用於微機電檢測之原子力顯微鏡探針定位及高效率掃描方法

Abstract

原子力顯微鏡是一種高精度的探針掃描儀器，能夠得到奈米等級的樣本三維表面輪廓。原子力顯微鏡能廣泛應用於不同領域，如奈米科技、半導體、微機電，生物科學等。為了得到大範圍樣本的表面輪廓，我們需要知道原子力顯微鏡的探針與樣本之間的相對位置，以將探針放在正確的地方掃描。以微機電檢測為例，經過半導體加工後得到的樣本，我們需要檢查樣本上的一些關鍵尺寸，以確保得到的樣本有符合當初設計的規格。 原子力顯微鏡的探針定位問題包含未知的初始位置和系統的不確定性。原子力顯微鏡的掃描範圍（例如幾十微米）遠小於樣本的尺寸（例如幾毫米），因此在沒有其他顯微鏡像是光學顯微鏡的輔助下，單單只靠原子力顯微鏡得到的影像來定位是非常困難的。原子力顯微鏡在微機電樣本上得到的掃描圖像通常是具有極少特徵的簡單幾何形狀，因此增加了定位的難度。除此之外，系統的不確定性包括壓電平台的磁滯現象、熱漂移和解析度較差之雙平台（例如長行程平台）都會影響到定位的精度。 在本論文中，參考宏觀機器人的定位方法，我們提出了一種使用粒子濾波器的原子力顯微鏡探針定位演算法。我們將原子力顯微鏡掃描得到的圖像當作唯一的傳感器，將樣本的設計圖當作地圖。粒子濾波器中的感知模型則是基於一個特徵提取演算法。定位完成後，使用一個快速掃描演算法來掃描大範圍樣本，該算法結合了即時變速掃描和基於機器學習的前饋控制。透過上述的這些方法，我們可以完成原子力顯微鏡在大範圍樣本上的自動定位和快速掃描。

Atomic force microscopy (AFM) is a powerful instrument that has the ability to characterize sample topography on nanoscale resolution. AFM is widely used in different fields, such as nanotechnology, semiconductor, MEMS, bioscience, etc. In the case of obtaining 3D topography of a large range sample, we need to know the relative position of the AFM probe to the sample so that the probe will be placed in the right place. Taking MEMS inspection for instance, we need to inspect some critical dimensions on the sample to ensure that the sample meets specifications after certain semiconductor process. The so-called AFM tip localization problem touches upon the issue of unknown initial position and system uncertainties. The scanning range of an AFM (e.g. tens of micrometers) generally is much smaller than the sample size (e.g. a few millimeters). Therefore, it is hard to localize the AFM tip position without other auxiliary microscopes, such as optical microscope. Moreover, the AFM scanned images on a MEMS sample typically involve only simple geometries with sparse features which usually leads to the difficulty of localization. Besides, the system uncertainties including piezoelectric scanner hysteresis, thermal drift, and coarse dual stage (e.g. long travel-range positioning stage) would affect positioning accuracy. In this thesis, we propose an AFM tip localization method using particle filter referring to macro robot SLAM. We take the AFM scanned image as the unique sensor and the sample layout as the map. The sensor model of the particle filter is based on a feature extraction algorithm. After localization, the large range sample is scanned using a novel fast scanning algorithm combining on-line variable speed scan and machine learning based feedforward control. To verify the effectiveness of the proposed methods, both simulations and experiments are conducted, and the tip localization as well as speed of sample area are highly promising.