Please use this identifier to cite or link to this item:
標題: 化學機械研磨時控製程參數最佳化技術
CMP Time-Optimal Process Control Technique
作者: 周孟賢
Chao, MengShian
關鍵字: OST;正交自調法;CDST;CMP;MRR;WIWNU;共軛自調法;化學機械研磨;材料移除率;晶片表面不均勻度
出版社: 機械工程學系
本研究中之主要目的乃是希望結合田口實驗法與類神經網路,在實驗次數減少但不影響CMP製程模型建立下,以類神經網路建立一包含時間函數之類神經網路式CMP製程模型,進一步提出一低成本且有效之CMP Time-Over 製程參數最佳化技術。
而為能精確建立CMP研磨製程的模型,我們發展出一套能隨學習軌跡之改變而自我調適參數之多層前饋式類神經網路(Feedforward Neural Network)適應學習法則。電腦模擬證明本研究中所提出之自調法則,在收斂成效與其它學習法則相當下,卻擁有參數完全自調之優點。

The application of CMP in semiconductor procedure is inevitable. So, in the process, a good detection system is the keypoint to improve the performance of CMP.
In this thesis, orthogonal array technique in the Taguchi method is combined with neural network and that is proposed to structure the CMP model. Decrease the experiment trial, the CMP model with time factor is built. In advance, a low cost and efficient systematic approach to achieve global optimal Time-Over CMP process is carried out
To detail build the CMP abrasion procedure model, a new and efficient adaptive-learning self tuner algorithm for multiplayer feedforward neural network is proposed. Computer simulation shows that the proposed approach have very good performance both in converging speed and converging number stdev compare with other algorithm.
The cost function J=a*MRR(S/N)+b*STDEV(S/N) is proposed to involve the performance index MRR and WIWNU which form a set of training pattern to the neural network, when built CMP model. With the neural network based system model, optimal CMP procedure parameters included time variable can be obtained through neural network based optimal design method. Setting the time variable with optimal parameters can achieve the CMP process when WIWNU fit the performance. Even thought the process time is reduced by 1/3 experimental results show that the proposed method can perform better than the original approach.
Appears in Collections:機械工程學系所

Show full item record

Google ScholarTM


Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.