| 摘要: | 線鋸切割已成為晶圓製造中一種成熟且廣泛採用的製程技術，特別是因為它能夠有效率地大量生產直徑大、厚度薄的晶圓。然而，隨著半導體產業的不斷發展，晶圓成品表面品質的要求日益提升，對這項成熟的技術提出了重大挑戰。為了滿足不斷變化的品質要求，開發能夠在整個切片過程中監測和預測晶圓表面狀況的預測系統已變得至關重要。此類系統不僅有助於製程優化，而且由於晶圓切片操作涉及大量的材料和加工費用，還能帶來顯著的成本節約效益。
為提升表面品質預測的準確性與可靠性，本研究設計了兩種機器學習架構。第一種為傳統的基礎模型，第二種則為堆疊模型。堆疊模型透過一個元模型整合多個基礎模型的預測結果，藉此提升整體預測效能。這兩種架構皆應用於由產業合作夥伴提供的實際晶圓切割資料，分別為A1與A2，並且針對四個關鍵表面品質指標上進行了評估，分別是總厚度變異(TTV)、翹曲(warp)、彎曲(bow)以及波紋指數(waviness)。本研究採用隨機森林模型分析製程中各感測器訊號的重要性，並從中篩選出對切割製程最具代表性的關鍵訊號。接著，透過訊號處理與特徵擷取技術，萃取出具代表性的統計特徵，並將這些特徵作為後續模型的輸入。模型的交叉驗證與測試結果顯示，本研究所提出的 AI 模型能夠有效預測四種晶圓切割製程中的關鍵品質指標。根據測試結果，堆疊模型在所有指標上的平均百分比誤差(MAPE)與決定係數(R²)表現均優於任一單一基礎模型。其中，TTV 的預測準確度最高，MAPE始終低於3%，R² 超過 0.9。Warp 與 waviness 的預測表現亦相當出色，MAPE 通常分別低於6% 和7.5%，R² 值接近0.9。Bow 的預測結果也維持穩定，R² 值皆高於0.85。
線性模型在本研究中扮演了關鍵的整合角色。儘管其作為單一預測模型時表現相對不佳，但在堆疊模型架構中擔任元模型時，卻展現出優異的整合效果。此現象顯示，線性模型雖不擅長獨立捕捉複雜的非線性關係，然而在彙總多個基礎模型的預測結果時，具備良好的泛化能力與穩定性。此外，模型在A1 與 A2 資料集對於各項品質指標上皆呈現高度一致的表現，進一步驗證本研究所建構模型架構之穩健性與實務應用潛力。
;Wire saw slicing has become a mature and widely adopted process in wafer manufacturing, particularly due to its efficiency in producing wafers with large diameters and thin thicknesses in high volumes. However, with the continuous advancement of the semiconductor industry, the demand for higher surface quality in wafers has become increasingly stringent, posing new challenges to this well-established technique. To reach these evolving quality requirements, it has become essential to develop predictive systems capable of monitoring and forecasting wafer surface conditions throughout the slicing process. Such systems not only support process optimization but also offer substantial cost-saving benefits, as wafer slicing operations involve considerable material and processing expenses.
To address the need for accurate surface quality prediction, this study develops two machine learning frameworks. One is a conventional stand-alone base model structure, and the other is a stacking ensemble model that integrates multiple base learners through a meta-model to enhance predictive performance. These models are evaluated using two real-world wafer slicing datasets, A1 and A2, which are provided by an industry partner. The prediction targets include four key surface quality indices, namely total thickness variation (TTV), warp, bow, and waviness. The study first applied a Random Forest model to analyze the importance of various process sensor signals and selected the most representative signals relevant to the slicing process. Signal processing and feature extraction techniques were then used to derive informative statistical features, which served as inputs to the subsequent prediction models.
Cross-validation and testing results show that the proposed AI models effectively predict all four critical quality indices in wafer slicing. The stacking model outperforms all individual base models across every metric, achieving superior performance in both MAPE and R². Among the indices, TTV demonstrated the highest prediction accuracy, with MAPE consistently below 3% and R² exceeding 0.9. Warp and waviness also performed well, with MAPE typically below 6% and 7.5%, respectively, and R² values approaching 0.9. Bow predictions remained stable, with R² values consistently above 0.85.
Linear models played a key role in the proposed ensemble architecture. Although their standalone prediction performance was relatively poor, they delivered excellent results when used as meta-models within the stacking framework. This finding indicates that while linear models may struggle to capture complex nonlinear patterns on their own, they are highly effective in aggregating and generalizing the outputs of diverse base learners. Furthermore, the consistent performance observed across both datasets A1 and A2 validate the robustness and practical applicability of the proposed prediction framework in real-world wafer slicing scenarios. |
