Professional Documents
Culture Documents
SCANNING TRANSMISSION
ELECTRON MICROSCOPY
School of Physics
The University of Melbourne
Australia
August 3, 2005
2
Abstract
This thesis explores the theory describing wavefunctions and images, both elastic
and inelastic, formed in scanning transmission electron microscopy.
A method is presented for calculating the elastic wavefunction based upon a new
formulation of the boundary conditions which couples the probe to Bloch states
within the crystal in a single step. Though this method is fundamentally equivalent
to previous approaches based upon the superposition of wavefunctions corresponding
to individual plane wave components in the incident probe, it provides new insight
into the some of the dynamics, allows for efficient calculations, and proves useful for
demonstrating well known results such as reciprocity relations. A formal inversion
technique is also presented that uses a collection of diffraction plane data in scanning
transmission electron microscopy to reconstruct the object potential, even in the
presence of strong multiple scattering.
The new form of the boundary conditions allows for a generalization of a cross-
section expression for calculating inelastic images, making use of the theory of mixed
dynamic form factors. This enables the simulation of images for a range of inelastic
mechanisms, including thermal scattering, used to simulate high-angle annular dark
field imaging, and inner-shell ionization, used to simulate electron energy loss spec-
troscopy images. A multislice form of this expression is given. Selection between the
methods can thus be based on the sample of interest: the Bloch wave method is very
efficient when the sample is crystalline; the multislice method is more appropriate
if the sample lacks periodicity.
The issue of cross-talk, where dynamical probe spreading may result in a signal
containing contributions from several columns and therefore confound direct inter-
pretation, is assessed for high-angle annular dark field imaging. Single atom images
are simulated to provide an estimate of the localization of signal in electron energy
loss spectroscopy, and confirm that the limitations of probe size generally outweigh
3
4
2. Due acknowledgment has been made in the text to all other material used.
3. The thesis is less than 100,000 words in length, exclusive of tables, bibliogra-
phies and appendices.
Scott Findlay
5
6
Acknowledgements
Deepest thanks first and foremost to Les Allen, supervisor, mentor and collaborator,
for his unwavering guidance, encouragement and patience throughout this work, for
many fascinating and insightful discussions, for knowing when I needed direction
and when I needed to find my own.
Special thanks to Mark Oxley, collaborator and mentor, for patient and generous
sharing of his time and knowledge from the nuts-and-bolts of computer code to
insights across the breadth of the research field.
Special thanks to Helen Faulkner, for her advice and empathy regarding the
doctoral student experience, for her humour, for much informing and enjoyable
correspondence.
Special thanks to Nicole O’Leary, for her sense of fun and of balance, for many,
many helpful discussions, for her encouragement.
Special thanks to Will McBride, for thoughtful discussions and new perspectives,
for his sense of humour.
Special thanks to Ben Carson, for his irrepressible energy and humour, for sup-
port and distraction as appropriate, for the daily tea ceremony.
Many thanks to Eireann Cosgriff, Adrian D’Alphonso, Andrew Martin and Chris
Witte, for helpful discussions, for being unafraid to ask difficult questions, and for
reminding me that physics is fun.
Many thanks to Steve Pennycook for his hospitality, guidance and support during
my visit to his research group at Oak Ridge National Laboratory. Thanks also to the
members of his research group: Andy Lupini and Maria Varela, collaborators who
showed immense patience introducing me to the realities of experimental materials
science and microscopy; Yiping Peng, for his correspondence and cheerful willingness
to wade through my rambling e-mails; Naoya Shibata, Albina Borisevich and Matt
Chisholm, for many helpful discussions; and the rest of the Oak Ridge electron
7
8
1 Introduction 19
1.1 Overview and motivation . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.2 Scanning transmission electron microscopy and the approach to
atomic resolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
1.3 Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
1.4 Outline of the thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
9
10
Bibliography 205
13
14
5.1 High-angle annular dark field simulations of ZnS using the 100 keV,
aberration-balanced probe, for single probe positions as a function
of thickness. The contribution from various columns is assessed. . . . 126
5.2 High-angle annular dark field simulations of ZnS using the 100 keV,
aberration-free probe, for single probe positions as a function of
thickness. The contribution from various columns is assessed. . . . . . 127
5.3 High-angle annular dark field simulations of ZnS using the 300 keV,
aberration-balanced probe, for single probe positions as a function
of thickness. The contribution from various columns is assessed. . . . 128
5.4 High-angle annular dark field simulations of InP using the 100 keV,
aberration-balanced probe, for single probe positions as a function
of thickness. The contribution from various columns is assessed. . . . 130
5.5 High-angle annular dark field simulations of ZnS using the 300 keV,
aberration-balanced probe and showing a decomposition of the
total intensity into separate contributions. . . . . . . . . . . . . . . . 132
5.6 K-shell electron energy loss spectroscopy images of single atoms
of O, Si and Ca. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
5.7 Experimental data plots showing the high-angle annular dark field
and Ti L-shell electron energy loss spectroscopy results from
SrTiO3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
5.8 Probe intensity profiles and single atom scanning transmission
electron microscopy images for Ti L-shell. . . . . . . . . . . . . . . . 139
15
5.9 Simulated data plots showing the Ti L-shell electron energy loss
spectroscopy results in the full nonlocal calculation for crystalline
SrTiO3 of thickness 200 Å using the three probes of figure 5.8 (a). . . 140
5.10 Ti L-shell electron energy loss spectroscopy line scans as a function
of depth in SrTiO3 for the 100 keV, aberration-free probe. . . . . . . 142
5.11 Ti L-shell electron energy loss spectroscopy line scans as a function
of depth in SrTiO3 for the 100 keV, aberration-balanced probe. . . . 143
5.12 Ti L-shell electron energy loss spectroscopy line scans as a function
of depth in SrTiO3 for the 100 keV, aberrated probe. . . . . . . . . . 144
5.13 Ti L-shell electron energy loss spectroscopy line scans as a function
of depth in SrTiO3 for the 100 keV, aberration-balanced probe in
the phase object approximation. . . . . . . . . . . . . . . . . . . . . . 146
5.14 High-angle annular dark field image showing an individual La
atom within a layered Lax Ca1−x TiO3 sample. . . . . . . . . . . . . . 148
5.15 High-angle annular dark field image showing an individual La
atom with electron energy loss spectra taken from different
columns. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
5.16 La M -shell electron energy loss spectroscopy line scans for a range
of dopant depths. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
5.17 La M -shell electron energy loss spectroscopy signal as a function
of impurity depth within the Ca column with the probe on the Ca
column, a nearest neighbour O column and a nearest neighbour
TiO column. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
5.18 High-angle annular dark field images of CaTiO3 with and with-
out Gaussian blurring and the corresponding single atom electron
energy loss spectroscopy images. . . . . . . . . . . . . . . . . . . . . . 155
5.19 Signal strength in raster scan regions as a percentage of that on
central Ca column. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
5.20 K-shell electron energy loss spectroscopy signals from a single P
atom in Si, in the [100] orientation, as a function of defocus. . . . . . 158
6.1 Real space intensity images in CaTiO3 , simulated using a 100 keV,
aberration-free probe with aperture semi-angle 25 mrad. . . . . . . . 169
6.2 Comparison of the spatial distribution of electron density in the
frozen phonon and absorptive models in CaTiO3 . . . . . . . . . . . . 172
6.3 Comparison of the spatial distribution of electron density in the
frozen phonon and absorptive models in SrTiO3 . . . . . . . . . . . . . 174
16
6.4 EELS line scan for Ag, M -shell and L-shell, showing the impor-
tance of accounting for thermally scattering electrons. . . . . . . . . . 182
6.5 Ag L-shell electron energy loss spectroscopy line scans testing the
convergence properties in the frozen phonon model. . . . . . . . . . . 185
6.6 Plots derived from the La M -shell electron energy loss spectroscopy
signal from a single atom impurity within CaTiO3 , comparing the
frozen phonon and absorptive models. . . . . . . . . . . . . . . . . . . 187
6.7 The scattering function model potential and assessments of the
validity of different approximations for the probing function. . . . . . 193
6.8 Ag L-shell electron energy loss spectroscopy line scans
simulated with scattering function model. . . . . . . . . . . . . . . . . 194
6.9 Ti L-shell electron energy loss spectroscopy images in SrTiO3
comparing the various models for the inclusion of thermal scattering. 196
List of Tables
17
18
¤£¡
ª@? ¡¢
R
R?
@ ª
¡
Chapter 1
©©
¼ ªH
¡ ?
RH
@ j
Introduction
19
20
energy electrons. Images formed in this mode typically consist of intensity measure-
ments of the wavefunction in an image plane or in the diffraction pattern (Buseck
et al., 1988; Spence, 1988; Shindo and Hiraga, 1998). Such images are – in intention
and often to a good approximation – coherent images, which is to say that the inter-
ference of the electron waves is essential to the form of the images. This is desirable
because for thin specimens it is known that the interaction with the crystal primar-
ily affects the phase of the wavefunction and therefore a method which depends on
this phase information directly, as is the case in holography, or indirectly, through
the mixing of phase and intensity information when the wavefunction propagates, is
necessary to retrieve the structure information. The interpretation of such measure-
ments is generally based upon relating them to the exit-surface wavefunction of the
specimen. This usually involves a phase retrieval method to take measured intensity
information and reconstruct the full complex wavefunction. With recourse to some
appropriate inversion scheme, the exit-surface wavefunction may then be used to de-
termine the projected potential of the specimen (Lentzen and Urban, 2000; Geuens
and Van Dyck, 2002; O’Leary and Allen, 2005).
The STEM geometry comprises a convergent probe focused upon a small area
of the sample. Such instruments may also be used to record coherent images, often
in the diffraction plane, and again these images are often analysed quantitatively.
Convergent beam electron diffraction is one such example (Spence and Zuo, 1992).
But the localization of this probe allows another type of image to be formed: the
integrated signal on some detector as a function of probe position. For example, the
signal may consist of all those electrons reaching an annular detector, or of those
electrons which reach an on-axis detector having undergone some particular energy
loss.
Chapter 1. Introduction 21
resolution.
The imaging of single uranium and thorium atoms on a thin carbon substrate
by Crewe et al. (1970), a major success for the new instrument, was explained
by the dominance of the elastic cross section for large angles.2 Ratios between
elastic and inelastic scattering provided in that work neglect, appropriately for that
sample, any channelling effects. However the focus on the importance of elastic
scattering to the detected images also dominated early considerations of imaging
crystalline materials (Cowley, 1976; Spence and Cowley, 1978), and using an annular
detector with inner angle barely exceeding the probe-forming angle was shown to
give very poor contrast in cases where Bragg scattering was significant (Donald
and Craven, 1979). Following the suggestions of Treacy et al. (1978) and Howie
(1979), high-angle annular detectors with inner angle significantly larger than the
probe forming angle began to be used (Pennycook and McMullan, 1983; Pennycook
et al., 1986). Such detectors not only eliminate the majority of the Bragg scattering
from crystalline specimens but also mean that the resulting images of the thermally
scattered electrons tend towards the Z 2 scaling of Rutherford scattering, where Z
denotes the atomic number, and thus allow for stronger Z contrast in imaging.
This innovation led to a renewed interest in STEM, particularly in materials
analysis, as it was found that the Z-contrast images were effectively incoherent
images and could be directly interpreted (Pennycook and Jesson, 1990, 1991; Loane
et al., 1992). Chemical analyses were performed, through Z-contrast imaging of
species with significantly different weights (Pennycook and Boatner, 1988; Voyles
et al., 2002), through direct use of spectroscopic methods (Muller et al., 1993), and
through the simultaneous use of both techniques (Batson, 1993; Browning et al.,
1993; Muller et al., 2004). The combination of techniques is particularly powerful for
profiling at the atomic scale. It has been used to determine the electronic structure
in gate oxides only a few interatomic distances thick (Muller et al., 1999), and for
the spectroscopic location and imaging of single atoms in the bulk (Varela et al.,
2004).
In parallel with the material science applications ran the push to higher resolu-
tions as determined by decreasing probe sizes (Pennycook, 1989; Shin et al., 1989;
McGibbon et al., 1995; Nellist and Pennycook, 1998b; Batson, 2003; Nellist et al.,
2004). Pivotal to this was the development of aberration correctors capable of con-
trolling the third order spherical aberration (Haider et al., 1998a,b; Krivanek et al.,
2
That analysis, based upon the theoretical cross-section calculations of Lenz, may not be ac-
curate in detail but the improvement in image quality of their ratio technique was clear (Treacy
et al., 1978).
24
1999; Dellby et al., 2001; Bleloch and Lupini, 2004), which had long been appreciated
as a major hindrance to resolution in STEM (Crewe, 1980). Sub-Ångstrom resolu-
tion has been demonstrated in a variety of contexts with such aberration-corrected
probes (Batson et al., 2002; Batson, 2003; Falke et al., 2004; Nellist et al., 2004;
Hutchison et al., 2005).
Such simulations are also presented in conjunction with, and as support for, ex-
perimental data. Much theoretical work has gone into the numerical processing and
tidying of experimental images (Nellist and Pennycook, 1998a; McGibbon et al.,
1999; Watanabe et al., 2002). But, with a few noteworthy exceptions (Loane et al.,
1992; Silcox et al., 1992; Watanabe et al., 2001a), relatively little quantitative com-
parison has been made between experiment and theory in HAADF imaging. Though
simulation is often used in analysing electron energy loss spectroscopy, the approach
Chapter 1. Introduction 25
3
For example, the assumed direct relation between signal and structure is the basis of the
Z-contrast tomography of Midgley et al. (2001).
26
1.3 Approach
Before embarking on an introduction to the concepts, techniques and tools that will
be used in what is to follow, a comment should be made on the methodology. The
treatment is given from a theoretical viewpoint. As such, the idealized assumptions
on which the work is based are emphasized at the outset. In many ways this approach
fails to do justice to several experimental details. This is not to say that such details
are uninteresting or insignificant. However this approach has been chosen to make
the conceptual underpinnings explicit and clear.
That said, for any theoretical endeavour to have practical significance, the syn-
thesis with experimental considerations must be made. Later chapters present direct
comparisons with experiment, and, through simulation, explore questions of both
theoretical and experimental interest.
