以元學習增強先進製程控制中預測模型的領域適應力

Abstract

在現今蓬勃發展的半導體產業中，各大製造企業都已具備奈米級別的製程能力，並仍不斷將製造尺度向下突破，持續為品質管理帶來挑戰，因應這些挑戰，製造商在基本的先進製程控制（Advanced Process Control, APC）框架下，透過當代領先的機器學習技術（如深度學習）發展出預測型延伸功能模塊（如預測性維護、虛擬量測、良率預測等），來增強先進製程控制框架的控制力，然而，為了確保這些預測模型的精準度，常限縮這些模型的建置領域與適用範圍，帶來大量的模型需求，並提高了預測性模型建置與維護的成本。 另一方面，元學習（meta-learning）又稱為「學習如何學習」的研究領域，近年以小樣本元學習（few-shot meta-learning）的研究分支引領起一波研究動能復興，該分支著重透過元學習理念達到小樣本學習的目標，非常符合上述問題情境，卻未有相關的應用研究。因此，本研究旨在透過小樣本元學習的方法增強先進製程控制框架中深度學習預測模型的領域適應能力，期望強化整體控制系統的敏捷度與反應力，進而向更高程度的自動化智慧製造邁進。 本研究以先進製程控制框架中衍生的虛擬量測（Virtual Metrology, VM）模塊與小樣本元學習中的模型不可知元學習（Model-Agnostic Meta-Learning, MAML）方法為例，進行小樣本學習的實驗，測試以MAML降低虛擬量測模型資料需求的成效。具體來說，我們以RNN（傳統直觀）、CNN（當代熱門）與注意力機制（具發展潛力）三種神經網路結構為基底，分別選擇長短期記憶（LSTM）、時間卷積網路（TCN）與Transformer編碼器三者建構三種虛擬量測模型，並使用PHM Data Challenge 2016資料集進行配置調整、MAML訓練與小樣本下的新領域適應實驗，最後以預訓練（pretrain）模型遷移法為對照組做比較。 虛擬量測模型的建置結果，在三種基底的虛擬量測模型在基本的結構設計下，皆能勝過物理原理模型、統計特徵基底的模型與決策樹的集成模型，我們驗證了注意力機制與Transformer編碼器對於虛擬量測模型的發展潛力，並意外的發現該模型原始的位置編碼設計在本研究的個案中影響甚小，對此，我們提供虛擬量測不需要位置資訊與模型能從時序資料學習位置資訊兩種推論；而少樣本實驗的結果，所有MAML增強下的少樣本訓練表現些皆勝過預訓練模型遷移的對照組，且兩組間的差異皆呈現統計顯著性。 根據本研究的個案實驗結果，我們持續相信並建議小樣本元學習作為增強先進製程控制框架中預測模型的適應能力是適用且具發展性的，並且，以注意力機制為基底的虛擬量測模型尚有改善空間，也有達到當代最佳表現的潛能，未來透過可解釋性人工智慧技術（Explainable AI, XAI），有望更近一步放大注意力機制的價值，此外，位置編碼對Transformer編碼器在虛擬量測的影響也是個有趣的深入探討議題。

In the semiconductor manufacturing industry, most competitive companies now process at nano-scales, while still race to break through their bottom limits of process scales. As the ongoing competition continues to bring challenge to quality management, it encouraged scholars to incorporate state-of-the-art predictive information technologies (e.g., deep learning) into the Advanced Process Control (APC) framework. However, in practical situations, the training scope of these predictive models are often scaled down in exchange for better accuracy, resulting in more need for model units. Resources expended on these predictive models while setting up and maintenance have become a new type of cost for semiconductor manufacturing. On the other hand, meta-learning, also known as 'learn to learn', has recently been reviving through the branch of few-shot meta-learning, which focuses on designing meta-learning methods to achieve the capability of learning from very few examples. Given the fact that we found no related work of applying few-shot meta-learning methods to predictive models in the APC framework, this study aims to enhance the domain adaptability of deep learning based predictive models in the APC framework through few-shot meta-learning approach. By doing so, we hope to enhance the agility and responsiveness of the whole control system, in order to move towards higher levels of automated intelligent manufacturing. In our work, we conducted few-shot training experiments to observe the effectiveness of enhancing the adaptability of Virtual Metrology (VM) models through Model-Agnostic Meta-Learning (MAML) as a case study. We targeted three types of neural network bases, namely RNN, CNN, and attention mechanism, which each represents the “traditionally intuitive”, “recently popular”, and “potentially promising” choices of building VM models. From the three bases, we chose Long Short-term Memory (LSTM), Temporal Convolutional Network (TCN) and Transformer encoder, and constructed three different VM model structures. We conducted few-shot training experiments on the PHM Data Challenge 2016 dataset and compare the results under the initialization provided by MAML in contrast to the pretrain method. As a result, all three of our models outperformed physical models, statistical feature-based models, and decision tree ensemble models, showing qualification for our further experiments. We verified that attention mechanisms and the Transformer encoder are promising to approach Virtual Metrology, and surprisingly found that positional encoding had little effect on VM performance in our case. For this phenomenon, we provide two of our best conjectures, namely, attention-based model does not require position information in VM, or that models can learn position information form VM data. In our experimental results, MAML outperforms the pre-trained method as VM adaptability enhancement, while having statistically significant differences between the two methods. From our case study results, we continue to believe and suggest that few-shot meta-learning is promising for enhancing the adaptability of predictive models in the APC framework. In addition, we believe that attention-based Virtual Metrology have potential to reach state-of-the-art performance. The impact of positional encoding on Transformer encoders in VM would also be an interesting topic for further studies.