| 摘要: | 線鋸切割是在晶圓製造上發展成熟的製程技術，由於該項技術能一次性生產眾多直徑大而且厚度薄的晶圓片，已被廣泛應用在半導體界中。然而，由於業界對晶圓成品的表面品質要求日趨嚴格，晶圓線鋸切割技術正在面臨嚴峻的挑戰。因此，線鋸切割製程有必要建立品質預測系統，並藉此跟上不斷攀升的品質標準。對於晶圓製造廠而言，由於晶圓線鋸在材料以及加工時間上都需要耗費鉅額成本，這樣品質預測系統的建立將能夠帶來大量的助益。
本研究採用了二種人工智慧(AI)模型，並利用來自合作廠商所提供的切割資料進行品質預測系統的建立，這套系統利用切割製程中所擷取的感測器訊號資料來預測晶圓成品的表面品質。首先，本研究採用隨機森林模型進行製程各感測器訊號的重要性分析，藉此篩選出對於切割製程具有高度代表性的感測器訊號。接著使用訊號處理與特徵擷取方法以提取訊號的重要統計特徵，並且將該特徵作為後續兩項AI品質預測模型的輸入。這兩項預測模型分別是k近鄰演算法(KNN)搭配粒子群最佳化演算法(PSO)，以及極限梯度提升演算法(XGBoost)。本研究所提出的這兩項AI模型利用與品質有高度關聯的溫度訊號特徵來預測晶圓線鋸切割成品的四種品質指標，分別是總厚度變異(TTV)、翹曲(warp)、彎曲(bow)以及波紋指數(waviness)。將合作廠商所提供的晶圓線鋸製程資料拆分成兩個數據集，分別用於AI模型的交叉訓練驗證及測試上。模型的交叉驗證與測試結果顯示，本研究所提出的AI模型，不論數據集來自單一取樣率或綜合取樣率的感測器訊號，皆能夠有效地預測四種品質指標。依據透過綜合取樣率的數據集所建立之模型，其預測品質指標的測試表現在TTV上之平均百分比誤差(MAPE)及決定係數(R2)分別達到3.71%與0.85，在warp上分別達到9.23%與0.8，在waviness上則分別達到12.25%與0.89，而對bow測試的R2能夠達到0.79。
;Saw-slicing is a mature manufacturing process for the fabrication of wafers. Due to its ability to produce multiple wafers of a large diameter and thin thickness in a single process, this technique is prevalent in the semiconductor industry. However, because of increasingly stricter demands for the surface quality of wafers, the wafer saw-slicing process is currently facing challenges to keep up with the higher quality standards. Therefore, applying a quality predicting system in the saw-slicing process is necessary. Such action is very beneficial for the wafer manufacturer, as the material and time costs of each slicing process are very high.
In this study, two artificial intelligence (AI) based models and real data from the wafer saw-slicing process were employed to establish the quality predicting system. The system performed predictions for the surface quality of the wafer based on the signals captured by the sensors during the process. The signal data were first selected by an important variable analysis using the Random Forest model. After that, signal processing and feature extraction methods were applied to the chosen signal, and the resultant statistical features served as inputs for the two quality-predicting AI models. The predictive models adopted were k-nearest neighbors (KNN) combined with particle swarm optimization (PSO) and XGBoost models. These AI models exploited the extracted signal features and predicted the four quality indices of sliced wafers, namely total thickness variation (TTV), warp, bow, and waviness. The collected datasets were divided into two parts for the training/cross validation and testing in the AI models. According to the results of the cross validation and testing, the proposed models could effectively predict the four quality indices, whether the dataset was from sensor signals of a single sampling rate or mixed sampling rates. Based on the performance of AI models trained with the dataset from mixed sampling rates, the testing scores in mean absolute percentage error (MAPE) and coefficient of determination (R2) respectively reached 3.71% and 0.85 for predicting TTV, and achieved 9.23% and 0.8 for predicting warp. For AI models that predicted waviness, the MAPE and R2 scores in testing reached 12.25% and 0.89, respectively. As for bow, the testing results of AI models showed an R2 score of 0.79. |