Chapter 2
©©
¼ ªH
¡ ?
RH
@ j
This chapter introduces the basic conceptual framework of theoretical electron mi-
croscopy, the means of modelling the interaction between high energy electrons and
material specimens, and the notation used to describe the process. While adding
little new to the field, this background theory is a necessary and integral part of
making explicit the theoretical assumptions underlying the new contributions of
later chapters.
29
30
Incident
wavefunction
Wavefunction
in the sample
Exit-surface
wavefunction
Figure 2.1: Idealization of the scattering geometry for (a) CTEM and (b) STEM.
The arrows are drawn after the manner of optical ray tracings and may be loosely
interpreted as representing momentum vectors.
potential approximation and the zero order Laue zone approximation, become ques-
tionable. For the present we will assume a monochromatic incident beam, though
this assumption will later be relaxed in order to investigate the role of temporal
incoherence on the images formed, particularly as it relates to chromatic aberration
in the probe-forming lens.
The electron beam incident on the crystal undergoes refraction, through inter-
action with the mean crystal potential, and diffraction, through interaction with
the varying part of the crystal potential. In CTEM, the probability density of the
incident fast electron is periodic – laterally, not so along the optical axis – within
the crystal with the periodicity of the crystal. This is in accordance with the Bloch
theorem which states that the electron wavefunction is equivalent in equivalent cells
up to a phase factor. In STEM, the probe does not have the translational invariance
of the CTEM regime and as such the electron intensity is not periodic within the
crystal, but rather localized.
Because the fast electrons begin very far from the crystal with constant energy E
(the accelerating potential), the motion of these electrons inside the crystal, which is
described by the solution to the Schrödinger equation, is in essence determined by the
constraint that the kinetic and potential energy sum to this fixed total energy. This
treatment makes definitive divisions of the problem into sections: (a) the electrons
prior to incidence; (b) the electrons inside the crystal; and (c) the electrons emitted
from the far side of the crystal. The treatment generally involves the paraxial
approximation. The Schrödinger equation is solved separately in the three regions,
and then the boundary conditions are applied at the entrance surface to determine
the wavefunction inside the crystal (with no account taken of reflection back to
the first region), and then again at the exit surface to determine the wavefunction
outside the crystal once more (again without possibility of back-reflection at the
interface). This treatment is consistent with the high energy approximation – it is
well established. More elaborate procedures are possible (Van Dyck, 1976; Kim and
Sheinin, 1982), but unnecessary for the geometry and samples we will use.
Finally, on exiting the crystal, post-specimen optics may be employed to form
images, obtain diffraction patterns, or perform cross-section measurements. Alter-
natively, post-specimen optics may be absent in some forms of diffraction pattern
acquisition and cross-section measurements.
In the following sections we will introduce the mathematical tools required to
describe the wavefunctions, the effect of the objective lens, and the scattering within
the specimen.
32
exp(2πik · r) . (2.1)
where r is the real space variable and k the reciprocal space variable. We will
generally adhere to the convention that functions in reciprocal space are denoted by
3
This point is crucial, and there have been many papers in the literature publishing tables
contrasting the various conventions. Unfortunately not all of these are correct. As pointed out by
us in Rossouw et al. (2003), our earlier paper, Allen et al. (2003b), used a convention for the lens
transfer function which was inconsistent with the sign convention used for the propagator through
the crystal. The ramifications of this error for the results presented in that paper were that the
probe used was the complex conjugate (or parameter sign reversal) of that claimed. Since the
resultant parameter set is achievable with Cs correctors, the physical deductions made there are
still significant. The error has been fully amended in all calculations and figures included in the
present document.
Chapter 2. Background theory: notation, conventions and models 33
upper case letters, in either the Greek or English character sets, while functions in
real space are denoted by lower case letters.
It is convenient to choose a coordinate system in which the optical axis is the
z-axis of a Cartesian system. In the typical case where the surface of the slab-like
sample is perpendicular to the optical axis, the xy-plane will be coincident with the
entrance surface of the crystal, with z increasing into the crystal.
General reciprocal space vectors will often be denoted by p or q. In relation
to the reciprocal lattice of the crystal, several different vector notations are used in
what follows and the distinctions are worth emphasizing. Capitalized vectors G and
H denote what will be referred to as physical reciprocal lattice vectors, those in the
reciprocal space lattice which is conjugate to the real space crystal lattice. We denote
the set of these vectors by {F}. A function written as a Fourier series with spatial
frequencies taken solely from the set {F} will have the periodicity of the crystal.
For plane wave incidence, the wavefunction at the surface is consistent with the
periodicity of the crystal, and Bloch’s theorem enables the resulting wavefunction
to be expressed in reciprocal space in terms of the set {F}.
However the probe in STEM does not have the periodicity of the crystal. In
principle, sampling this wavefunction requires a continuous set of reciprocal space
frequencies. Numerical implementation is necessarily based on a finite set, but the
mesh will be finer than that of the set {F}. The lower case vectors g and h will be
used to denote reciprocal space vectors of this finer mesh, as required to sample the
transverse momentum components in a STEM probe. We denote the set thereof by
{f }.
This notation that parallels the set of physical reciprocal lattice vectors is chosen
to emphasize that the set used is still discrete (cf. a general reciprocal space vector
set). Indeed it may be equivalently generated by an m × n supercell, that is to say
an effective unit cell formed by tiling the natural unit cell of the crystal m times
along one axis and n along the other. By using all the physical reciprocal lattice
vectors of the supercell, a much finer sampling of the reciprocal mesh is achieved
than would be the case for the (smaller) conventional unit cell.
It will be seen in what follows that Fourier transforms are intimately involved in
propagation of the wavefunction. Numerical evaluation via discrete Fourier trans-
form introduces an effective periodicity. In STEM this is unphysical. Thus the
choice of supercell size should be such that the spurious repetitions of the probe
introduced by the periodic treatment are sufficiently far apart that they do not sig-
nificantly interact. The same approach is taken in handling crystal defects in the
34
1 q = (000)
q = ±(½00)
gy
0 q = ±(0½0)
q = ±(½½0)
-1
-2
-3
-3 -2 -1 0 1 2 3
gx
Figure 2.2: The vector set {f } generated using a 2 × 2 supercell (or equivalently the
direct sum of the four q vectors shown and the vector set {F}).
In the high energy approximation, where the Bloch states may be determined
from reciprocal lattice vectors in the zero order Laue zone, these vectors will be
taken to lie parallel to the crystal surface.4
4
In assuming collinearity of real space and reciprocal space lattice vectors we implicitly limit
the crystal geometries which are properly described by the formulae developed. This is done for
simplicity of notation. The generalization to further crystal types follows from careful bookkeeping
in the usual manner.
Chapter 2. Background theory: notation, conventions and models 35
h2
− ∇2 ψ(r) − eV (r)ψ(r) − ieV 0 (r)ψ(r) = eEψ(r) , (2.4)
8π 2 m
where V (r) denotes the elastic crystal potential, taken as positive when electrons
are attracted to the atoms; V 0 (r) denotes the absorptive potential; and E denotes
the accelerating potential, all in electron volts. The electron charge e in the above
equation is meant in magnitude only, its sign is taken as positive (e = |e|).
Significant assumptions are implicit in writing equation (2.4). The absorption
term indicates that there are many processes going on in the crystal. Though we are
at present interested in determining the form of the elastic wavefunction, we must
account for the reduction of the number electrons in the elastic channel with prop-
agation through the crystal due to the portion of electrons which undergo inelastic
scattering and are removed from the elastic beams. Various levels of approximation
describing where such electrons go will be central to much of our work, but at this
stage we note these processes only in their effect upon the elastic channel. The
most general form would make use of a nonlocal potential, however it is generally
accepted that the dominant absorption mechanism is thermal or phonon scattering.
In the high energy approximation, and using an Einstein model, this leads to a local
effective scattering potential. We will assume an absorptive potential arising from
an Einstein model for thermal scattering.6
Let us define a modified wavefunction via
Using equation (2.5) and the standard high energy approximations (Humphreys,
1979a), equation (2.4) may be simplified to
dφ(r) i £ 2 ¤
= ∇⊥ + W (r⊥ ) φ(r) , (2.6)
dz 4πK
where
W (r⊥ ) = U (r⊥ ) + iU 0 (r⊥ ) , (2.7)
with
X Z t
2πiG·r⊥ 8π 2 me 1 £ ¤
U (r⊥ ) = UG e = V (r⊥ , z) − V dz , (2.8)
G6=0
h2 t 0
X Z t
0 0 2πiG·r⊥ 8π 2 me 1
U (r⊥ ) = UG e = V 0 (r⊥ , z)dz , (2.9)
G
h2 t 0
where V denotes the mean elastic potential. Note that the so-called reduced poten-
tials U (r⊥ ) and U 0 (r⊥ ), which have dimensions of Å−2 , can be written as Fourier
series. The coefficient U0 is excluded from the potential because it has been in-
cluded in the refraction-corrected wavevector. Note too that by averaging over t in
equations (2.8) and (2.9) we have adopted the projected potential approximation.
Taking advantage of the periodicity of the crystal, the eigenstates of equation
(2.6) are Bloch states. They may be expressed in Fourier series form as
i
X
φiK (r) = e2πλ (K⊥ )z i
CG (K⊥ )e2πi(K⊥ +G)·r⊥ , (2.10)
G
where λi = γ i + iη i is the eigenvalue for Bloch state i. The quantities γ i are called
the anpassung and the quantities η i are called the absorption coefficients (Allen
and Rossouw, 1989). We have labelled the Bloch states by the incident wavevector
K, and written the Fourier coefficients and eigenvalues of the Bloch state with an
explicit dependence on K⊥ , the component of the wavevector perpendicular to the
optical axis. We will tend to drop this explicit wavevector dependence excepting
in cases where doing so might cause confusion. The Bloch state coefficients and
eigenvalues satisfy the periodicity conditions (Kästner, 1993)
i i
CG−H (K⊥ + H) = CG (K⊥ )
λi (K⊥ + H) = λi (K⊥ ) . (2.11)
In the absence of absorption, the Fourier coefficients of the Bloch state satisfy the
Chapter 2. Background theory: notation, conventions and models 37
orthogonality conditions
X j
i
CG CG = δi,j (2.12)
G
X
i i
CG CH = δG,H . (2.13)
i
0
where WG = UG + iUG .
The dynamical equations may be recast into matrix form (Humphreys, 1979a):
£ ¤
AC = C 2Kλi D . (2.16)
Writing simultaneous equations in matrix form allows much freedom as to the or-
dering of rows and columns in the matrix. Allen and coworkers (Allen et al., 1998,
1999, 2000) write the structure (Bethe) matrix A as
.. .. .. .. ..
. . . . .
··· P (K⊥ , H) WH−G WH WH+G W2H
···
··· WG−H P (K⊥ , G) WG W2G WG+H
···
A=
··· W−H W−G P (K⊥ ) WG WH ,
···
· · · W−G−H W−2G W−G P (K⊥ , −G) W−G+H ···
··· W−2H W−H−G W−H W−H+G P (K⊥ , −H) · · ·
.. .. .. .. ..
. . . . .
(2.17)
where P (K⊥ , G) = −(K⊥ + G)2 + iU00 , since it makes the symmetry across the
anti-diagonal more evident.
38
The diagonal matrix [2Kλi ]D contains the complex eigenvalues of A. The matrix
C contains as its columns the eigenvectors of A:
.. .. .. ..
. . . .
1
CH CH 2
··· CH i
···
1 2 i
CG CG ··· CG ···
C=
C01 C02 ··· C0i ··· .
(2.18)
1 2 i
C−G C−G ··· C−G ···
C1 2 i
···
−H C−H · · · C−H
.. .. .. ..
. . . .
where notation [C −1 ]im denotes the element in the i th row and m th column of the
inverse matrix of C.
and hence let us define a notation for the elements of the inverse matrix through the
G
inversion of the scripts: [C −1 ]i j ≡ [C −1 ]i j . In particular, the excitation amplitude
for plane wave incidence may now be written as αi = [C −1 ]i (N +1)/2 = [C −1 ]0i .
Let us summarize thus far. The Schrödinger equation has been written in a matrix
form. For the purposes of the direct problem, the structure is assumed to be known,
and so, recalling that the incident electron wavevector is controlled experimentally,
all the elements in the matrix A are known. The eigenvalues and eigenvectors can
be determined numerically and so the eigenvalues λi and the matrix C are known.
Hence all the quantities in equation (2.20) are known. The wavefunction in the
crystal has been determined; the Schrödinger equation has been solved.
Let us introduce the vector v = (vG ), the vector representation of the exit-surface
wavefunction, and u, the vector representation of the entrance-surface wavefunction
which for normal plane wave incidence is given by u = (δG,0 ). The scattering matrix
S is then defined through
v = Su . (2.22)
i
X
= Cni exp(2πiλi t)[C −1 ]i (N +1)/2
i
= {exp [(iπt/K)A]}G,0 , (2.23)
where the final line follows from the use of the spectral representation for a matrix
exponential and comparison with equation (2.19). Note that, from this principal
direction, only the single central column of the S matrix has been obtained. It
40
seems natural to extend this form to define the whole matrix through the identity
£ ¤
S = exp [(iπt/K)A] = C exp(2πiλi t) D C −1 . (2.24)
Using the periodicity relations of equation (2.11) it may be shown that the math-
ematically natural extension used to define all of S supports the physical interpre-
tation of equation (2.22) for incident plane waves tilted with respect to the surface
normal by a reciprocal lattice vector: u = (δG,H ) (Allen et al., 1999). If the inci-
dent plane wave has a tangential component K⊥ = q then the scattering is again
i
described by equation (2.24) constructed using elements CG (q) and λi (q). These in-
terpretations immediately extend to an incident probe formed by the superposition
of such waves. Together these ideas provide the framework for the generalization to
a STEM probe or indeed to any arbitrary wavefunction incident upon the crystal.
reciprocal space as
where the Fourier transform is taken only over r⊥ , that is to say in two dimensions.
