1. Charles Lyman Lehigh University, Bethlehem, PA
Spectrum Imaging
Charles Lyman Lehigh University, Bethlehem, PA
Based on presentations by John Hunt (Gatan, Inc.), John Titchmarsh (Oxford University), and Masashi Watanabe (Lehigh University)

2. Spectrum Imaging (SI) x E Collect entire spectrum at each pixel
Scan y
Incident
electron probe
“x-y-energy” data cube
Collect entire spectrum at each pixel
No a priori of specimen knowledge required
Can detect small amounts of elements in local regions of x-y images
Away from microscope:
Repeatedly apply sophisticated spectrum processing
“Mine the data cube” for features
Concept
Jeanguillaume & Colliex, Ultramicroscopy 28 (1989), 252
Demonstration
Hunt & Williams, Ultramicroscopy 38 (1991), 47

3. Elemental Maps from Data Cube
X-ray map x y
Energy
200
400
600
800
1000
1200
1400
1600
1800
2000
Energy x y
X-ray Spectrum
Specimen: polished granite
Data courtesy of David Rohde

4. Quantitative Phase Analysis
Sum spectra for pixels within box
Enough counts for quatitative analysis
Specimen: polished granite
Data courtesy of David Rohde

5. Compositional Maps in TEM/STEM
Collection by:
STEM X-ray
Sequentially acquire EDS x-ray spectrum at each pixel (original concept)
Each x-ray entering detector assigned “x-y-energy” tag (Mott & Friel, 1999)
STEM EELS
Sequentially acquire EELS spectrum at each pixel
EFTEM (Energy-filtered imaging)
Sequentially acquire images at specific energies
One energy window for each energy channel in spectrum (DE)

6. A few Words about EFTEM Elemental Maps without Employing Spectrum Imaging

7. EFTEM: In-Column and Post-Column Energy Filters
Omega Filter
Gatan Imaging Filter (GIF)
From Williams and Carter, Transmission Electron Microscopy, Springer, 1996

8. Energy-Filtered TEM (EFTEM) Element Maps - Not Spectrum Images
Elemental Maps of a SiC/Si3N4 ceramic Short Acquisition Time (3 maps, 250K pixels) = 50s
Carbon
Nitrogen
Oxygen
RGB composite
Courtesy John Hunt, Gatan 3

9. Energy-Filtering TEM Images of only a small range of energies
Energy window of 1-100eV
Just above or just below energy-loss edge
EFTEM compositional mapping
Elemental maps using multiple energy-filtered images
2 images to determine background before edge
Scale background and subtract to obtain elemental signal
1 image to collect elemental signal (edge above background)
Only one electron energy can be precisely in focus
All other energies will be suffer resolution loss (blurring)
The blurr is given by:
d = Cc *b *DE/E
Cc = chromatic aberration constant
b = the acceptance angle of the objective aperture
DE = range of energies contributing to the image
Blurr will be especially large for thick, high-Z specimens.
Reduce blurr by:
Using a small energy window (DE)
Select energy loss DE by changing the gun voltage (vary kV)
This is an introductory slide. More in-depth discussion is given further on in this presentation.

10. EFTEM Elemental Mapping
Three-Window Method
Subtract edge background using two pre-edge images (dotted line)
Element concentration proportional to area of edge above background (outlined in red)
Absolute concentration can be determined if thickness and elemental cross-sections are known
Courtesy John Hunt, Gatan

11. EFTEM Elemental Mapping: Example 1
Aluminum
Titanium
6 layer metallization test structure
3 images each around:
O K 532 eV
Ti L eV
Al K 1560 eV
1 µm
Oxygen
Superimpose three color layers to form RGB composite O Ti Al
Courtesy John Hunt, Gatan

12. EFTEM Elemental Mapping: Example 2
BF image N Ti O Al Si
Unfiltered bright-field TEM image of semiconductor device structure and elemental maps from ionization-edge signals of N-K, Ti-L, O-K, Al-K, and Si-K.
Color composite of all 5 elemental maps displayed on the left,showing the device construction.
Courtesy John Hunt, Gatan

