Multivariate statistics applications in phase analysis of STEM-EDS spectrum images Chad M. Parish n,1 , Luke N. Brewer Sandia National Laboratories, Albuquerque, NM 87185, USA article info Article history: Received 26 February 2009 Received in revised form 28 September 2009 Accepted 13 October 2009 Keywords: STEM X-ray microanalysis Spectrum imaging Spectral imaging Quantification Multivariate statistical analysis Principal component analysis PCA abstract Spectrum imaging (SI) methods are displacing traditional spot analyses as the predominant paradigm for spectroscopic analysis with electron beam instrumentation. The multivariate nature of SI provides clear advantages for qualitative analysis of multiphase specimens relative to traditional gray-scale images acquired with non-spectroscopic signals, where different phases with similar average atomic number may exhibit the same intensity. However, with the improvement in qualitative analysis with the SI paradigm has come a decline in the quantitative analysis of the phases thus identified, since the spectra from individual pixels typically have insufficient counting statistics for proper quantification. The present paper outlines a methodology for quantitative analysis within the spectral imaging paradigm, which is illustrated through X-ray energy-dispersive spectroscopy (EDS) of a multiphase (Pb,La)(Zr,Ti)O 3 ceramic in scanning transmission electron microscopy (STEM). Statistical analysis of STEM-EDS SI is shown to identify the number of distinct phases in the analyzed specimen and to provide better segmentation than the STEM high-angle annular dark-field (HAADF) signal. Representa- tive spectra for the identified phases are extracted from the segmented images with and without exclusion of pixels that exhibit spectral contributions from multiple phases, and subsequently quantified using Cliff–Lorimer sensitivity factors. The phase compositions extracted with the method while excluding pixels from multiple phases are found to be in good agreement with those extracted from user-selected regions of interest, while providing improved confidence intervals. Without exclusion of multiphase pixels, the extracted composition is found to be in poor statistical agreement with the other results because of systematic errors arising from the cross-phase spectral contamination. The proposed method allows quantification to be performed in the presence of discontinuous phase distributions and overlapping phases, challenges that are typical of many nanoscale analyses performed by STEM-EDS. & 2009 Elsevier B.V. All rights reserved. 1. Introduction A spectrum image (SI, sometimes called a hyperspectral image), is a dataset produced by the acquisition and storage of full spectra at many spatially distinct points in a sample [1–3]. Typical materials science applications of spectrum imaging include secondary ion mass spectroscopy (SIMS) [4], proton- induced X-ray emission (PIXE) [5], and X-ray photoelectron spectroscopy (XPS) [6]. In electron microscopy, techniques commonly performed in a spectrum imaging mode include energy dispersive spectroscopy (EDS) [7,8], electron backscatter diffrac- tion (EBSD) [9], cathodoluminescence (CL) [10], and electron energy loss spectroscopy (EELS) [2,3,11–13]. An important and growing application of scanning transmis- sion electron microscopy (STEM)-EDS spectrum imaging is in the quantification of atomic species present at each pixel. Hunneyball et al. [14] produced the first quantitative STEM-EDS maps. Later authors [7,15–19] were able to push quantitative mapping to near- one-nanometer spatial resolution and very high chemical sensi- tivity. As shown by Watanabe et al. [20], using principal component analysis (PCA) [a type of multivariate statistical analysis (MVSA) ] as a noise-filter allows very minor ( o1 wt%) components to be examined at high spatial resolution ( o5 nm), especially when combined with the very bright probe of an aberration-corrected cold-field-emission STEM. Following that work, we have recently used PCA noise-filtering to produce quantitative maps of cation fractions in (Pb,La)(Zr,Ti)O 3 (PLZT) thin-film materials [21]. Traditionally, quantitative STEM-EDS studies were performed by analyzing individual points, or a few points in a linescan [22], rather than quantifying pixel-by-pixel in a map or SI. These results ARTICLE IN PRESS Contents lists available at ScienceDirect journal homepage: www.elsevier.com/locate/ultramic Ultramicroscopy 0304-3991/$ - see front matter & 2009 Elsevier B.V. All rights reserved. doi:10.1016/j.ultramic.2009.10.011 n Corresponding author. E-mail address: parishcm@ornl.gov (C.M. Parish). 1 Present address: Oak Ridge National Laboratory, Oak Ridge, TN 37831, USA. Ultramicroscopy 110 (2010) 134–143