The Fourier transform pair p and P effect propagation in the paraxial approxi-
mation through a distance δzn . The reciprocal space form is
£ ¤
P (q, δzn ) = exp −πiλq2 δzn . (2.27)
The Fourier transform pair q and Q effect the change in amplitude and phase
occasioned by traversing the nth slice of the crystal. The real space form is
where Wn (r⊥ ) is the reduced potential projected over the nth slice,7 and σ = πλ is
a constant.8
The effect of a lens in the imaging of a wavefunction in the electron microscope may
be described by multiplying the reciprocal space representation of that wavefunction
by a contrast transfer function T (q). This transfer function serves to truncate the
extent of the wavefunction in reciprocal space by imposing an aperture. It also
modifies the phase of the wavefunction in reciprocal space in accordance with the
lens aberration function.
The contrast transfer function is given by
· ¸
2π
T (q) = A(q) exp −i χ(q, λ) . (2.29)
λ
7
Often this potential is simply the projected potential of the entire crystal, but in some cases it
is convenient to use the projection for a thinner slice, the full cell containing an integer number of
such (possibly different) sub-slices. This is the motivation for retaining a symbolic n dependence;
the projected potential used for different depths may indeed be different.
8
In the literature σ often includes additional factors such that it is the projected potential rather
than the reduced projected potential which appears in equation (2.28).
42
in which qmax is the reciprocal space aperture cutoff, and c is a constant. In CTEM
c = 1 must be chosen in order to use T (q) as the transfer function to image an
exit-surface wavefunction. However for STEM we reserve the option of choosing c
differently in order to enforce a particular form of normalization.
The most general form of χ(q, λ) = χ(q, θ, λ) is (Krivanek et al., 1999)
In practice the effects of the astigmatic terms are often directly visible and may
be corrected manually. Neglecting these, we identify the defocus, ∆f ≡ C10 , third
order spherical aberration, Cs ≡ C30 , and fifth order spherical aberration C5 ≡ C50 ,
and write
1 1 1
χ(q, λ) = ∆f (λq)2 + Cs (λq)4 + C5 (λq)6 , (2.32)
2 4 6
In STEM the lens does not modify the exit-surface wavefunction, in the manner
in which it is utilized in CTEM, but rather is used to form the incident convergent
probe. For this purpose the object being imaged is ideally an effective point source.
Since the Fourier transform of a point source equally weights all frequencies, the
Fourier transform of the convergent probe wavefunction is precisely the transfer
function of the lens (Spence and Cowley, 1978). The incident wavefunction so formed
Chapter 2. Background theory: notation, conventions and models 43
in which R denotes the position on the surface about which the probe is centred.
Note that we have used a discrete summation notation rather than the more for-
mally correct integral notation – some discretization is necessary for computational
purposes, and, subject to sufficient sampling, is accurate. This expression implicitly
ignores effects due to chromatic aberration and finite source size, but these points
will be discussed later.
Scherzer (1949) analyzed the dependence of the lens transfer function on coher-
ent aberrations. His work is best known for the optimum conditions – now known
as Scherzer conditions – for imaging in CTEM, where a fixed value for Cs was as-
sumed and in terms of which the optimum values of defocus and aperture semi-angle
were given. Resolution is primarily a function of the number of Fourier frequencies
present, and thus resolution may be improved if the aperture is made wider. The
limitation is that if there is significant variation in the aberration function then
adjacent frequencies are transmitted without a close phase relationship and this
leads to considerable difficulty in interpreting the images. The Scherzer condition
in CTEM imaging prescribes a value of defocus such that the first term in equation
(2.32) is balanced against the second such that the aberration function varies slowly
within the prescribed aperture, allowing for the reliable transfer of a larger range of
frequencies.
Scherzer also gave optimal criteria for what we will later describe as incoherent
imaging. In STEM the lens is used to form a probe, and the ideal probe is as narrow
as possible without possessing significant subsidiary peaks, often referred to as probe
tails (Scherzer, 1949; Pennycook and Nellist, 1999).9 Loosely speaking, to obtain
column imaging in STEM it is necessary for the probe width to be smaller than the
inter-column spacing. Scherzer noted that the limitation on both techniques was
the value of Cs , which was unavoidable given the manner of the lens typically used
in electron microscopes at the time.
The recent emergence of Cs correctors has been accomplished by the use of non-
spherical lenses (Haider et al., 1998a; Krivanek et al., 1999). With such aberration
9
Seeking a probe as fine as possible yet without probe tails is predicated on balancing resolu-
tion – measured by the highest spatial frequencies present – with the conditions allowing direct
interpretation. However see Van Aert et al. (2002) for an alternative perspective on optimality
criteria more in keeping with the goal of quantitative analysis.
44
Intensity
Intensity
Intensity
0.6 0.6 0.6
2 2 2
1 1 1
Phase
Phase
Phase
0 0 0
-1 -1 -1
-2 -2 -2
-3 -3 -3
0.0 0.5 1.0 1.5 2.0 2.5 3.0 0.0 0.5 1.0 1.5 2.0 2.5 3.0 0.0 0.5 1.0 1.5 2.0 2.5 3.0
χ/λ
χ/λ
0.0 0.0 0.0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
Figure 2.3: Intensity (top row), phase (middle row) and aberration function (bottom
row) for: the aberrated probe, with ∆f = −496.7 Å, Cs = 0.5 mm, and aperture
semi-angle of 14 mrad (left column); the aberration-balanced probe, with ∆f =
62 Å, Cs = −0.05 mm, C5 = 63 mm, and aperture semi-angle of 20 mrad (centre
column); and the aberration-free probe, with aperture semi-angle of 25 mrad (right
column).
3
m− 2 ). This is often remedied by adopting a box normalization, where the wavefunc-
√
tion expression would contain a pre-factor of 1/ V for some appropriate volume V .
The volume of the crystal would be a natural choice here. However the notational
increase occasioned by including such a factor in all equations, when it is frequently
uninformative, has led to its omission in most of the theoretical literature on electron
diffraction.10
46
We will adopt this convention for simplicity of notation. However relating the
cross-section expression to be presented in chapter 4 to experimental results in a
quantitative fashion requires a rigorous handling of the wavefunction normalization.
We will therefore adopt the practice whereby the normalization is explicitly inserted
as a pre-factor to expressions requiring normalization. Thus the wavefunction in the
crystal should, in the absence of absorption, be normalized such that
Z
1
|ψ(R, r)|2 = 1 . (2.35)
V V
The probe wavefunction can be extended to three dimensions using the paraxial
form of the free-space propagator [cf. equation (2.27)] to give
X 2
ψ(r⊥ , z) = T (q)e−πiλq z e2πiq·(r⊥ −R) . (2.36)
q
from which equation (2.35) readily follows if c in equation (2.30) is chosen such that
X
|T (q)|2 = 1 . (2.38)
q
In the STEM work to follow, the normalization conditions of equations (2.35) and
10
This trend aids a related subtlety in the high energy diffraction literature, that images are
given as intensities of the wavefunction in the plane rather than as currents. This approach is
justified in the paraxial approximation since they are, to an excellent approximation, proportional
to one another. However this elision has aided the habit of not explicitly giving the units used.
Chapter 2. Background theory: notation, conventions and models 47
Sr
Ti
Zn
S
Ag
Figure 2.4: (a) Crystal structure for cubic perovskite, (b) labelled as for SrTiO3 .
(c) The [001] zone axis projection of (a), with a dashed line indicating a [110] scan
line. (d) Crystal structure for zinc-blende material, (e) labelled as for ZnS. (f) The
[110] zone axis projection of (d), with a dashed line indicating a [001] scan line. (g)
Crystal structure for face-centred cubic material, (h) labelled as for silver. (i) The
[001] zone axis projection of (g), with a dashed line indicating a [100] scan line.
been introduced into the sorts of signals we will consider when sensible estimation
of these parameters is made.
Chapter 2. Background theory: notation, conventions and models 49
ing electrons which have not lost any energy in their interaction with the crystal.
Conversely, inelastic scattering involves a loss of energy on the part of the electrons
and some change of state in the crystal. We have already mentioned three possible
mechanisms of inelastic scattering: thermal diffuse scattering, plasmon scattering,
and inner-shell ionization.
Two wavefunctions are said to be coherent if their superposition is achieved by
adding complex amplitudes. The result is that interference, be it constructive or de-
structive, may occur. Conversely, two wavefunctions are said to be incoherent if the
superposition is achieved by adding the intensities. The result is that no interference
between the wavefunctions can occur.11 Holography, for instance, requires coherence
between the scattered wave and the reference wave in order that the superposition
causes interference, the resultant intensity being what is recorded in forming a holo-
gram. Inelastic scattering is generally considered to be incoherent with respect to
the elastic channel, and thus in determining the intensity from many such channels
it is the intensities, and not the complex wavefunctions, which are added.
These concepts need to be comprehended in order to appreciate the manner in
which the expressions coherent and incoherent imaging are used in the literature.
These appellations are applied based on the visual appearance of the images, ir-
respective of the underlying physical mechanisms. Thus images are referred to as
coherent if they display interference fringes and contrast reversals and other such
features which generally result from coherent addition. Similarly, images are referred
to as incoherent if they do not display such features.
The subtlety arises in attempting to relate these ideas. Some connections are
reliable matters of definition. For example, coherence is a requirement in order to
get coherent images. Some connections are consequences of the underlying physics,
for example inelastic channels are incoherent with respect to the elastic channel.
Some connections are subtle, for example it is possible that incoherent images may
result from coherent interference, which is to say that while the interaction is coher-
ent, the resulting images do not display the usual hallmarks of coherent interaction
– interference fringes or contrast reversal. Some connections are complicated by
the constraints of experimental apparatus, for example thermal scattering, while
formally involving an energy loss and hence incoherent with respect to the elastic
beams, involves such a small loss of energy that the contribution from thermally
scattered electrons may not easily be experimentally separated from the contribu-
11
For simplicity we have described only the two limiting scenarios: complete coherence and
complete incoherence. The spectrum between these extremes, described as partial coherence, may
be realized in practice (Born and Wolf, 1999).
Chapter 2. Background theory: notation, conventions and models 51
where the sum over n encompasses all atoms with scattering factors fn in positions
τ n in the unit cell of volume Vc . The Debye-Waller factor Mn (q) = 2π 2 hu2n iq 2 , where
hu2n i is the projected mean square thermal displacement. If the integration is taken
over the complete solid angle then the MDFFs are proportional to the absorptive
potential for thermal diffuse scattering (Allen and Rossouw, 1990). By integrating
over smaller solid angles (defined by the detector), MDFFs can be calculated which
54
The MDFFs for inner-shell ionization of atoms of a particular type are written
1 X
µh,g = exp[−Mn (g − h)] exp[2πi(g − h) · τ n ]F (h, g) , (2.40)
2πkVc n
where the sum over n now refers only to the pertinent atom species. The Debye-
Waller term accounts for thermal smearing of the ionization potential (Cherns et al.,
1973). The atomic scattering factor for ionization F (h, g) contains a product of a
pair of non-diagonal transition amplitudes, each derived from an isolated transition
matrix element for an (e,2e) interaction (Maslen and Rossouw, 1983; Allen and
Josefsson, 1995, 1996; Oxley and Allen, 2000). This may be written as (Oxley and
Allen, 1998, 2001)
Z Z
1 0 M(κ, q+ g)M∗ (κ, q + h)
F (h, g) = 3 2 k dκdΩ , (2.41)
2π a0 |q + g|2 |q + h|2
where a0 is the relativistic Bohr radius, k 0 is the wavevector after inelastic scattering
and the transition matrix element
Z
M(κ, q + g) = uf (κ, r) exp[2πi(q + g) · r]ui (r)dr . (2.42)
Here ui (r) and uf (κ, r) denote the bound and continuum state wavefunctions of
the target electron, ejected with momentum κ. These MDFFs can be used to
simulate the contrast by electron energy loss spectroscopy or energy dispersive x-ray
analysis. Calculations in this thesis of inner-shell ionization are carried out using
the transition matrix elements based upon realistic wavefunctions in an angular
momentum representation (Saldin and Rez, 1987; Oxley and Allen, 1998, 2000).
Two points which were emphasized by Allen and coworkers (Allen et al., 2003b;
Rossouw et al., 2003) deserve brief mention here. The first is that the quantum
mechanical phase of the matrix elements M(κ, q + g) appearing in equation (2.41)
is vital for correctly describing the scattering. The second relates to whether and
when the effective scattering potential is a nonlocal potential. Pragmatically, a
nonlocal potential requires two vector subscripts to correctly describe the MDFFs,
Chapter 2. Background theory: notation, conventions and models 55
µh,g . A local potential is one for which µh,g ' µh−g,0 and so may be described by a
single vector subscript. The thermal diffuse scattering and energy dispersive x-ray
spectroscopy interactions are adequately described by local effective interaction po-
tentials, while the electron energy loss spectroscopy interactions generally require a
nonlocal effective interaction potential for accurate description (Allen and Josefsson,
1995; Rossouw et al., 2003). The nonlocal nature of the effective interaction poten-
tial arises from the analysis whereby inelastic signals are indirectly deduced from the
elastic wavefunction alone, i.e. without directly calculating inelastic wavefunctions
(Allen and Josefsson, 1995).
These points are not controversial, though they were emphasized by Allen and
coworkers (Allen et al., 2003b; Rossouw et al., 2003) because it was felt that these
points were not sufficiently appreciated in the literature. All calculations in this
thesis have appropriately dealt with these matters.
2.9 Conclusion
Having first introduced the conceptual model for transmission electron microscopy,
this chapter has explored the technical means by which the interaction between fast
electrons and the crystal is described. The theory for calculating the wavefunction
in the specimen has been provided and the notation to be used for this purpose
carefully explained. The description of the lens transfer function and its role in
forming fine probes has been reviewed. The significance of the MDFF elements,
which figure centrally in the calculation of signals derived from inelastic scattering
events, has been sketched. Having thus established the basic scaffolding, both of
notation and theory, we may now construct a theoretical edifice to describe imaging
in the scanning transmission electron microscope.
56
¤£¡
ª@? ¡¢
R
R?
@ ª
¡
Chapter 3
©©
¼ ªH
¡ ?