13. EFTEM detection limits
Typically 2-5% local atomic concentration of most elements
1% is attainable for many elements in ideal samples
10% for difficult specimens that are thick or of rapidly varying thickness
Sensitivity limited by:
Diffraction contrast
Small number of background windows
Signal-to-noise
Thickness
Artifacts
If you can see the edge in the spectrum, you can probably map it
EFTEM spectrum image can map lower concentrations than the 3-window method
Better background fits because there are more fitting channels
Courtesy John Hunt, Gatan

14. STEM & EFTEM EELS Spectrum Imaging

15. STEM spectrum image acquisition
acquired by stepping a focused electron probe from one pixel to the next
The spectrum image data cube is filled one spectrum column at a time
In STEM it is possible to collect x-ray, EELS, BF, and ADF simultaneously
Use of the ADF or SE signal during acquisition permits spatial drift correction
EDX
STEM
EELS DF
Specimen x y E
Courtesy John Hunt, Gatan

16. EFTEM spectrum image acquisition
Acquire an image containing a narrow range of energies
The spectrum image data cube is filled one energy plane at a time
Image plane retains full spatial resolution of TEM image y x E
image at E1
image at E2
image at Ei .
Courtesy John Hunt, Gatan

17. STEM EELS spectrum imaging
EELS STEM SI acq. at 200keV (cold FEG)
xy: 50*29 pixels
E: 1024 channels (75eV, D=0.5eV)
Acquisition time: ~ 5 minutes
Processing time: ~ 5 minutes
Courtesy John Hunt, Gatan

18. Quantitative EFTEM Spectrum Imaging
EFTEM Spectrum Image
2.9 nm resolution
Si-L23 : eV {3eV steps} (1.5 min)
N-K, Ti-L, O-K : eV {5eV steps} (8 min)
FEI CM120 + BioFilter
120keV
Corrections: x-rays, MTF, spatial drift
Scaled by hydrogenic x-sections
Courtesy John Hunt, Gatan

19. STEM vs. EFTEM Spectrum Imaging
Quantitative elemental mapping
Both STEM SI and EFTEM SI can do this
EELS STEM Spectrum Imaging
Good quality spectra
All artifacts / instabilities correctable
Usually safer w/unknowns
EFTEM Spectrum Imaging
Fast mapping
Uncorrected artifacts / instabilities are very dangerous
Very useful for well characterized systems
Excellent spatial resolution

20. X-ray Spectrum Imaging

21. Masashi Watanabe Lehigh University
Mining the SI Data Cube
Multivariate Statistical Analysis of X-ray Spectrum Images
Nb(wt%)
Nb(wt%)
1.5
1.5
Masashi Watanabe
Lehigh University

22. X-ray Spectrum Imaging
Specimen: Ni-based superalloy
Collection of SI
Huge data set
e.g. 256x256 = 65,536 spectra
each spectrum 1024 channels
cannot analyze manually
Noisier spectrum
for XEDS than EELS
Many possible variables
composition, thickness, multiple phases
100 nm
NiKa
AlKa
CrKa
What can we do?
TiKa
FeKa
Courtesy M. Watanabe

23. Multivariate Statistical Analysis
Multivariate statistical analysis (MSA) is a group of processing techniques to:
identify specific features from large data sets (such as a series of XEDS and EELS spectra, i.e. spectrum images) and
reduce random noise components efficiently in a statistical manner.
Problems for which MSA may be useful
Investigation of data of great complexity
Handling large quantities of data
Simplifying data and reducing noise
Identifying specific features (components) can be interpreted
in useful ways
E.R. Malinowski, Factor Analysis in Chemistry, 3rd ed. (2002)

24. Nb map in Ni-base superalloy
original
MSA-processed
Nb(at%)
Nb(at%) 1 1
100 nm
Multivariate Statistical Analysis
identify specific features in the spectrum image
reduce random noise
Courtesy M. Watanabe

25. The Data Cloud Find greatest variancein data
x1, x2, x3 are first three channels of spectrum or image
Manipulate matrices
Principal component analysis finds new axes for data cloud that correspond to the largest changes in the data
These few components can represent data

26. Principal Component Analysis (PCA)
PCA is one of the basic MSA approaches and can
extract the smallest number of specific features
to describe the original data sets.
The key idea of PCA is to approximate the original
huge data matrix D by a product of two small
matrices T and PT by eigenanalysis or singular value
decomposition (SVD)
D = T * PT
D: original data matrix (nX x nY x nE)
T: score matrix (related to magnitude)
PT: loading matrix (related to spectra)
Courtesy M. Watanabe

