1098 (2012).
[14] G. E. Hinton* and R. R. Salakhutdinov “Reducing the Dimensionality of Data with Neural Networks” 28 JULY 2006 VOL 313 SCIENCE
[15] Arden Dertat “Applied Deep Learning - Part 3: Autoencoders” Oct 3, 2017
[16] Supriyo Datta, “Non-equilibrium green's function (NEGF) method: a different perspective”, Computational Electronics (IWCE), 2015 International Workshop on
[17] Zhibin Ren et al., “nanoMOS 2.5: A Two-Dimensional Simulator for Quantum Transport in Double-Gate MOSFETs”, IEEE TRANSACTIONS ON ELECTRON DEVICES, VOL. 50, NO. 9, SEPTEMBER 2003
[18] Supriyo Datta, “The Non-Equilibrium Green’s Function (NEGF) Formalism: An Elementary Introduction”, Electron Devices Meeting, 2002. IEDM '02. International.
[19] James Patterson and Bernard Bailey, “Solid-State Physics: Introduction to the Theory”, Springer
[20] Hsien-Ching Chung, “Electronic and Optical Properties of Monolayer and Bilayer Graphene Nanoribbons”, doctoral dissertation
[21] Charles Kittel, “Introduction to Solid State Physics”, 6th edition, John Wiley & Sons
[22] Edward McCann and Mikito Koshino, “The electronic properties of bilayer graphene”, 2013 IOP Publishing Ltd, Reports on Progress in Physics, Volume 76, Number 5
[23] V Nair, GE Hinton “Rectified linear units improve restricted boltzmann machines” Proceedings of the 27th international conference …, 2010
[24] S Ioffe, C Szegedy “Batch normalization: Accelerating deep network training by reducing internal covariate shift” arXiv preprint arXiv:1502.03167, 2015
[25] Albert Barabasi et al. “The human disease network” Proceedings of the National Academy of Sciences 104(21):8685-90 · June 2007Plos.org
[26] Alireza Makhzani, Jonathon Shlens, Navdeep Jaitly, Ian Goodfellow, Brendan Frey “Adversarial Autoencoders” arXiv:1511.05644v2 [cs.LG] 25 May 2016