電路布局擺置後的結果，會嚴重影響繞線的方式。為了讓電路維持在最佳的效能以及排除非理想效應，在做佈局擺置的時候應該要有繞線成本的預估。在佈局擺置的階段，現行的繞線預估大多採用半周長來做推測，在多端點情況下可能不太正確。而且，對於較敏感的類比電路而言，繞線成本不單單只有線長而已，導線轉彎數以及線導孔的數量等參數也都會電路效能產生的影響。
本論文利用機器學習的技術，協助設計者進行早期電路繞線成本的預估，希望能提早知道符合實際情況的繞線成本，幫助設計者在佈局擺置時進行對應的調整，以避免掉一些不必要的重複設計。我們將利用人工神經網路的方式做機器學習，藉由擺置的資訊來預測線長、轉彎數、以及線導孔數量等繞線成本。如同實驗數據所示，經過學習的神經網路將能精準預估佈局擺置過後的繞線成本，提供設計者有用的參考資訊。

The placement results have large impacts on routing results. In order to keep circuit performance and eliminate non-ideal effects, we have to predict routing cost at layout placement stage. Most of current approaches use semi-perimeter method to predict the routing cost at placement stage. It might not be correct in multi- terminal routing cases. Moreover, for sensitive analog circuits, routing cost considers more than wire length only. The turn numbers of each metal line and the via numbers of each net will also effect circuit performance.
In this thesis, we use machine learning technique to help designer predict the routing cost at placement stage. With the predicted routing cost, we can make proper adjustment in advance to avoid unnecessary design iterations. Using artificial neural networks for machine learning, we can use the placement information to predict the routing cost, such as wire length, via numbers, and turn numbers. As shown in the experimental results, we can accurately predict wire length, via number, turn number base on the neural network models. They can be good references for designers to determine a good layout placement.

[1] Milan Hajek, “Neural Networks”, University of KwaZulu-Natal, 2005
[2] Wikipedia, “Physical Design (electronics)”, https://en.wikipedia.org/wiki/Physical_design_(electronics)
[3] Wikipedia, “Standard Cell”, https://en.wikipedia.org/wiki/Standard_cell
[4] 紀浩瑜 “以減少導線負載為目標的效能導向之類比電路繞線方法” 國立中央大學電機工程學系碩士論文, July, 2017
[5] Wikipedia. “Hopfield Network”, https://en.wikipedia.org/wiki/Hopfield_network
[6] Faith Ertam, Galip Aydin, “Data Classification with Deep Learning using Tensorflow”, 2017 International Conference on Computer Science and Engineering (UBMK)
[7] Google TensorFlow, https://www.tensorflow.org/
[8] Nvidia CUDA, https://developer.nvidia.com/
[9] Po-Hsun Wu, Mark Po-Hung Lin, and Tsung-Yi Ho, ‘Analog Layout Synthesis with Knowledge Mining’, 2015 European Conference on Circuit Theory and Design (ECCTD)
[10] Bo Liu, Francisco V. Fernandez, andGeorges Gielen, ‘Fuzzy Selection Based Differential Evolution Algorithm for Analog Cell Sizing Capturing Imprecise Human Intentions’, 2009 IEEE Congress on Evolutionary Computation (CEC 2009)
[11] Sheng-Yuan Lin, Jing-Yi Chen, Jin-Cheng Li, Wan-Yu Wen, and Shih-Chieh Chang, ‘A Novel Fuzzy Matching Model for Lithography Hotspot Detection’, 2013 50th ACM/EDAC/IEEE Design Automation Conference (DAC)
[12] Wen-Hsiang Chang, Chien-Hsueh Lin, Szu-Pang Mu, Li-De Chen, Cheng-Hong Tsai, Yen-Chih Chiu, and Mango C.T. Chao, ‘Generating Routing-Driven Power Distribution Networks With Machine-Learning Technique’, 2017 IEEE Transactions on Computer-Aided Design of Integrated Circuits and System
[13] Michal Puheim, Ladislav Madarász, “Normalization of Input and Outputs of Neural Network Based Robotic Arm Controller in Role of Inverse Kinematic Model”, SAMI 2014 IEEE 12th International Symposium on Applied Machine Intelligence and Informatics
[14] 李牧勳, “Front-End Supply Current Waveform Models for Dynamic IR-Drop Analysis”, 國立中央大學電機工程學系博士論文, June, 2012
[15] Wikipedia, “Semi-supervised learning”, https://en.wikipedia.org/wiki/Semi-supervised_learning