27. Practical Operation of PCA
eigenanalysis
or SVD
original data
score
loading nE nE nE nX D T PT
line profile
PCA = nX * nY
nX x nY
nX x nY nE
eigenvalues
D: original data matrix (nX x nY x nE)
T: score matrix (related to magnitude)
PT: loading matrix (related to spectra)
D = T * PT
spectrum image
Courtesy M. Watanabe

28. Spectrum Image of Ni-Base Superalloy
matrix
NiKa
FeKa
CrKa g’
NiKa
NbLa
AlKa
TiKa
M23C6
CrKa
spectrum image:
256x256x1024
dwell time: 50 ms
20 eV/channel
100 nm
Courtesy M. Watanabe
Reconstructed spectra

29. Results of PCA 1 #1: average #2: M23C6 scree plot #3: g’ Score Loading
STEM-ADF
Score
Loading
#1: average
Cr Ka
Ni Ka
200 nm
#2: M23C6
scree plot
Cr Ka
Ni Ka
#3: g’
Fe Ka
Ni Ka
Noise
Cr Ka
Al Ka
Ti Ka
Courtesy M. Watanabe

30. Results of PCA 2 #4: absorption #5: noise scree plot #6: noise Score
STEM-ADF
Score
Loading
#4: absorption
Cr Ka
Ni Ka
Ni La
200 nm
#5: noise
scree plot
#6: noise
Noise
Courtesy M. Watanabe

31. Compositional fluctuations below 2 wt% can be revealed
Comparison of Maps Al Nb
wt%
wt% 2
1.5
Original
wt%
wt% 2
1.5
Reconstructed
100 nm
Compositional fluctuations below 2 wt% can be revealed
Courtesy M. Watanabe

32. Application to Fine Precipitates
Irradiation-induced hardening in low-alloy steel
is caused by fine-scale precipitation
Average precipitate size: 2-5 nm
X-ray mapping in VG HB keV STEM
BF-STEM image
ADF-STEM image
100 nm
Burke et al. J. Mater. Sci. (in press)

33. Application to Fine Precipitates in Steel
Burke et al. J. Mater. Sci. (in press)
STEM ADF
Thickness Fe Cr
50nm 10 20 85 95 1 5
(nm)
(wt%)
(wt%) Ni Mn Cu Mo 8 2 3
0.5 1
(wt%)
(wt%)
(wt%)
(wt%)
Too noisy

34. Application of MSA to Fine Precipitates
Burke et al. J. Mater. Sci. (in press)
STEM ADF
Thickness Fe Cr
50nm 10 20 85 95 1 5
(nm)
(wt%)
(wt%) Ni Mn Cu Mo 8
1.5 3
0.8 1
(wt%)
(wt%)
(wt%)
(wt%)

35. Some References to MSA Procedures
Multivariate statistical analysis – in general
S.J. Gould: “The Mismeasure of Man”, Norton, New York, NY, (1996).
E.R. Malinowski: “Factor Analysis in Chemistry, 3ed ed.”, Wiley, New York,
NY, (2002).
P. Geladi & H. Grahn: “Multivariate Image Analysis”, Wiley, West Sussex,
UK, (1996).
For microscopy applications
P. Trebbia & N. Bonnet: Ultramicroscopy 34 (1990) 165.
J.M. Titchmarsh & S. Dumbill: J. Microscopy 184 (1996) 195.
J.M. Titchmarsh: Ultramicroscopy 78 (1999) 241.
N. Bonnet, N. Brun & C. Colliex: Ultramicroscopy 77 (1999) 97.
P.G. Kotula, M.R. Keenan & J.R. Michael: M&M 9 (2003) 1.
M.G. Burke, M. Watanabe, D.B. Williams & J.M. Hyde: J. Mater. Sci. (in press).
M. Bosman, M. Watanabe, D.T.L. Alexander, and V.J. Keast: Ultramicroscopy
(in press)

36. Summary Spectrum Imaging Mining the data cube
the way serious microanalysis should be done
Mining the data cube
MSA is applicable for large data sets such as line
profiles and spectrum images
The large data sets can be described with a few
features by applying MSA
PCA is useful for noise reduction of data sets.
Be aware -- MSA can provide only hints of significant
features in the data sets (abstract components)