Title: 應用粒子群神經網路構建積體電路之良率預測模式
Developing a Wafer Yield Prediction Model Using Particle Swarm Optimization Neural Networks
Authors: 彭御哲
Yu-Che Peng
唐麗英
Lee-Ing Tong
工業工程與管理學系
Keywords: 積體電路;缺陷點;群聚現象;群聚指標;粒子群神經網路;粒子群最佳化;良率預測模式;Integrated circuit;defect;cluster;cluster index;Particle Swarm Optimization Neural Network;Particle Swarm Optimization;yield model
Issue Date: 2007
Abstract: 晶圓上的良率(yield)是積體電路業者衡量其製程能力的重要指標。當晶圓上缺陷點(defects)的群聚(clustering)現象不明顯時,晶圓良率的高低與晶圓上的缺陷點數成反比,但是當晶圓面積越作越大時晶圓上缺陷點的群聚現象就越來越明顯,晶圓良率與缺陷點數之間的反向關係也就越不明顯,因此群聚現象會導致卜瓦松良率模式預測不準確。針對此問題,中外文獻提出了一些機率模式如負二項良率模式(Negative Binomial yield model) 、Murphy模式等,或用類神經網路(Artificial Neural Network)等方法來構建良率預測模式,但這些良率預測模式仍各有一些缺失。因此本研究的主要目的為應用粒子群神經網路(Particle Swarm Optimization Neural Networks, PSONN),並以缺陷點數、缺陷點群聚圖案種類、群聚指標 值為依據,建構出一個精確的良率預測模式。目前的類神經網路良率模式中大都以倒傳遞網路良率模式為主,訓練倒傳遞網路的方法是最陡坡降法(Steepest Decent Method),但最陡坡降法之解有著可能僅是局部最佳解及依賴網路初始權重值等缺點,往往在訓練過程當中無法收斂於最佳解處,而影響所建構預測模式的準確性。粒子群神經網路是採用多點搜索的方式,因此能快速的移動至最佳解處,可避免陷入局部最佳解,找出最佳的網路結構與連接權重值,進而提升網路模型預測的準確性。本研究最後以新竹科學園區某積體電路公司所提供之實際晶圓資料以及模擬資料來驗證本研究所發展之良率模式確實有效可行。
For integrated circuit (IC) manufacturers, the wafer yield is a key index to evaluate their process capability. As the wafer size increase, the defects clustering phenomenon tends to be significant, and the phenomenon leads to the inverse relationship between the wafer yield and the defects counts. In this case the conventional Poisson yield model will frequently underestimate the actual wafer yield. Although many modified yield models or yield models constructed by Neural Networks have been developed to overcome the clustering problem. However, these models still have some shortcomings. Therefore, the objective of this study is to develop an IC yield model using Particle Swarm Optimization Neural Networks (PSONN). The cluster index , the cluster patterns and the defect counts are considered as the input variables for the proposed model. Back-Propagation Neural Network (BPN) is often applied in Neural Networks to construct the yield model, and the commonly used training method is BP algorithm. Because BP algorithm is a kind of the Steepest Decent Method, it has some disadvantages. That is, it can obtain only the local optimum solution and it is sensitive to initial connecting weights. Consequently, this study employs Particle Swarm Optimization (PSO) in Network training to determine the connecting weights between layers of Neural Networks. PSO is a population based optimization method, it can avoiding a local optimum solution and has higher prediction accuracy than BP. A simulated wafer data and a real case from a Taiwan IC Company are applied to demonstrate the effectiveness of proposed model.
URI: http://140.113.39.130/cdrfb3/record/nctu/#GT009533528
http://hdl.handle.net/11536/39159
Appears in Collections:Thesis