用於提升微機電側邊量測精準度之原子力顯微鏡新式掃描方法

Abstract

原子力顯微鏡是一種利用探針描繪樣本三維表面輪廓的高精密度掃描儀器，其精密度可達到奈米等級。原子力顯微鏡在許多領域中皆被廣泛應用，例如奈米科技、半導體製程、微機電、生物醫學等。在掃描樣本的過程中，探針需要保持固定的振幅以得到準確的測量結果。隨著製程技術的進步，垂直的側壁在許多半導體與奈米元件中趨於常見，而製造上的誤差及瑕疵往往在側壁附近出現，因此做為高精準度檢測儀器的原子力顯微鏡需要能夠在這些區域執行檢測，確保樣本符合規格。 在樣本側邊掃描所遭遇的問題包含一般使用的回授控制器皆會因為劇烈的高度落差產生延遲，而在微機電樣本中側邊往往高度變化大且角度接近垂直，這增加了一般回授控制器的追蹤難度。因此在沒有其他硬體設備改良如可旋轉式掃描平台及特別構造探針的協助下，獲得的側邊掃描結果不僅無法反映實際樣本狀況，甚至也可能讓探針在掃描過程中毀損或斷裂。除此之外，系統的不確定性例如壓電平台遲滯現象以及熱飄移與外在的雜訊影響皆可能降低側邊掃描準確度。 本論文設計一針對樣本陡峭側壁的檢測流程。首先，我們提出了能依據樣本樣貌逐條調整前饋誤差的疊代學習控制器，藉由線上訓練類神經網路預判回授誤差值平移量來解決疊代學習控制器無法追蹤非週期性訊號的限制。再經過疊代學習控制器掃描後，接下來為在搜尋區域內找尋側邊位置。搜尋過程中，我們透過蟻群演算法規劃搜索點的順序節省移動時間，同時利用疊代學習控制器提供的資訊來加速各搜索點垂直探索，最後以支撐向量機進行離散搜尋點分類估測側邊位置。搜尋完成後執行結合式掃描，在鄰近側邊區域讓探針定點收斂到正確位置，其他位置則維持原本的掃描速度。透過上述檢測流程，我們可以讓原子力顯微鏡有效率的在陡峭側邊提升追蹤表現及表面測量準確度。

Atomic force microscopy (AFM) is a high precision scanning probe measurement instrument that has the ability to obtain 3-D sample topography under nanoscale resolution. AFM is widely utilized in various fields, such as nano-technologies, semiconductor industry, Micro-Electro-Mechanical System (MEMS), bioscience, etc. Through scanning operation, oscillation amplitude of the scanning probe is required to maintain a constant set-point value to provide accurate scanning results. With the advance of manufacturing technologies, it gradually becomes more common that many semiconductors and nano-components are manufactured with perpendicular sidewalls. However, sidewall regions are prone to defects and process errors, which gives rise to the need of AFM to perform high precision topography measurement at these regions to ensure consistency between the actual sample product and its original design. There are several problems AFM would encounter during the process of sidewall inspection. Rapid height change would cause delay in conventional feedback schemes, making it incapable of tracking desired set-point amplitude. Therefore, in absence of hardware modifications such as additional rotating platform and customized scanning tips, not only would the obtained scanning results fail to provide accurate topographical information but also, the scanning probe might be under the risk of permanent fracture. The fact that MEMS sample often contains sidewall that has large height and nearly perpendicular angle that aggravates the difficulty. Moreover, system uncertainties including hysteresis in the piezoelectric stage, thermal drift, and environmental noise would also degrade sidewall scanning performance. In this thesis, we propose a scanning procedure aimed to enhance the accuracy at steep sidewall regions. First, an iterative learning controller (ILC) that could adjust feedforward data according to topography change along each scanning line is proposed. Through integrating an on-line trained edge prediction model based on a neural network (NN), we are able to obtain the amount of feedback error shift in advance, which in turn lifts the ILC limitation of periodic input signals. The next step is to define searching region according to the ILC scanning results and search for the exact sidewall location. To boost up efficiency of the searching process, we arrange the order of locations for probing by applying ant colony optimization (ACO), speed up the probing process at each location by information provided by ILC, and obtain a boundary that represents the estimated sidewall by supported vector machine (SVM) classification. With the exact sidewall location being made available, a hybrid scanning strategy is proposed to halt the scanning probe at regions adjacent to the sidewall to provide smooth and steady convergence process. Finally, to verify the proposed procedure of sidewall accuracy enhancement procedure does improve tracking performance and topographical scanning quality at sidewall regions, both extensive simulations and experiments have been conducted.