應用資料探勘技術於晶圓測試回測率預測之研究 – 以K公司為例

Abstract

半導體晶圓測試中有許多因素會影響到晶圓測試良率的好壞，以晶圓測試廠來說時間就是金錢，如何能在最短時間內讓晶圓能最快速的完成測試，便能提高公司產能最大化以增加公司營收及提高公司競爭力。 本研究主要於建立一回測率預測模型用於預測每批晶圓投入量產後所需時間，以便產線排程人員能夠做機台最佳化之排程，將測試機台時間有效利用以增進機台產能並提高公司競爭力。研究中採用三種數值預測演算法來做比較，分別為線性迴歸分析、隨機森林迴歸分析及GBDT迴歸分析，並調教相關參數來達成最佳預測。實驗中發現若先做分群後進行預測可得到更佳的預測結果，因此先使用K-Means分群演算法分群後再進行預測。實驗結果顯示隨機森林及GBDT迴歸分析的預測結果相當接近，但是GBDT迴歸分析的最大誤差值比較小，因此三種數值預測演算法中GBDT迴歸演算法的效果優於隨機森林迴歸與線性迴歸。

There are many factors in the semiconductor wafer test that will affect the wafer test yield result. For the wafer test factory, time is money; thus, complete the chip test in shortest time can increase the company's capacity to maximize the revenue and improve the company's competitiveness. This study mainly focused on building a retest rate prediction model to predict the time required for each batch of wafer input; so that the production line scheduling staff can optimize the schedule of wafer test and effectively use the machine to enhance the capacity of the machine and improve the competitiveness of the company. In the study, three prediction methods, including the linear regression, random forest and GBDT regression analysis, are adopted to build the prediction model by coordinating the relevant parameters to achieve the best prediction result. The experimental result shows that the clustering-based prediction models can achieve better prediction results. Thus, the K-Means clustering algorithm is performed before building the prediction models. The Experimental results of random forest and GBDT regression analysis are quite close, but the maximum error value of GBDT regression analysis is relatively small. The GBDT regression method performs better than the random forest and linear regression methods.