RH
@ j
In the previous chapter we described the conceptual set-up required to model conven-
tional transmission electron microscopy (CTEM) and scanning transmission electron
microscopy (STEM). The theoretical tools – the Bloch wave method, the dynam-
ical equations, the lens transfer function, inelastic scattering coefficients or mixed
dynamic form factors – were developed with reference primarily to CTEM and its
literature.
In this chapter we will focus on STEM. From the foundations of the previous
chapter the theory needed to simulate wavefunctions in STEM is quickly built up.
In the Bloch wave formulation, the primary difference is in the form of the boundary
conditions. A new form of the boundary conditions is presented, one which regards
the transition across the boundary as a unified step. The discussion ranges from the
theoretical description through to efficient methods of performing calculations.
Wavefunctions have, of course, been simulated previously in STEM. It is demon-
strated that the method presented here is formally equivalent to those used previ-
ously. However, strong emphasis is made upon the conceptual differences between
these approaches, and the different physical insights obtained in each.
The principal result of the thesis, the cross-section expression for inelastic scat-
57
58
tering and its use for simulating incoherent images, will be delayed until chapter 4.
In this chapter the new expression for the wavefunction is used to derive expressions
for coherent imaging and recast some well known results for coherent imaging in
the present formulation. In particular, the reciprocity theorem, relating bright field
STEM imaging and CTEM, will be derived.
crystal wavefunction which incorporates all the transverse momenta in the incident
probe. All such momentum components are accommodated in the propagator within
the crystal, as in a multislice formulation of STEM contrast (Loane et al., 1988;
Hillyard et al., 1993; Anderson et al., 1997). This method is equivalent to that in
the previous literature, and we will see that the crystal wavefunctions produced in
these seemingly disparate approaches are identical. However the two formulations
offer different insights into the physics of the problem. A significant point of the
new technique is that it leads to a different form of the Bloch state excitation
amplitudes, which prove to be dependent on probe position and focus, as well as
crystal orientation. We will see in the next chapter that the current method lends
itself readily to calculating incoherent images.
For a focused coherent electron probe, the reciprocal lattice mesh associated with
the crystal must be adequate not only for a sampling of the structure of the crystal
but also for sampling of the transverse momentum components of the probe. This
results in a structure matrix A which contains elements usually excluded when plane
wave incidence is considered. This A-matrix can be of large order, increasing the
numerical demands of the eigenvalue problem to be solved considerably. However
the A-matrix may be block diagonalized in the case of perfect crystals.
Block diagonalization leads to a natural subdivision of the wave function into
pieces based on sets of reciprocal lattice vectors differing by what we have referred
to as physical reciprocal lattice vectors (cf. section 2.2). The block diagonaliza-
tion provides insight into the physics involved in the coupling of the probe to the
wavefunction in the crystal. It also lends itself to ready proofs of reciprocity for
coherent contrast and to a demonstration of the lack of contrast in lattice imaging
for apertures which do not encompass a physical reciprocal lattice vector.
We will conclude the chapter by considering the possibility of using coherent
imaging in STEM as a means of solving the inverse scattering problem, that is to
say as a means to determine the projected potential of the structure.
where R denotes the position of the probe and r⊥ is a general position vector in
the xy-plane from the origin of the unit cell. Excitation amplitudes αi (R) are to be
derived from boundary conditions alone, and it is anticipated that these will depend
on the coherent aberrations of the lens as well as the probe position. This is in
contrast to previous Bloch wave approaches in which a spherical wave is constructed
from a coherent summation of phase-linked plane waves. In that case, the excitation
amplitudes αi = [C −1 ]0i , which in the absence of absorption reduce to the more
familiar αi = C0i∗ , are associated with each phase-linked component (Pennycook
and Jesson, 1991; Nellist and Pennycook, 1999; Yamazaki et al., 2000b). We will
assume at present that the probe is symmetric about the optical axis and will leave
out explicitly writing the dependence upon the wavevector.
Recall from equation (2.10) that Bloch states for propagation of the wavefunction
in the crystal may be written in the form
X
φi (r⊥ , z) = exp(2πiλi z) Cgi exp(2πig · r⊥ ) . (3.2)
g
Note that we are using reciprocal space vectors from the set {f }, that is to say
corresponding to the supercell, as per the discussion in section 2.2.
Recall from equation (2.34) that the real space probe amplitude may be con-
structed from the Fourier transform
X
T (r⊥ ) = T (p) exp[2πip · (r⊥ − R)] , (3.3)
p
where the summation over p is taken over the mesh in g defined by the supercell.
R
The excitation amplitude is given by the overlap integral A1 φi∗ (r⊥ , 0)T (r⊥ )dr⊥
(Broeckx et al., 1995). This complex conjugate notation is appropriate in the ab-
sence of absorption, but to include absorption we make the generalization to the
inverse matrix elements (Allen and Rossouw, 1989), as discussed in section 2.3 when
describing the eigenvector matrix and its inverse matrix. Taking this one step fur-
ther, use of equations (3.2) and (3.3) yields the excitation amplitude2
X X Z
1
i
α (R) = [C −1 ]gi T (p) exp(−2πip · R) exp[2πi(p − g) · r⊥ ]dr⊥
g p
A
X
= [C −1 ]gi exp(−2πig · R)T (g) . (3.4)
g
2
See section 2.3 for clarification of the inverse matrix notation [C −1 ]gi adopted.
Chapter 3. Wavefunctions and elastic imaging 61
This equation, in which each Fourier component [C −1 ]gi is coupled via a position-
dependent phase factor to the lens transfer amplitude T (g), is the key result under-
pinning our description of coherent and incoherent STEM contrast.
It is informative to explore the general result by considering special cases. That
of an incident plane wave and of a point source are now examined. Plane wave inci-
dence is obtained by setting T (g) = δg,0 , leading to the usual plane wave boundary
condition
αi = [C −1 ]0i (plane wave) . (3.5)
The key result is that the excitation amplitudes depend on probe position and
lens aberrations. Figure 3.1 shows the excitation amplitudes for a few select Bloch
states in ZnS, viewed along the [110] zone axis. For simplicity we will neglect
absorption and so [C −1 ]0i = C0i∗ . We use two STEM probes from figure 2.3, the
100 keV aberration-balanced probe (first column) and the 100 keV aberration-free
probe (second column), as well as a point-source probe as per equation (3.6) (third
column).
In figure 3.1, the number below the state label for each row is the value of |C0i∗ |,
which is the magnitude of the excitation amplitude for plane wave incidence. Thus
in CTEM with normal incidence, the first four Bloch states in figure 3.1 account for
99% of the electron density. However, depending on the probe position, it may be
seen that in the STEM case these states may not be significantly excited. Similarly,
state 5, which being antisymmetric has zero excitation amplitude for normal plane
wave incidence, may be significantly excited for certain probe positions. It must
be stressed that all available antisymmetric states are excited depending on probe
position, whereas these remain unexcited for plane wave incidence on a crystal in
the symmetric zone axis orientation. Contributions from all states must therefore
be taken into account for the correct dispersion of the probe in real space with
increasing depth.
Relative magnitudes are shown below the images. State 1 may be regarded as
an s-state for the zinc column, following the parlance of Buxton et al. (1978), and
state 2 may be regarded as an s-state for the sulfur column (though in both cases
3
Of necessity for the δ-function probe, we have adopted for this equation, and for figure 3.1, a
different normalization to that of equation (2.38): we set c in equation (2.30) to unity.
62
0.380
0.424
0.468
0.669
0.000
Figure 3.1: Bloch state excitation amplitudes in ZnS, viewed along the [110] zone
axis, using 100 keV STEM probes. The rows are labelled with an identifying state
number, below which is given the magnitude of C0i∗ which would be the excitation
amplitude in the case of normal plane wave incidence. The first column corresponds
to the aberration-balanced probe, the second an aberration-free probe, and the third
a δ-function probe. Maximum and minimum values are given below the images to
provide a sense of scale.
faint contributions should be noted on the adjacent column site). In these states
especially, though also in the others, the variations in the excitation amplitudes
with probe position become more pronounced for the finer probes. The excitation
amplitudes suggest that the s-states incorporate most of the electron density when
the probe is situated upon the column. Much has been made of s-state models in
Chapter 3. Wavefunctions and elastic imaging 63
trying to find schemes which balance accuracy with ease of interpretation. In chapter
5 we will consider the decomposition of a signal into s-state contributions and show
that such a division does indeed prove useful for the case of the probe above the
column. However it is clear from figure 3.1 that for other probe positions it will
be necessary to include many more Bloch states for an adequate description of the
wavefunction and the signals arising from it (Allen et al., 2003b; Anstis et al., 2003).
Put another way, only for the case of plane wave incidence (in symmetrical zone axis
orientation) is it true that the wavefunction may be adequately represented by just
a few states near the top of the dispersion curve. In particular, the extent of the
range of Bloch states excited is critical in assessing the spreading of the wavefunction
about the probe location.
X
φi (r⊥ , z) = exp(2πiλi z) Cgi exp(2πig · r⊥ ) , (3.8)
g
and
X
αi (R) = [C −1 ]gi exp(−2πig · R)T (g) . (3.9)
g
Initially it may seem at odds with Bloch’s theorem to be using a mesh finer than
the physical reciprocal lattice for the Fourier representation of the crystal potential.
However, while it is the physical reciprocal lattice that determines the crystal repeat
distance, the incident probe no longer has the translational symmetry of the plane
wave case. The wavefunction in the crystal is not the same in adjacent cells up
to a phase (Bloch’s theorem), as this does not take the boundary conditions into
account. Rather, the intensity of the wavefunction should tend to decrease with
increasing distance from the probe position defined by the vector R, albeit not
64
As done for plane waves in section 2.3, the Bethe equations may again be written
in the form
AC = 2KC[λi ]D , (3.10)
where C is the matrix of eigenvectors Cgi and [λi ]D is the matrix containing the
eigenvalues, the subscript D indicating that it is diagonal in form. For STEM the
structure matrix A may be written
.. .. .. .. ..
. . . . .
··· 2 0
−h + iU0 Wh−g Wh Wh+g W2h ···
··· Wg−h −g2 + iU00 Wg W2g Wg+h ···
A=
··· W−h W−g iU00 Wg Wh ···
,
··· W−g−h W−2g W−g −g2 + iU00 W−g+h ···
··· W−2h W−h−g W−h W−h+g −h2 + iU00 ···
.. .. .. .. ..
. . . . .
(3.11)
which differs from equation (2.17) only in that the vectors used in the diagonal ele-
ments, and as subscript for the potential elements, are now based upon the reciprocal
lattice vectors of the supercell rather than of the physical lattice. For simplicity of
notation we will assume that the supercell results from an n × n tiling of the unit
cell, though the conclusions we will reach are equally valid for any rectangular tiling.
However, for crystalline specimens, the off-diagonal elements Wg−h are non zero
only if the vector difference g − h is a physical reciprocal lattice vector, i.e.
The presence of many zero elements in the A-matrix leads to further simplifications
in the description of the wavefunction within the crystal.4
Recalling that we may uniquely express g = G + ql and h = H + qm , where the
subscripts l and m describe elements of the set {q}, then the off-diagonal elements
of the A-matrix may be rewritten as
Because the supercell is n × n it follows that l and m run over n2 values. Let us
suppose that we include N significant physical reciprocal lattice vectors. The A-
matrix of equation (3.11) is thus of order n2 N . Using equation (3.13) in equation
(3.11), the A-matrix may be reordered in such a way that it is block diagonal since
the elements at the intersection of different groups, described by the subscripts l
and m, are all zero. The matrix equation becomes
[A(q1 )] [ 0 ] ··· [ 0 ] [C(q1 )] [ 0 ] ··· [ 0 ]
[ 0 ] [A(q2 )] ··· [ 0 ] [ 0 ] [C(q2 )] ··· [ 0 ]
..
. ..
.
··· ··· ··· ··· ··· ···
··· ··· ··· [A(qn )] ··· ··· ··· [C(qn )]
[C(q1 )] [ 0 ] ··· [ 0 ] [λ(q1 )]D [ 0 ] ··· [ 0 ]
[ 0 ] [C(q2 )] ··· [ 0 ] [ 0 ] [λ(q2 )]D · · · [ 0 ]
= 2K
..
..
,
··· ··· . ··· ··· ··· . ···
··· ··· ··· [C(qn )] ··· ··· ··· [λ(qn )]D
(3.14)
4
This simplification requires that the crystal be periodic. For imperfect crystals equation (3.12)
does not hold and so neither will the diagonalization property and the dividends in computational
efficiency which follow from it. Wavefunctions and cross sections can, of course, still be calculated
for such materials via the techniques in this thesis since they apply to non-periodic specimens
just as well as periodic ones: periodicity allows for simplifications which take advantage of the
symmetry, but does not change the physics used.
66
The superscript on the eigenvalues λk has been changed to emphasize the fact that
k now labels the N columns within the sub-matrix specified by the label l. Note
that the sub-matrix A(ql ) is equivalent to that describing plane wave incidence with
tangential component K⊥ = ql .
With the reordering used in equation (3.14) we now replace the index i in equa-
tion (3.10), which implicitly labels the columns of the complete C-matrix, with the
indices l, describing each sub-matrix C(ql ), and k, which denotes the columns of
each sub-matrix. Rewriting the equation for the total wavefunction, equation (3.7),
in terms of summations over l and k
2
n X
N
X
ψ(R, r⊥ , z) = αl,k (R)φl,k (r⊥ , z) , (3.16)
l=1 k=1
From the block diagonal form in equation (3.14) it immediately follows that Cgl,k = 0
for any reciprocal lattice vector g 6= G + ql [i.e. those elements not contained in the
sub-matrix C(ql )]. Thus the sums over g reduce to sums over G, giving
X l,k
φl,k (r⊥ , z) = exp(2πiλl,k z) CG+q l
exp[2πi (G + ql ) · r⊥ ] , (3.19)
G
and
X
αl,k (R) = [C −1 ]G+q
l,k
l
exp[−2πi (G + ql ) · R] T (G + ql ) . (3.20)
G
We now use the fact that sub-matrix A(ql ) is equivalent to that describing plane
wave incidence with tangential component K⊥ = ql . From equation (3.15) it is
clear that the sub-matrices C(ql ) and λ(ql ) must also contain the eigenvectors and
Chapter 3. Wavefunctions and elastic imaging 67
and
X
αl,k (R) = [C −1 (ql )]G
k exp[−2πi (G + ql ) · R] T (G + ql ) . (3.22)
G
The total wavefunction given by equation (3.16) is hence a sum over n2 independent
wavefunctions, each associated with a plane wave incidence, with the tangential
component of the plane wave being K⊥ = ql . Moreover, these wavefunctions do
not interact with one another. Thus, while the probe contains a continuum of
transverse momentum components, inside the crystal these components only interact
with others differing in value by physical reciprocal lattice vectors G.
