微影製程之疊對控制

Overlay Control for Lithography Process

林祖強(Tsu-Chiang Lin)

指導教授 ： 張耀仁

中原大學/工學院/機械工程研究所/碩士(2005年)

https://doi.org/10.6840/cycu200500664

摘要

微影是IC製程中影響關鍵尺寸最直接的因素之一，光罩圖案轉移的優劣會直接影響後續製程的進行，而步進機台、晶圓本身以及製程環境都會造成層間光罩圖案疊合時產生位移和誤差，這稱為疊對誤差。隨著關鍵尺寸不斷縮小、晶圓尺寸變大以及光罩數目的增加，微影疊對誤差容忍度也越來越嚴苛，當微影製程的疊對誤差超過誤差容忍度則層間設計電路可能因為位移發生斷路或是短路而無法通過電性測試而報廢，進而影響產品良率，故疊對誤差控制是控制生產良率的重要指標，故準確的濾除不確定及隨機因素並估測出精確的最佳化可調整參數來穩定地降低疊對誤差是本文的目的。
本文利用類神經網路對未知非線性函數的逼近能力，以適當神經元模擬環境或晶圓本身所產生的隨機誤差，並將之濾除，再配合渴望函數與智慧型基因演算法求取最佳的可調整參數，以求出不受隨機誤差影響的最佳化參數解。
徑向基底網路快速學習的特點，可以提高估測的時效性，而智慧型基因演算法的最佳解可以使疊對誤差最小化，由兩者合併的結果顯示，本方法可以濾除隨機誤差的影響，並估測出最佳化的誤差可調整參數以提供工程人員做為補償疊對誤差時的重要參考。

關鍵字

最小平方法 ； 類神經網路 ； 智慧型基因演算法 ； 疊對 ； 最陡坡降法 ； 代價函數

並列摘要

Photolithography is the most critical step in semi-conductor industrial that influence the limit of critical dimension. Overlay errors are caused by mismatching between reticle and wafer or between lens and reticle. There are various kinds of overlay errors including trapezoid, rotation, expansion, magnification, translation error and so on. Summation of those items is called total error. With decreasing of critical dimension, increasing of reticles and increasing of wafer size, tolerance of overlay error is getting more and more critical. When errors overpass the tolerance of overlay error, it’ll cause short-circuit or broken circuit, and the products may fail in ADI ( After Develop Inspection ). In order to precisely control the overlay errors steadily, filtering the random noise out and describing a precise mathematical model is necessary.
Neural network has strong powers in approximation of unknown non-linear function. Simulating random errors caused by wafer or environment with appropriate neuron , filtering it out and then combining with Desirability Function and intelligent genes algorithm(IGA) to compute the optimal tunable parameter of stepper.
Learning quickly is one of the features of radial basis function network (RBFN). Moreover it filters the random noise and uncertainly factor out and simplifies the overlay error model. Intelligent genes algorithm is applied to computing the optimal tunable parameters. Combining RBFN and IGA and Desirability function can compute optimal parameters as reference for engineers when tuning. Finally the optimal tunable parameters are available.

並列關鍵字

least square ； Neural Network ； overlay ； Gradient Descent Method ； Cost Function ； intelligent genes algorithm

參考文獻

35. 趙育漢，”以類神經網路分析微影覆蓋幾何誤差”，中原大學，2004

14. Bode, Christopher A. ; Toprac, Anthony J. ; Edwards, Richard D. ; Edgar, Thomas F, “ Lithography overlay controller formulation “, Proceedings of SPIE - The International So- ciety for Optical Engineering, Vol.4182, p 2-11, 2000.

1. D. S. Perioff, “A Four-Point Electrical Measurement Technique for Characterizing Mask Superposition Errors on Semiconductor Wafer”, IEEE Journal of Solid-State Circuits, Vol.SC-13, No.4, pp.436-444 , August, 1978.

2. D. MacMillen, and W. D. Ryden, “Analysis of image field placement deviations of a 5X Micro-lithography reduction lens”, Proceedings of SPIE, Optical Micro-lithography-Technology, pp.78-89 1982.

3. C. K. Peski, “Minimizing Patter Registration Error Through Wafer Stepper Matching Techniques”, Solid State Technology, pp.111-115, May 1982.