基於機器學習之kMLLS聚類演算法及其在薄膜電子損失譜之應用

The Introduction to kMLLS Clustering and It's Applications to EELS Analysis of Thin Films

近年來，球面像差校正穿透式電子顯微鏡影像搭配能譜分析技術，已逐漸成為材料分析中不可或缺的工具。從能譜影像中，藉由適當的數據處理，我們除了原子特徵之外，更可以得到化學元素及電子結構等材料訊息。在本文中，我們介紹一種新穎的演算法──kMLLS聚類，其結合了k平均聚類與複線性最小平方擬合的優點，使我們可以正確地萃取出終端材料能譜，並得到其分佈情形。利用kMLLS聚類，未來可應用在線上檢測，使研究人員即時得到更深入的材料訊息。

In recent years, combining a Cs-corrected scanning transmission electron microscope with an EDS and/or EELS detector has become an indispensable tool for material characterization. With a proper data processing, atomic structures, chemical compositions, and electronic configurations of materials can be resolved in a spectrum image. In this article, we introduce a novel algorithm - kMLLS clustering, which combines the advantages of k-means clustering and multiple linear least squares fitting, to accurately extract the spectra of the endmembers and the corresponding distribution from a spectrum image. kMLLS clustering has the great potential to the in-line application and provides significant insights into materials.