In reference to figure 3.1 we discussed the Bloch states and Bloch wave amplitudes
for normal plane wave incidence, K⊥ = 0, and noted that the first five Bloch states
accounted for over 99% of the incident intensity in the example given there. Now
we see that not only may the significant intensity be spread over many more Bloch
states but the block diagonal form of the eigenvalue problem indicates there will
be n2 − 1 other such sets of significant Bloch wave amplitudes, which would not
normally be excited for plane wave incidence. In some cases these may be very
similar, to the extent that some can be joined together in a branch cluster (Bird,
1989; Pennycook and Jesson, 1991), but are formally distinct.
It will now be shown that the phase-linked plane wave approach is formally
equivalent to that described above, despite the conceptually different foundation.
As such, both techniques offer different insights into the physics of the problem.
It is worth noting that in calculation of convergent beam electron diffraction pat-
terns, multislice methods have been applied with both approaches to the boundary
conditions (Self and O’Keefe, 1988).
Equation (5) of Nellist and Pennycook (1999) expresses the total wavefunction
obtained in the phase-linked plane wave boundary conditions approach. That equa-
tion, re-drafted for consistency with the notation and sign conventions used here,
reads
Z XN
ϕ(R, r⊥ , z) = β k (R, p)φk (p, r⊥ , z)dp , (3.23)
k=1
where β k is the excitation amplitude of Bloch wave φk . Note that the superscript
k is used here, denoting that the sum is over the N columns of the plane wave C-
matrices. Bloch waves are resolved for each transverse component p of the focused
probe:
X
φk (p, r⊥ , z) = exp[2πiλk (p)z] k
CG (p) exp[2πi(p + G) · r⊥ ] . (3.24)
G
Replacing the integral in the total wavefunction equation (3.23) by a sum over
the nearly-continuous p and breaking it up into a sum over physical vectors H and
a sum over l labelling the vectors of the set {q} gives
2 N
n XX
X
ϕ(R, r⊥ , z) = β k (R, H + ql )φk (H + ql , r⊥ , z) , (3.26)
l=1 H k=1
where
φk (H + ql , r⊥ , z) = exp[2πiλk (H + ql )z]
X
k
× CG (H + ql ) exp[2πi(G + H + ql ) · r⊥ ] . (3.27)
G
Using the periodicity relationships of equation (2.11) in equation (3.27), the vector
H may be brought out of the argument and into the subscript. This results in an
Chapter 3. Wavefunctions and elastic imaging 69
expression in the sum over G of terms depending on G+H. This is a basis shift,
and we may re-label the sum as follows:
X
φk (H + ql , r⊥ , z) = exp[2πiλk (ql )z] k
CG+H (ql ) exp[2πi(G + H + ql ) · r⊥ ]
G
X
= exp[2πiλk (ql )z] k
CG (ql ) exp[2πi(G + ql ) · r⊥ ] . (3.28)
G
where again we have made use of the periodicity relations. Now note that the above
equation is identical to the block diagonal expression for the excitation amplitudes,
P
equation (3.22). So H β k (R, H + ql ) ≡ αl,k (R). Thus equations (3.26), (3.28)
and (3.29) describing the wavefunction are identical to equations (3.16), (3.21) and
(3.22). The wavefunctions in the two formulations are equivalent.
In obtaining the block diagonal form, it was shown that no connection exists
between wavefunctions described by plane waves with wavevectors which do not
differ by a physical reciprocal lattice vector, although it initially seemed implied by
that model. In working the phase-linked plane wave approach towards an equivalent
form it was shown, by subsuming the sum over H into the terms β k and φk , that a
meaningful physical connection exists between incident plane waves with wavevec-
tors differing by a physical reciprocal lattice vector, although it initially seemed that
this was not implied by that model. This is the insight gained through the compar-
ison of these models. Thus the useful physical picture and efficient coding method
(i.e. that taking advantage of the block diagonal property of the supercell structure
matrix) is a half-way point between the two formulations.
70
Replacing the sum over l by a sum over q, equations (3.16), (3.21) and (3.22) may
be rewritten as follows:
XX
ψ(R, r⊥ , z) = αk (R, q)φk (r⊥ , z, q) , (3.30)
q k
X
φk (r⊥ , z, q) = exp[2πiλk (q)z] k
CG (q) exp[2πi(G + q) · r⊥ ] , (3.31)
G
X
αk (R, q) = [C −1 (q)]H
k exp[−2πi(H + q) · R]T (G + q) . (3.32)
H
where in the last step the scattering matrix S has been identified [cf. equations
(2.23) and (2.24)].
As discussed in the previous section it is seen that the wave function divides into
non-interacting components associated with each transverse momentum component
q in the first Brillouin zone. Thus we may write
X
ψ(R, r⊥ , z) = ψq (R, r⊥ , z) , (3.34)
q
Chapter 3. Wavefunctions and elastic imaging 71
where
( )
X X
ψq (R, r⊥ , z) = SG,H (q, z) exp[−2πi(H + q) · R]T (H + q)
G H
× exp[2πi(G + q) · r⊥ ] . (3.35)
This layered equation has a simple, physical interpretation. The expression SG,H (q, z)
describes the scattering probability that a wave with incident momentum compo-
nent H + q has scattered to a wave with momentum component G + q at thickness
z. The incident probe comprises plane waves with incident momentum compo-
nent H + q weighted according to the lens and position by the complex factor
exp[−2πi(H + q) · R]T (H + q). Thus the total scattering from the incident probe
to a wave with momentum component G + q at thickness z is obtained by summing
the product of the probe amplitude and phasing factors and the scattering matrix
elements – this is the meaning of the expression in braces in equation (3.35). The
wavefunction is then composed by the Fourier transform of this momentum space
description, as per the phased sum over G in equation (3.35). Since there is no
interaction between the wavefunctions corresponding to different q values, equation
(3.34) expresses the total wavefunction as a sum over the component wavefunctions.
This intuitive reformulation will be useful in discussing coherent imaging and
general wavefunction properties.
A ronchigram (Saxton, 1978; Spence, 1999) is obtained from the intensity distribu-
tion |Ap (R)|2 as a function of p for fixed R. Consider a point detector located at
the origin (p = 0), and suppose that the probe-forming aperture is sufficiently large
that overlap between neighbouring fundamental reflections H occurs at the centre
of the zeroth order disk (Spence and Cowley, 1978). A coherent STEM lattice image
|A0 (R)|2 may be calculated as a function of probe position R. The zeroth beam
72
expression simplifies to
X
A0 (R) = S0,H (0, t) exp[−2πiH · R]T (H) . (3.37)
H
and it readily follows that |AG+q (R)|2 is independent of R, the condition for no
contrast. This conclusion applies to equation (3.37), which assumes a point detector
at the origin, provided that the aperture admits only the central physical reciprocal
lattice vector; in this case there may be some disc overlap but unless it occurs at
the site of the detector then the image will still contain no contrast.
3.5.2 Reciprocity
The Bloch wave formulation presented here for the STEM wavefunction in no way
makes use of assumptions of reciprocity in its derivation. However the result of
reciprocity is well established (Cowley, 1969; Wang, 1995; Kirkland, 1998); the new
technique must reproduce it. It will now be shown that the phenomenon of reci-
procity follows naturally from this formulation. Figure 3.2 shows schematically the
formation of a diffraction pattern in STEM and an image in CTEM.
Consider first the STEM case, taking the incident wave to originate from a point
source. The reciprocal (Fourier space) representation of the wavefunction at the
source, labelled (a) in figure 3.2, is unity
Ψa (p) = 1 . (3.39)
Transmission through the lens, in the reciprocal space notation, is obtained by mul-
tiplication with the transfer function of the lens, denoted T (p). We also introduce
the phase factor exp (−2πip · R) to account for the position of the centre of the
Chapter 3. Wavefunctions and elastic imaging 73
a c’
b b’
c a’
STEM CTEM
lens with respect to a chosen origin on the crystal surface [cf. equation (2.34)]. The
reciprocal space representation of the wavefunction at (b) is then
and the real space wavefunction is simply the Fourier transform thereof:
X
ψb (r) = T (p) exp (−2πip · R) exp (2πip · r) . (3.41)
p
Recalling that the S-matrix relates the incident and exit surface waves, the recip-
rocal space representation for the exit-surface wavefunction may be related to the
reciprocal space representation for the entrance-surface wavefunction at (b) given in
equation (3.40) via
X
Ψc (q) = Sq,p T (p) exp (−2πip · R)
p
" #
X X ¡ ¢ i∗
= Cqi i
exp 2πiλ t Cp T (p) exp (−2πip · R) . (3.42)
p i
74
Note that we have written Cpi∗ rather than [C −1 ]pi . We are neglecting absorption.
This assumption is necessary here since absorption removes the idealized result of
reciprocity [though see Pogany and Turner (1968) for a discussion of approximate
reciprocity including inelastic scattering].
The bright field STEM amplitude is simply the zeroth component of the above
expression
" #
X X ¡ ¢
A0 STEM = Ψc (0) = C0i exp 2πiλi t CH
i∗
T (H) exp (−2πiH · R) , (3.43)
H i
where we have used the block diagonalization result to revert to using the physical
reciprocal lattice vector notation now that the value of q has been fixed.
Now consider CTEM. The above treatment followed the standard convention
that the z-direction be identified with the incident beam direction – down in figure
3.2 for STEM. For the following treatment of CTEM we will break that; instead we
will retain the z-direction as down, despite the incident beam direction being up in
figure 3.2. Thus we must propagate backwards. In doing so we will be guided by
the rationale that the operations performed should, if applied to the STEM case
described above, be those which undo the operations performed there.
With this in mind, propagating backwards through the crystal is accomplished
in a matrix sense by multiplying by S −1 . Since we are neglecting absorption, the
S-matrix is unitary: S −1 ≡ S † .
Careful consideration of the matrices then gives
£ ¤
S = C exp(2πiλi t) D C †
£ ¤
⇒ S −1 = S † = C exp(−2πiλi t) D C † , (3.44)
†
X ¡ ¢
Sg,h ≡ Cgi exp −2πiλi t Chi∗ . (3.45)
i
In CTEM, as shown in figure 3.2, the incident plane wave at (a’) is given by the
reciprocal space representation
The reciprocal space representation for the wavefunction at (b’) may then be ob-
†
tained by interpreting Sg,h as the transition amplitude for incident plane wave la-
belled h to the exit plane wave labelled g. So
X † †
Ψb0 (h) = Sh,p δp,0 = Sh,0
p
X ¡ ¢
i
Ψb0 (H) = CH exp −2πiλi t C0i∗ , (3.47)
i
where we have changed notation h to H to reflect that in the plane wave case the
exit-surface wave comprises only plane waves related to the incident plane wave by
tilts of physical reciprocal lattice vectors.
As in STEM, the positioning of the lens means that the reciprocal space repre-
sentation of the image results from multiplying the reciprocal space wavefunction
at (b’) with the complex conjugate of the transfer function (corresponding to prop-
agating backwards through the lens)
X ¡ ¢
i
Ψc0 (H) = CH exp −2πiλi t C0i∗ T ∗ (H) . (3.48)
i
Finally, the real space wavefunction in the image plane at (c’) is constructed by
inverse Fourier transform, giving
" #
X X ¡ ¢
i
ATEM (R) = ψc0 (R) = CH exp −2πiλi t C0i∗ T ∗ (H) exp (2πiH · R) .
H i
(3.49)
Note that R has now been used to describe the position in the image. This usage
is different to that in the STEM discussion. However this notational subtlety has
been tolerated because it aids a comparison between equations (3.43) and (3.49).
Inspection shows that they are the complex conjugates of one another, in particular
they have the same magnitude, and therefore intensity. Thus the reciprocity theorem
for the equivalence of CTEM and STEM as appearing in the literature (Cowley,
1969; Wang, 1995; Kirkland, 1998) has been demonstrated from the Bloch wave
expression for the wavefunction.
The finite size of the source means that different points on the tip can emit inco-
herently. The lens, either by faithful reproduction of the range of incident energies
occasioned by the energy spread from the tip or through its own current instabili-
ties, provides another form of incoherence. Other effects which do not correspond
to incoherently related wavefunctions but produce sufficiently similar effects in the
images that they come under the umbrella term of incoherence in the imaging sys-
tem include instability in the position of the probe and the recording properties of
the detector.
From equation (3.36) we may write the intensity in the diffraction pattern at
point G + q for a given probe position R as
X
∗ 0 ∗ 0
I(G + q, R) = SG,H (q, z)SG,H0 (q, z) exp[−2πi(H − H ) · R]T (H + q)T (H + q) .
H,H0
(3.50)
The effect of finite tip size is to lead to a set of incoherent probes for slight
displacements ∆R about the probe position R. The effect of position instability
of the probe or sample also leads to a spread about some median R value. These
effects may be incorporated by convolving the bright field intensity in equation
(3.50) with a suitable distribution function in R space. The effect of detector size
is also a convolution, but leading to an integration over diffraction points G + q
corresponding to the detector pixel size. A more detailed description of these effects
is given by Nellist and Rodenburg (1994).
