新穎的微影技術製程熱點偵測平臺

Abstract

在先進製程下，不斷增強的次光波長效應會導致晶片上不必要的圖形扭曲。雖然設計規則檢查（DRC）和光罩解析度增強技術（RET），諸如光學鄰近校正（OPC）和次解析輔助特徵（SRAF），可以緩解圖樣是否可成功微影的問題，但晶片上許多區域仍易受到微影製程所影響。這些有問題的區域，即所謂的製程熱點，必須在光罩製作前被檢測出來並且加以修正以保證晶片功能的正確性。因此，製程熱點偵測在實體驗證上是一不可或缺的任務。

製程熱點偵測，又稱為圖樣比對，可分為三大類：1）精確圖樣比對是用來萃取晶片佈局中與給定的熱點圖樣完全相同的區域。2）範圍圖樣比對則著重於尋找和給定圖樣有類似拓撲結構的熱點，藉以發掘更多的熱點區域。3）模糊圖樣比對則根據已知的熱點圖樣來預測從未見過的熱點。有時找到的熱點在拓撲結構上和給定熱點可能有巨大的不同。

為了準確且有效地解決這三類圖樣比對問題，需要好好處理以下兩個關鍵要素：1）一個好的熱點圖樣表示法和2）熱點關鍵特徵的選擇。一個好的圖樣表示法可以使熱點拓譜結構中關鍵特徵的抽取更準確便利。而當熱點的拓撲結構只使用真正關鍵的特徵來表示時，後續熱點偵測的流程才可大大地加快且不失精確性。

此論文中，我們在製程熱點的偵測上提出了一個完整的解決方法。1）針對精確圖樣比對，我們建構出以DRC 為基礎的新熱點偵測平臺，且提出修正的遞移閉包圖（MTCG）來簡潔地表示一個給定的熱點圖樣。現今其他以DRC 為基礎的偵測方法中，會直接提取所有的拓撲特徵來轉成設計規則。不同於他們，我們只提取關鍵拓撲結構來描述熱點圖樣。藉由我們的DRC 熱點偵測平臺，使用者可精確且有效地定位出所有熱點。2）針對範圍圖樣比對，當前最前瞻的方法係以字串比對為基礎，由於先天上的限制，此法主要著重在圖樣可變動邊緣的處理上，他們無法正確有效的處理不完整定義的熱點圖樣。在通過延伸MTCG 表示法和特徵提取規則後，我們的平臺不僅可以處理圖樣中可變動的邊緣也能完美處理不完整定義的熱點圖樣，進而萃取更多熱點。通過圖樣列舉，圖樣關鍵特徵提取，DRC 搜索空間縮減技術以及兩階段的過濾流程，我們可以在很短的運行時間內檢測所有的熱點，這是目前字串比對為基礎的方法做不到的。3）於模糊圖樣比對中，當前最先進的方法統合了圖樣比對和機器學習引擎來彌補相互的不足，但他們忽略了機器學習所擁有的潛力，而我們平臺中提出嶄新的技術來充分發揮機器學習的優勢。訓練資料的平衡和資料雜訊去除是達到高命中/誤報比的關鍵。通過整合拓撲分類，關鍵特徵提取和反饋訓練，我們的多核心熱點偵測平臺可達到非常高的精度，極低的誤報數，並縮短當前機器學習熱點偵測方法的運行時間。實驗結果應證了我們製程熱點偵測平臺的有效性和效率。

In advanced process technology, the ever-growing sub-wavelength lithography gap causes unwanted shape distortions of the printed layout patterns. Although design rule checking (DRC) and reticle/resolution enhancement techniques (RET), such as optical proximity correction (OPC) and subresolution assist features (SRAF), can alleviate the printability problem, many regions on a layout may still be susceptible to lithography process. These problematic regions, so-called lithography hotspots, have to be detected and corrected before mask synthesis to guarantee the functional correctness of the printed layout. Hotspot detection, therefore, is an essential task in physical veri?fication.

Lithographic hotspot detection, as known as pattern matching, can be classifi?ed into three categories: 1) Exact pattern matching, which extracts only hotspots in a layout with exactly the same dimension as a given hotspot pattern. 2) Range pattern matching, which focuses on hotspots with identical or similar topologies to a given pattern to identify more hotspot locations. 3) Fuzzy pattern matching, which seeks the unseen hotspot patterns based on known hotspots. The topologies between detected hotspots and given hotspots may have huge di?fferences.

To accurately and e?fficiently solve these three types of pattern matching, two key issues need to be well handled: 1) A good hotspot pattern representation and 2) critical feature selection. A good pattern representation can facilitate the critical feature extraction. By using only critical features to express pattern topologies, one can greatly accelerate the subsequent hotspot detection process without loss of accuracy.

In this dissertation, we propose a comprehensive solution for hotspot detection. 1) For exact pattern matching, we propose a good pattern representation called modifi?ed transitive closure graph (MTCG) to succinctly represent given hotspot patterns in our DRC-based hotspot detection framework. Unlike existing DRC-based works, which extract all topological features as design rules, we extract only critical design rules to express the topological features of hotspot patterns. By adopting a two-stage ?filtering process, we can locate all hotspots accurately and e?fficiently. 2) For range pattern matching, unlike the state-of-the-art string-matching based method, which mainly focuses on edge tolerances, we extend MTCG and the extraction rules to handle not only edge tolerances but also incomplete speci?fications to capture more hotspots. By combining pattern enumeration, critical feature extraction, DRC searching space reduction techniques and two-stage fi?ltering, we can detect all hotspots in a short running time. 3) For fuzzy pattern matching, unlike current state-of-the-art works, which unite pattern matching and machine learning engines, we fully exploit the strengths of machine learning using novel techniques. Population balancing and noise removal are the keys to reach high hit/extra ratio. By combing topological classi?fication, critical feature extraction and feedback kernels, our multiple kernels hotspot detection framework achieves very high accuracy, low false alarm, and shorter running time. Experimental results show the e?ffectiveness and effi?ciency of our proposed lithographic hotspot detection framework.