以主動式學習引導次級解析輔助特徵圖案擺置

Abstract

隨著半導體製程技術的演進，特徵尺寸已遠小於曝光光源波長，迫使成像明顯偏離原來的設計圖案。因此光學微影解析度增強技術在製造可行性領域中益發重要。其中，次級解析輔助特徵圖案的擺置能有效提高目標圖案的焦深及其製程視窗的品質。為了突破業界慣用的以模型為基礎的高計算成本以及以準則為基礎的過大的查找表，近期研究多專注於使用機器學習的預測來減少計算時間。然而，現有機器學習模型的表現高度仰賴足夠的訓練樣本，在先進製程中，雖有龐大的解空間，但僅有少量標記資料。雖然可以透過收集更多的樣本來解決，但決定樣本資訊性的訣竅仍鮮少被討論。在本篇論文中，我們的貢獻有三：一、提出新穎的基於變分自動編碼器的主動式學習框架來主動選擇具資訊性的樣本並減少所需的標記資料量。二、提出區域性同心圓取樣表示法避免資訊丟失。三、提出聚類法決定最終次級解析輔助特徵圖案的擺置。實驗結果顯示，我們提出的框架僅使用40%的訓練樣本即能達到較現有方法優異的製程變異帶寬與邊緣放置誤差。

As the feature size keeps shrinking in the modern semiconductor manufacturing process, resolution enhancement techniques (RETs) are crucial to improve the manufacturing yield. Sub-resolution assist feature (SRAF) insertion is one of the RETs that can improve the target pattern printability and lithographic process window. Model-based SRAF generation achieves a high accuracy but with a high computational cost, while rule-based SRAF insertion may require a huge look-up table to handle complex patterns. Thus, recent works focus on reducing runtime by using machine-learning based models, and they rely on sufficient training samples to generalize the trained models and achieve high performance. Nevertheless, in advanced lithography, we may have a huge solution space but may have few labeled training samples. Although we can simply gather more samples, it is difficult to determine the most informative samples. Therefore, in this thesis, our contributions are threefold: First, we propose an active learning framework based on Variational Auto Encoders (VAEs) to actively select informative samples used to trained our model to guide the SRAF insertion. Second, we propose a region-based concentric circle area sampling representation to avoid information loss. Third, we propose a clustering-based scheme to determine the final placement of SRAFs. Experimental results demonstrate that, compared with state-of-the-art works, our framework uses 40% training samples and improves process variation (PV) band and edge placement error (EPE).