The only dependence on defocus in equation (3.50) is through the aberration function
in the lens transfer function [cf. equation (2.29)]. Thus we seek to evaluate
Z ∞ ½ ¾
2π 0
J≡ p(F ) exp i [χ(H , ∆f + F ) − χ(H, ∆f + F )] dF . (3.52)
−∞ λ
Let us assume a Gaussian form for the defocus spread, with a 1/e width of ∆, such
that
1 F2
p(F ) = √ e− ∆2 . (3.53)
∆ π
Chapter 3. Wavefunctions and elastic imaging 77
Using equation (3.53) in equation (3.52), and using equation (2.32) for the form of
the aberration function, it may be shown that
½ ¾
© ª 2π
J = exp −π 2 λ2 ∆2 (H2 − H02 )2 /4 exp i [χ(H0 , ∆f ) − χ(H, ∆f )] , (3.54)
λ
and thus
X
∗ 0
I(G + q, R) = SG,H (q, z)SG,H0 (q, z) exp[−2πi(H − H ) · R]
H,H0
2 λ2 ∆2 (H2 −H02 )2 /4
×T (H + q)T ∗ (H0 + q)e−π , (3.55)
which is the standard result, familiar from both STEM and, because of the similari-
ties already noted in discussing reciprocity, CTEM. Note that unlike equation (3.50),
equation (3.55) cannot be written as the modulus squared of a wavefunction because
the final exponential term, the damping factor resulting from incoherent effects, does
not factorize neatly. This increases the computational complexity of evaluating the
diffraction pattern since it becomes necessary to perform two summations instead of
one. In some cases the problem may be simplified by the approximate factorization
(Coene et al., 1996)
2
e−π
2 λ2 ∆ 2
(H2 −H02 ) /4
' e−π
2 λ2 ∆2 H4 /4
e−π
2 λ2 ∆2 H04 /4
. (3.56)
The majority of coherent imaging in STEM is done via the diffraction pattern.
It is only very recently that microscopes with optics capable of atomic resolution
both pre- and post-specimen have come on-line and offer the possibility of coherently
imaging wavefunctions formed with a focused incident probe (Hutchison et al., 2005).
We will restrict the current discussion to techniques which utilize diffraction pattern
intensities.
When the phase object approximation is valid and the probe wavefunction is
known, determining the transmission function for the specimen is accomplished by
dividing the exit-surface wavefunction obtained via phase retrieval by the probe
wavefunction. The possibility of implementing such an approach is a recent devel-
opment. It utilizes improved methods of diffraction pattern phase retrieval based on
over-sampling in the diffraction plane and effective compact support on the wave-
function. This support might be real, using a post-specimen aperture (Faulkner and
Rodenburg, 2004; McBride et al., 2004), or effective, using the probe intensity itself
as the constraint (Rodenburg and Faulkner, 2004).
Suppose that the facility is available to record ronchigrams, that is to say the
complete diffraction pattern. For simplicity let us extract from this diffraction pat-
tern those points in the set {F}, i.e. those corresponding to q = 0, where we are
assuming that the angular calibration of the detector in the diffraction pattern has
Chapter 3. Wavefunctions and elastic imaging 79
Now suppose that we record such a diffraction pattern for a sufficient number of
probe positions that we may calculate the Fourier transform of equation (3.58) over
probe position. Thus we obtain
© ª X
FR→p |Ψ(R, G, t)|2 = ∗
SG,H (t)SG,H+p (t)T (H)T ∗ (H + p) . (3.59)
H
Let us consider a special case of equation (3.59), that of the phase object approxima-
tion. The phase object approximation is usually written in a real space formulation;
in the multislice method it amount to neglecting the propagator. However, it is
obtained as the solution to the Schrödinger equation in the limit where the kinetic
energy term is neglected. In the A-matrix formulation this is equivalent to setting
all the diagonal elements to zero. This has the consequence that AG,H = AG−H,0 .
It is easy to show that the product of matrices with this property also possesses this
property, and thus in the phase object approximation SG,H = SG−H,0 ≡ SG−H . In
this approximation, equation (3.59) may be written
© ª X
FR→p |Ψ(R, G, t)|2 ' ∗
SG−H (t)SG−H−p (t)T (H)T ∗ (H + p)
H
≡ D(G, p) ⊗G GA (G, p) , (3.60)
In its strictest sense, the phase object approximation has a fairly limited range
of validity, though in CTEM the qualitative resemblance between the phase and the
potential persists beyond the validity domain of the phase object approximation.
What might be done to effect inversion beyond the phase object approximation?
In CTEM techniques which extend somewhat beyond this range exist (Lentzen and
Urban, 1996, 2000; O’Leary and Allen, 2005), and might be adapted to STEM. These
techniques apply to a single exit-surface wavefunction in CTEM, which is effectively
a single column of the S-matrix. A formal inversion method exists for the case that
the entire S-matrix is known (Allen et al., 1999, 2000). This approach has some
very attractive features: it is linear and it is not restricted to thin specimens, indeed
it does not require the thickness to be known. In CTEM, determining the entire
S-matrix requires a series of experiments to be performed for different tilts of the
incident plane wave. The technique has not been realized because performing this
finely controlled series of tilts is experimentally very challenging. In this regard
STEM looks very attractive, because the convergent probe comprises precisely a set
of tilted plane waves.
We therefore seek a method which will enable the extraction of the S-matrix ele-
ments from measurement of diffraction patterns in STEM. However, unlike equation
(3.60) for the phase object approximation, the disentanglement of the general scat-
tering matrix from the probe amplitudes in equation (3.59) is less straight-forward.
Some explicit form of phase retrieval will be necessary.
Chapter 3. Wavefunctions and elastic imaging 81
where the dependence of the lens transfer function on the defocus of the probe-
forming lens has been explicitly emphasized since its variation will be the basis of the
phase retrieval method used. Note that this intensity is effectively four dimensional.
For each physical probe position R in a two dimensional scan we get a physical
diffraction pattern with points labelled by the two dimensional reciprocal lattice
vector G.
In the previous paragraph we emphasized the word “physical”. The experiment
dictates an obvious interpretation of I∆f (R, G, t): one should view R as a parameter
and G as the variable describing the image. In principle the machine could be shut
down and restarted between recording such a data set and recording that for the
next position R: there is no dynamical connection between these measurements.
However, in a formal, mathematical sense, equation (3.63) possesses another
interpretation. We might instead treat G as a parameter and R as the variable;
we could form intensity images as a function of R for a fixed value of G. This is
not new. Taking G = 0 and forming an image as a function of R is known as a
bright field image, and we have already considered this in previous sections when
discussing reciprocity and coherent imaging.
The dividend of thinking about equation (3.63) in this way is that it motivates
an alternative option for phase retrieval. The obvious option is to perform phase
retrieval on a physical wavefunction, on the images or diffraction patterns formed for
a fixed value of R. But, if we recorded several images at different values of defocus,
we may use phase retrieval on equation (3.63) taking images as a function of R for
a fixed value of G. To elucidate this point let us rewrite equation (3.63) as
¯ ¯2
¯X ¯
¯ ∗ ∗ 2πiH·R ¯ 2
I∆f ,G (R, t) = ¯ SG,H (t)T∆f (H)e ¯ ≡ |Φ∆f,G (R)| , (3.64)
¯ ¯
H
where
∗ ∗
FR→H {Φ∆f,G (R, t)} ≡ SG,H (t)T∆f (H) . (3.65)
82
Having retrieved the phase, shifting back to the original basis is straightforward.5
Thus for a fixed value of G a through-focal series technique (Faulkner, 2003;
Allen et al., 2004a) may be used to retrieve the reciprocal space “wavefunction”
∗ ∗
SG,H (t)T∆f (H) which may then be aberration corrected to yield SG,H (t). This then
gives, correctly phased, the elements within one row of the scattering matrix (cf.
within a column, as usually obtained with the conventional transmission electron
microscope). Repeating the procedure for different values of G determines the full
scattering matrix. The different rows are not automatically phased correctly, but, at
least for K⊥ = 0, this may be fixed via the anti-diagonal symmetry of the scattering
matrix (Allen et al., 2000; Findlay, 2005).
The inversion method of Allen and co-workers (Allen et al., 1999, 2000), who
showed that if the entire scattering matrix is obtained then the inversion problem
may be reduced to a linear problem which is therefore unique and well defined,
may now be applied. The approach is based on the appreciation that the scattering
matrix and the structure matrix have the same eigenvector spectrum [cf. equations
(2.19) and (2.24)]. So the eigenvectors of the structure matrix are obtained immedi-
ately on finding the eigenvectors of the scattering matrix. Obtaining the eigenvalues
of the structure matrix from those of the scattering matrix is hampered because there
is a uniqueness problem in taking the complex logarithm [cf. equation (2.24)]. In-
stead the eigenvectors may be used together with the known diagonal elements and
symmetries of the structure matrix to generate from equation (2.19) a set of linear
5
That such a procedure should improve the stability, as gauged by the success of the retrieval
or the rate of convergence, is not obvious. Moreover it is not general. It seems to result from
∗
the form of SG,H (t), which is approximately symmetric about H = G and not H = 0. It is not
necessary in practice to use only one phase retrieval routine for all through-focal series. Rather
one could subject a problematic series to a range of retrieval methods until success is achieved.
The approach taken here should not be regarded as the “best” method, but simply as what worked
well with our data set.
Chapter 3. Wavefunctions and elastic imaging 83
equations satisfied by the eigenvalues (Allen et al., 2000). The diagonal terms from
equation (2.19) provide inhomogeneous equations of the form
N
X £ ¤
2K Cni λi C −1 in = Ann = P (K⊥ = 0, Gn ) = − (Gn )2 + iU00 , (3.67)
i=1
where the final identification is drawn from equation (2.17) and it should be re-
membered that, as per the discussion above, we have chosen to work with the set
K⊥ = 0. It is unlikely that U00 will be known a priori and so we will simply neglect
it. Doing so is known not to significantly affect the accuracy of the reconstruction
(Allen et al., 2000). The symmetries off the diagonal lead to homogenous equations.
So supposing that Amn = Apq then it follows from equation (2.19) that
N ³
X £ ¤ £ ¤ ´
Cmi C −1 in − Cpi C −1 iq λi = 0 . (3.68)
i=1
It is interesting to note that this approach does not require the thickness to be
known, since it appears only with the unused eigenvalues of the scattering matrix.
Thus a complete knowledge of the scattering matrix allows the eigenvector matrix
to be determined,6 and so the structure matrix to be reconstructed. The Fourier
coefficients of potential are then read off.
In summary, we have a method of determining the entire scattering matrix, and
therefore carrying out the inversion procedure, within the scanning transmission
electron microscope. The requisite data set is a set of diffraction patterns for dif-
ferent positions of the probe upon the surface and for a range of defocus values.
This data set is no less daunting than was required in CTEM. Indeed it is larger
than the formidable set described by Rodenburg and Bates (1992), since that only
required a single defocus value. However this approach is not restricted to the phase
object approximation and does not require modification of the geometry of existing
microscopes.
3.6.3 Truncation
The inversion procedure, which requires the determination of eigenvectors, must
be applied to a square matrix. In principle images may be collected in the form
6
In practice the “absolute phase” and “absolute magnitude” of the scattering matrix, i.e. relative
to the incident beam, cannot easily be determined experimentally. Note however that this does
not prevent accurate determination of the eigenvector matrix, since S(t) and αS(t) have the same
eigenvectors, for any complex number α .
84
of equation (3.64) out to very large values of G, but in performing the aberration
correction step it becomes clear that only those values of SG,H for which T∆f (H) 6= 0,
for which H fits in the probe-forming aperture, may be obtained. The largest
square scattering matrix is limited to this resolution. As per equation (2.24), the
scattering matrix is formed from the matrix exponential of the structure matrix. The
structure matrix is formally of infinite order, and therefore so too is the scattering
matrix. Numerical modelling requires treating finite matrices, but it may prove
that adequate convergence requires them to be of large order. By performing the
inversion on a truncated scattering matrix we treat a model problem with scattering
and structure matrices of relatively small order. The results will only be valid if the
elements in the retrieved, finite structure matrix are good approximations to those
of the full structure matrix. This will be the case if the scattering matrix formed
by the truncated structure matrix almost converges to the truncated portion of the
scattering matrix as calculated from the full structure matrix. If this is not the case
then aliasing effects serve to distort many of the Fourier coefficients of potential
retrieved.
This problem, called the truncation problem, has been discussed for CTEM in the
systematic row case by Allen and Oxley (2001). The systematic row case converges
much faster than the general zone axis case, but the qualitative conclusions drawn
remain the same. It was found that the aliasing acts from the off-diagonal corners
of the structure matrix inwards, affecting the Fourier coefficients of the potential,
and the problem becomes more significant for more severe truncations. Truncation
tends to significantly over-estimate the magnitude of higher order Fourier coefficients
and so, with the expectation that such Fourier coefficients should tend to decrease
with increasing spatial frequency, it may be possible to identify suspect Fourier
coefficients of the retrieved potential by inspection. Truncation will prove to be the
most severe limitation on the inversion procedure in practice and will be discussed
further subsequent to presenting a case study.
et al. (2004a,b) is another possibility, which may handle noise and partial coherence
more rigorously, though we will not pursue this more elaborate course here.
While incoherent imaging in STEM seems to be fairly robust in the presence
of partial coherence (Nellist and Pennycook, 1998b), the coherent imaging mode is
affected by it. It may readily be shown from equation (3.64) that the effect on the
intensity images of the spread in defocus values is identical to its effect in CTEM
(Allen et al., 2004a). Specifically, averaging equation (3.64) over defocus using a
Gaussian distribution with 1/e value ∆ centred at ∆f gives (cf. section 3.5.3)
X
∗ ∗
I∆f ,G (R, t) = SG,H (t)SG,H0 (t)T∆f (H)T∆f (H0 )
H,H0
2
0
e2πi(H−H )·R e−π
2 λ2 ∆2
(H2 −H02 ) /4
. (3.69)
As a consequence of the cross terms in the final exponential, this result is no longer
exactly expressible as the modulus-squared of a single wavefunction. We will sim-
ulate using the full form above, but when attempting to correct for the spread in
defocus we will make the factorization approximation of equation (3.56), and de-
convolve to remove the resultant envelope function (Coene et al., 1996; Allen et al.,
2004a).
The effect of the finite source size is somewhat different in STEM, being rather
equivalent to the effect of finite detector pixel size in CTEM (Nellist and Rodenburg,
1994). This can be described as a convolution over probe position of the image with
a distribution resulting from the finite source size. Provided it can be quantified,
this issue can be corrected by deconvolution applied to the images prior to phase
retrieval and as such we will not include this effect.7
There is an issue relating to the ability to determine the precise locations of the
diffraction beams G: the possibility of some systematic offset, measuring beams
G + q rather than G. In this case one would record an image
¯ ¯2
¯X ¯
¯ ∗ ∗ ¯
I∆f,G+q (R) = ¯ SG+q,H+q (t)T∆f (H + q)e2πi(H+q)·R ¯ . (3.70)
¯ ¯
H
The scattering matrix SG+q,H+q (t) is equivalent to SG,H (q, t) as seen previously [cf.
equation (3.33)] and is formed from equation (2.24) but using a structure matrix with
diagonal entries shifted as per G → G + q. This does not affect the symmetries off
the diagonal of the structure matrix. But it does mean the assumption of its diagonal
7
This is not strictly permissible when noise is present but we will not explore this issue here.
86
sim
where SG,H (t) denotes an element of the simulated scattering matrix using the
exp
retrieved potential and guessed thickness t while SG,H denotes the corresponding
scattering matrix element of the experimental data. Both quantities have been
normalized by the central scattering matrix element because the absolute magnitude
and phase of the experimental data is likely unknown. This was not a problem for the
inversion method, since multiplicative constants do not affect the eigenvectors, only
the unused eigenvalues. Minimization and non-linear equation solving approaches
to the inversion problem must also handle this unknown complex multiplicative
constant, and further discussion may be found in the literature (Lentzen and Urban,
1996; O’Leary and Allen, 2005).
As a prescription for limiting the effects of aliasing resulting from truncation we
will only take from the retrieval those Fourier coefficients of potential which lie in
Chapter 3. Wavefunctions and elastic imaging 87
sim
the central column of the structure matrix. For the purposes of forming SG,H (t) we
will therefore make a scattering matrix from these Fourier coefficients only; elements
which are not by symmetry identical to any in the central column are set to zero.
We will also force the known diagonal elements but set the mean absorption U00 to
zero since the normalization will eliminate it anyway.
(a) (b)
120 120
l (eV)
l (eV)
80 80
Potentia
Potentia
40 40
5 5
0 4 0 4
Å)
)
(Å
3 3
(
-40 -40
ion
ion
2 2
3 3
sit
1
sit
1
Y po 2 Y po 2
po
po
1 0 sitio 1 0
sitio 0 n (Å 0
X
n (Å
X
) )
(e) (f)
8 6
Reference Reference
4
Magnitude (eV)
Magnitude (eV)
6 Retrieved Retrieved
2
4
0
2
-2
0 -4
-2 -6
0 20 40 60 0 20 40 60
0.12
0.08
0.04
Ga As 0.00
50 100 150 200 250
t (Å)
Figure 3.3: (a) Intensity of exit surface wavefunction for 152 Å thick GaAs in [110]
orientation. (b) Phase of same exit surface wavefunction. (c) True potential using
first 69 Fourier coefficients. (d) Potential retrieved by 69 beam truncation of a 261
beam scattering matrix, including 1% noise and temporal incoherence. (e) Real
component of Fourier coefficients. (f) Imaginary component of Fourier coefficients.
(g) Sketch of projected GaAs structure. (h) F(t) as defined by equation (3.71).
Chapter 3. Wavefunctions and elastic imaging 89
structure matrix. The retrieved potential is in fair agreement with the reference
structure, though the retrieved potential erroneously has the arsenic columns less
pronounced than the gallium columns. Figures 3.3 (e) and (f) show, for the real and
imaginary part of the Fourier coefficients of the potential, the comparison between
the input and retrieved data. The effects of noise are manifest in distinguishing
between coefficients which in the perfect structure would be identical. Figure 3.3
(h) shows F (t) as defined by equation (3.71), and the global minimum at a thickness
of 150 Å is in excellent agreement with the true thickness of 152 Å.
3.6.7 Discussion
We suggested earlier that the greatest limitation on the method was the error intro-
duced by the truncation. In the Bloch wave method, the number of beams is selected
according to some convergence criterion. For plane wave incidence the number of
beams should be large enough that the central column, which contains the Fourier
coefficients of the exit surface wavefunction, is converged. The inversion procedure
requires not a single column but a central sub-matrix. Simulating a 69 beam matrix
then applying the inversion to that matrix is a self-consistent problem. But the con-
verged calculation would require simulating for a matrix sufficiently large that the
central 69-by-69 beam sub-matrix is essentially converged. This was approximated
in the previous section by truncating a 261 beam matrix. Such a truncated problem
would only be exactly consistent if the sub-matrix appeared in a block diagonal
structure, since only in such a case would the selected elements be independent of
the omitted ones. However this will never quite happen in practice, because the rep-
etition of elements in the structure matrix ensures that low order Fourier coefficients
recur in structure matrices of all orders.
That truncation is unavoidable begs the question of whether the inversion is
stable. Unfortunately this is not guaranteed. The advantages garnered by converting
the problem into a set of linear equations for the unknown eigenvalues come through
the assumption that the eigenvectors of the scattering matrix are identical to those
of the structure matrix. The inversion is therefore unavoidably limited by how well
the eigenvectors of the scattering matrix, subsequent to truncation and containing
the errors incurred by noise and incoherent effects, match those of the structure
matrix.
To demonstrate this possible instability in the inversion, we will simulate the
inversion step for a 300 Å thick sample of ZnS in the [110] orientation using 300 keV
electrons. The different sample and parameters were chosen to illustrate the insta-
90
bility, which does not occur in the case study of the previous section. The scattering
matrix was simulated using 459 beams. Figure 3.4 (a) shows the true projected po-
tential using 103 Fourier coefficients (corresponding to a 36 mrad aperture) on the
central column of the structure matrix. Figure 3.4 (b) shows the potential retrieved
by inversion of the scattering matrix truncated to 103 beams. Figures 3.4 (c) and
(d) show similar results for 63 beams (corresponding to a 26 mrad aperture). In a re-
trieval from experiment, one would not have the true result available for comparison.
With this in mind it is clear that qualitative interpretations made from the 63 beam
case would be more sound than those made from the 103 beam case. The retrieval
from 103 beams contains significant fluctuations and little differentiation between
the two column types. The 63 beam inversion, while notably over-estimating the
magnitude of the potential, is in good qualitative agreement with the true structure.
(a) (b)
160 160
V)
V)
120
Potential (e
120
Potential (e
80 80
40 40
0 5 0 5
4 4
)
)
(Å
(Å
-40 3 -40 3
n
n
3 2 3 2
io
io
2 1 2 1 sit
si t
1 1
po
Yp 0
po
Yp 0
osit
osit 0 0
X
X
ion i on
(Å) (Å)
(c) (d)
V)
V)
160 160
Potential (e
Potential (e
120 120
80 80
40 40
0 5 0 5
4 -40 4
)
-40
)
(Å
(Å
3 3
n
n
3 2 3 2
io
io
2 1 2 1
sit
sit
1 1
po
Yp 0
po
Yp 0
osit
osit 0 0
X
X
i on i on
(Å) (Å)
Figure 3.4: (a) True potential for ZnS using first 103 Fourier coefficients. (b) Po-
tential retrieved by 103 beam truncation of a 459 beam scattering matrix. (c) True
potential using first 63 Fourier coefficients. (d) Potential retrieved by 63 beam trun-
cation of a 459 beam scattering matrix. Using a 300 Å thick sample of ZnS, viewed
along the [110] zone axis, with 300 keV electrons.
Chapter 3. Wavefunctions and elastic imaging 91
Strict rules for when one should expect an inversion to succeed or when it might
struggle have not been established. One indicator is the spread of significant mag-
nitudes of scattering matrix elements away from the diagonal. For a sub-matrix to
behave as though it came from a block diagonal matrix, the intensity near to top left
and bottom right corners of the matrix must be concentrated near the diagonal. If
this is so, the elements discarded by the truncation are small. The author has found
cases for a Si3 N4 simulated inversion in which there was significant spread away
from the diagonal axis in the scattering matrix that showed unstable oscillatory be-
haviour persisting to sub-matrices of orders much larger than those corresponding
to realistic objective aperture sizes. That said, a variety of samples and thicknesses
were found to be amenable to this inversion procedure as was a tendency for the
inversion to improve, though not monotonically, with increased aperture size.
Thus far we have determined a fair estimate for a range of lower order Fourier
coefficients (range depending on aperture size) and also for the thickness. The first
of these stages was deterministic. The inversion procedure produces a unique result,
though this does not give an assurance of high precision in the retrieved structure
because of the effects of truncation. The second stage was based on minimization
for a single variable, but in all our simulations this minimum was well defined. The
quantity of experimental data suggests that we might now switch to an optimiza-
tion procedure to refine our values, thus improving the good qualitative agreement
of figure 3.4 (d) to good quantitative agreement, and even extend the resolution
(i.e. obtain some potential coefficients off the truncated central column). Previous
optimization approaches were limited because they required a reasonable starting
guess which was obtained either by a single scattering approach or by a channelling
approach, both of which have a thickness-limited domain of validity (Lentzen and
Urban, 1996, 2000; O’Leary and Allen, 2005). The result of our current method
could thus be combined with the experimental data and the techniques of previous
authors to extend the precision of the retrieved potential.
The inversion problem will not be pursued further here. The tendency in STEM
is to use incoherent imaging mechanisms on the basis that the structure determi-
nation is almost a matter of inspection. However the information contained in the
wavefunction for STEM is tantalizing, and such an approach would represent a
considerable advance in quantification over visual interpretation methods.
92
3.7 Conclusion
In this chapter a new formulation of STEM which uses continuity of the wavefunc-
tion across the boundary in a single step has been described. The mathematical
consequences of this were that the Bloch state excitation amplitudes become explic-
itly dependent upon the position of the probe and the coherent aberrations of the
lens. By plotting the excitation amplitudes as a function of probe position we saw
that many more states may be required for an adequate description of the STEM
wavefunction than were necessary in CTEM calculations.
This approach of treating the wavefunction as a single entity is, of necessity,
equivalent to the phase-linked plane waves method upon which previous Bloch wave
treatments of STEM were based. The current formulation has some attractive fea-
tures for analytical work, such as the ready demonstration of the reciprocity relation
between bright field STEM and CTEM and of the absence of contrast in the absence
of disk overlap within the detector.
In the case of crystalline specimens we saw that the structure matrix could be
block diagonalized, allowing for significantly faster calculations of the wavefunction
relative to the nonperiodic case. Beams which differ by a physical reciprocal lattice
vector are connected and interfere with one another. Beams which do not differ by a
physical reciprocal lattice vector are not so connected and cannot interfere with one
another. This understanding both simplifies subsequent calculations and represents
a basic concept in the physics of STEM imaging.
Using simulations which include realistic constraints such as counting noise and
temporal incoherence, we explored a structure retrieval method which uses diffrac-
tion pattern data collected using the scanning transmission electron microscope in
coherent imaging mode to reconstruct the structure matrix. Though requiring the
collection of a large data set, the method can handle multiple scattering and does
not require that the specimen thickness be known – indeed it is able to provide
an excellent estimate for the thickness after the fact. This approach removes the
need to take data at carefully controlled incident orientations which has thus far
prevented the inversion method being applied in the CTEM geometry.
Though such an inversion procedure as described here offers much in the way
of quantitative structure analysis, it is not the key result of this thesis. It proves
to have several limitations, most notably the large size of requisite data and its
inherent instability. In the next chapter we will turn our attention to incoherent
imaging modes which offer more promise for analysing a much broader range of
specimen properties.
¤£¡
ª@? ¡¢
R
R?
@ ª
¡
Chapter 4
©©
H
¼ ª?
¡ RH
@ j
93
94
where P (R, r) is a quantity related closely to the incident probe while the object
function O(r) should be a property of the crystal (save perhaps that it may depend
on the energy of the incident electrons). We will see presently that inelastic images
are always incoherent images by this definition if the effective scattering potential
is local. The term incoherent image is often used in a restricted sense, namely for
images that may be described by the form
Z
I(R) = |P (r⊥ − R)|2 O(r⊥ )dr⊥ , (4.2)
The adverbs used to qualify the description of equation (4.1) warrant further
scrutiny. By “usefully” we mean that for the incoherent imaging conceptualization
to offer insight into the images formed, the quantities P (R, r) and O(r) should
both have a physical meaning of some generality. For instance, if we assume that
|P (r⊥ − R)|2 is the probe intensity on the surface then from a given experimen-
tally obtained image I(R) it would be possible to determine the function O(r⊥ ) of
equation (4.2) by deconvolution. But in order to regard the result as a property
of the specimen, and the nature and geometry of the detector, this same function
O(r⊥ ) should be obtained for a wide range of probe functions |P (r⊥ − R)|2 and the
correspondingly measured images I(R). If this is not the case then the incoherent
imaging assumption may not be valid. It could, of course, be that the choice of
interpreting |P (r⊥ − R)|2 as the probe intensity is poor, but unless other reason-
able interpretations are available then this simple incoherent imaging model ceases
to have any predictive or explanatory power. By “adequately” we accept that the
result need not be exact. For instance, it will be generally assumed that the incident
probe intensity is approximately axially symmetric with a single central peak which
tends to drop off, though not strictly monotonically, with increasing radial distance.
96
If equation (4.2) held for such a class of common probe distributions, even if it did
not hold for more exotic intensity distributions, then the simple incoherent imaging
interpretation would still be useful.
Incoherent imaging is a desirable property because it often justifies the direct
interpretation of images. However whether a particular experimental image is well-
described by the assumption of incoherent imaging should be determined on the-
oretical grounds. Thus the term will be used in this discourse only as something
to be established by more rigourous methods, something to be demonstrated and
justified rather than simply assumed.
It has been shown that elastic images formed using high-angle annular detectors
can be justifiably interpreted as incoherent images (Nellist and Pennycook, 1999;
Peng et al., 2004). Previous authors have demonstrated, and the explorations in
this chapter will again demonstrate, that inelastic images may often be justifiably
interpreted as incoherent images, at least for thin specimens and local scattering
potentials. However this is not always the case, and later chapters will present some
important examples of inelastic images for which the incoherent imaging assumption
is poor.
so that the probability density over distance t along the optical axis is unity. The
numerator can be evaluated from the Yoshioka equations (Yoshioka, 1957) using
conservation of electrons and, with slightly different multiplying factors, the result
is formally a cross section (Allen and Josefsson, 1995). Because of this historical
link we will sometimes refer to the set of expressions derived from this form as
cross-section expressions, though the units of all calculations performed and figures
presented will be that of the fractional intensity.
Equation (4.3) may be evaluated using the Bloch wave representation for STEM
in a manner entirely analogous to the CTEM case considered by Allen and Josefsson
(1995), differing only in that the Bloch state excitation amplitudes are now those
pertinent for STEM and so depend on probe position and the coherent aberrations
of the probe. The resulting fractional intensity, denoted I(R, t) to emphasize its
dependence on probe position and sample thickness, is written
X X
I(R, t) = t B ij (R, t) Cgi Chj∗ µh,g . (4.4)
i,j g,h
The only difference between equation (4.4) and the plane wave case is in the explicit
R-dependence of the interference term B ij (R, t) via the excitation amplitudes:
So much for the dynamical term. What, though, of the diffuse term derived by
Allen and Josefsson (1995)? If the detected signal being considered originates from
thermal scattering then a diffuse term is seldom necessary, although its inclusion
was important in the back-scattering simulations of Rossouw et al. (1994). However
if we are detecting, say, energy loss corresponding to inner-shell ionization then the
diffuse term is relevant: electrons which have undergone thermal scattering are still
in the crystal and may cause ionization events. In principle one could form further
contributions, for instance from those electrons which had previously undergone
plasmon scattering. However, for most crystals considered in high-resolution imag-
ing, it is generally accepted that thermal diffuse scattering is the most significant
inelastic scattering mechanism for producing background electrons.
In the cases to which the incident plane wave formulation was applied the dif-
fuse term was a significant and important contribution, necessary to obtain good
agreement with experimental data. However, the diffuse background contribution
from dechannelled or absorbed electrons as derived in that original work is not well-
suited to the case of fine probes in STEM. We will return to the issue of the diffuse
background and how it should be adapted for STEM in chapter 6, but will neglect
it for the present. This will prove not to be a limiting assumption for the majority
of cases we will consider, since the diffuse contribution is often very small.
Equations (4.4) and (4.5) provide a framework for calculating the image arising
from any incoherent scattering process under highly dynamical diffraction condi-
tions. We will be particularly interested in HAADF images, formed with electrons
thermally scattered to high angles, and EELS images, obtained from electrons with
an energy loss characteristic of inner-shell ionization. The rewriting of excitation
amplitudes as a function of the coherent aberrations and probe position enables this
generalization of the theory for incoherent scattering, and encompasses all possible
scenarios from that of plane wave incidence to that of a point source focused within
a unit cell.
0.290 0.017
(c) Zn L–shell EELS (d) S K–shell EELS
Figure 4.1: STEM images for the case of electrons incident along the [110] zone
axis of ZnS, thickness of 122.4 Å, with incident energy of 100 keV. The probe is
aberration-balanced. (a) Elastic potential. (b) HAADF image with 60–160 mrad
detector. (c) Zn L-shell EELS image. (d) S K-shell EELS image. (e) Zn K-
shell EDX image. (f) S K-shell EDX image. Maximum and minimum fractional
intensities are given below the images.
Several points of interest may be identified in these figures. The HAADF images
give the finest feature size, and the EDX images tend to be more localized to the
known locations of the atoms than do the EELS images, especially for the L-shell in
zinc. Indeed this last becomes somewhat difficult to interpret directly at the larger
Chapter 4. Incoherent and inelastic imaging 101
t = 306 Å
0.363 0.050
(c) Zn L–shell EELS (d) S K–shell EELS
Figure 4.2: STEM images for the case of electrons incident along the [110] zone
axis of ZnS, thickness of 306.0 Å, with incident energy of 100 keV. The probe is
aberration-balanced. (a) Elastic potential. (b) HAADF image with 60–160 mrad
detector. (c) Zn L-shell EELS image. (d) S K-shell EELS image. (e) Zn K-
shell EDX image. (f) S K-shell EDX image. Maximum and minimum fractional
intensities are given below the images.
thicknesses. We will have more to say about issues of localization in EELS images
in the next chapter.
The two columns in the HAADF image are best distinguished for the thinner
specimens. As the thickness increases, the difference between the relative intensi-
ties and size of the spots diminishes. The simple Z-contrast model, which predicts
that the signals should scale as some power of the atomic number Z, is adequate
in this case for qualitative analysis but not quantitative analysis, a point which has
been made previously by Ishizuka (2001). Note from the intensity maximum the
saturation of the images – the signal does not increase greatly with the increase
in thickness. This property is shared by the EELS and EDX images. It suggests
that the majority of the contribution derives from near the top surface of the spec-
102
imen, that is to say that upon which the probe is incident. In a single channelling
model, in which the inelastically scattered waves are treated as plane waves (cf.
Bloch waves), this is unsurprising. Absorption means that fewer electrons remain in
the elastic channels at greater depths and so the probability of inelastic scattering
is correspondingly reduced. The single channelling model, by approximating the
inelastically scattered electrons as plane waves, assumes that the scattered electrons
are not affected by the quantity of crystal they must propagate through prior to
exiting the specimen. This is not true in detail, but for moderately large detectors
proves to be a good approximation (Josefsson and Allen, 1996).1
Comparing figure 4.1 and 4.4, those for the same thickness but using differ-
ent probes, we find unsurprisingly that the finer probe leads to finer features in
all the images. This is the idea driving the improvements in aberration correc-
tors, monochromators and system stability, the idea that finer probes lead to better
resolved and more interpretable images. In chapter 5 we will discuss further the va-
lidity of this idea and some limitations which may be placed upon it by the nature
of the interaction. However a few comments on this point are worth making here.
For thin crystals where the incoherent model is expected to be most justified,
because the probe has undergone little channelling and thus little change, we may
expect a reasonable correlation between the width of the probe and the width of the
STEM image. This property may persist to thicker specimens in so much as we have
noted that the greatest contribution to the signal for thicker crystals may still come
from the topmost layers. But in terms of channelling along columns, the picture
usually provided to explain column-by-column resolution, we do not expect that
increasing the aperture indefinitely will result in ever increasingly resolved images.
This may be argued geometrically, because widening the aperture corresponds to
admitting plane waves with larger transverse momentum components and as such
these may be less bound to the columns.
Channelling along the columns is often described in terms of an s-state model.
The Bloch states which are localized to the columns may be labelled in analogy to
bound atomic orbitals (Buxton et al., 1978), and so s-states are the Bloch states
most tightly localized upon given columns. Most of the recent development of such
channelling models has been geared towards CTEM (Van Dyck and Chen, 1999;
1
This model is justified precisely through the observation that it gives a good approximation
to the more rigourous scenario in which the final states are Bloch waves – and so are modified
by the amount of crystal through which they propagate – but where the integration over large
detectors tends to average out the detailed effects of this subsequent channelling. Ultimately it is
the extent to which the scattered wavefunctions are altered by further channelling that will provide
a thickness limit to the validity of the single channelling approximation.
Chapter 4. Incoherent and inelastic imaging 103
t = 459 Å
0.378 0.081
(c) Zn L–shell EELS (d) S K–shell EELS
Figure 4.3: STEM images for the case of electrons incident along the [110] zone
axis of ZnS, thickness of 459.0 Å, with incident energy of 100 keV. The probe is
aberration-balanced. (a) Elastic potential. (b) HAADF image with 60–160 mrad
detector. (c) Zn L-shell EELS image. (d) S K-shell EELS image. (e) Zn K-
shell EDX image. (f) S K-shell EDX image. Maximum and minimum fractional
intensities are given below the images.
Geuens and Van Dyck, 2002), but there is also a significant STEM literature on the
topic (Anstis et al., 2003). Since Bloch states are eigenstates then whatever portion
of the probe couples to the s-states will, absorption aside, stay there throughout
the propagation. While the probe intensity involves the coherent superposition of
all excited Bloch states,2 it is still reasonable to assert that the localization of the
wavefunction for a wide range of depths will be aided by coupling as large a portion
of the probe as possible to s-states. Peng et al. (2004) plot the excitation amplitude
for the s-state as a function of probe forming semi-angle and find that there is an
2
And, since it is the intensity which appears in equation (4.2), the width of the s-state is not,
as some authors have suggested, the fundamental limit to resolution in electron microscopy, since
the coherent superposition of Bloch states can lead to an intensity distribution peaked on a smaller
scale.
104
t = 122.4 Å
0.308 0.022
(c) Zn L–shell EELS (d) S K–shell EELS
Figure 4.4: STEM images for the case of electrons incident along the [110] zone
axis of ZnS, thickness of 122.4 Å, with incident energy of 100 keV. The probe
is aberration-free. (a) Elastic potential. (b) HAADF image with 60–160 mrad
detector. (c) Zn L-shell EELS image. (d) S K-shell EELS image. (e) Zn K-
shell EDX image. (f) S K-shell EDX image. Maximum and minimum fractional
intensities are given below the images.
angle of maximum coupling, with coupling strength dropping off monotonically for
both smaller and larger probe forming semi-angles. The angle of optimum coupling
depends on both the probe properties, such as coherent aberrations and accelerating
voltage, and the column properties, such as element, density and vibrational ampli-
tude. It will therefore vary from sample to sample. But the coupling is greater for
the probe used to form figure 4.4 than that used for 4.1, and this also likely aids the
increased resolution of the image. Having determined which Bloch states correspond
to the s-states, it is straight-forward to plot |αi (R)| as a function of aperture cutoff
to determine the aperture for optimum coupling
Chapter 4. Incoherent and inelastic imaging 105
Earlier in this section we reverted to treating Bloch states on the supercell, and thus
the reciprocal lattice vectors g and h used were those corresponding to the supercell.
However we saw in chapter 3 that for crystal specimens the wavefunction could be
decomposed into a set of wavefunctions associated with vectors in the set {q}, and
each such wavefunction could then be described using only the physical reciprocal
space lattice vectors. This block diagonalization of the A matrix afforded insight
into the physics but also facilitated the computation of the wavefunction. It is of
interest therefore to ask whether a similar rationale can be applied to, and similar
dividends obtained from, the cross section and fractional intensity expressions.
The MDFFs, including the site term for the crystal structure, obey an identity
equivalent to equation (3.12) for the optical potential, specifically:
N
XX X
I(R, t) = t B hk (R, q, t) h
CG k∗
(q)CH (q)µq+H,q+G . (4.7)
q h, k G,H
A more elaborate derivation of this result, cast in matrix form for the purposes of
providing a guide to the efficient implementation of the algorithm, is provided in
appendix C. In matrix form this result is expressed as
X £ ¤
I(R, t) = t Tr B(q)C † (q)U(q)C(q) , (4.8)
q
Substituting this into equation (3.22) for the excitation amplitude in the block di-
agonal formulation
Thus all the R dependence is coupled to terms ql . But the block diagonal formu-
lation has shown, equations (4.7) and (4.8), that the contributions from different
ql are independent of one another. For any given wavefunction ql , the terms B hk
contain R dependence only through the expression αh (R)αk∗ (R) which, from the
form of equation (4.10) above, is independent of R. Thus if the aperture is too
small to admit two vectors differing by a physical reciprocal lattice vector then the
resultant images contain no contrast.
Chapter 4. Incoherent and inelastic imaging 107
qmax = 0.20 Å–1 = 7.4 mrad qmax = 0.16 Å–1 = 5.9 mrad
C = 1.9 C=1
Figure 4.5: HAADF STEM image contrast for the case of electrons incident along
the [110] zone axis of ZnS, thickness of 122.4 Å, with incident energy of 100 keV,
for different aperture sizes. The aberration-balanced probe was used. The HAADF
detector spans 60–160 mrad. The C value given is the ratio of the maximum value
to the minimum value, a rudimentary definition for contrast. The aperture size
of 0.16 Å−1 is fractionally too small to encompass reciprocal vectors differing by a
full physical reciprocal lattice vector and as such the resulting STEM image has no
contrast.
108
Recognizing that the integral over z may be replaced by a sum over slices, equa-
tion (4.11) gives a multislice evaluation equivalent to the Bloch wave expression
equation (4.4) in the full nonlocal model. Note that Ψ(R, h, z) can be obtained via
fast Fourier transform.
Appendix E shows how these expressions reduce in the case of the local approx-
imation µh,g ≈ µh−g,0 . The result is
Z ·Z ¸
1X t
2
I(R, t) = µh,0 |ψ(R, r⊥ , z)| dz exp(2πih · r⊥ )dr⊥ . (4.13)
A h A 0
Note that the area integral is simply a Fourier transform of the depth integrated
intensity and thus rapidly calculable via fast Fourier transform.
Identifying the coefficients µh,0 to be the Fourier coefficients of a local, effective
inelastic potential Vinel , equation (4.13) is a re-statement of the relation (Cherns
et al., 1973) Z tZ
I(R, t) ∝ |ψ(R, r⊥ , z)|2 Vinel (r⊥ )dr⊥ dz . (4.14)
0 A
Chapter 4. Incoherent and inelastic imaging 109
A similar expression has been presented recently by Ishizuka (2001, 2002) in a mul-
tislice formulation of scattering by local potentials as regards STEM imaging. The
interpretation of this expression is that the contribution to the image from a given
slice in the crystal is the product of the probe intensity and the inelastic scatter-
ing potential Vinel (r⊥ ) at that slice. For HAADF imaging, the local approximation
is sufficient. However in some circumstances, in particular for the case of EELS
(Allen and Josefsson, 1995, 1996; Oxley and Allen, 2001), the local approximation
is inadequate and the nonlocal algorithm should be used.
Note that equations (4.13) and (4.14) have precisely the form required for an
incoherent image as per the definition of equation (4.1). If the crystal is sufficiently
thin that the intensity of the wavefunction does not differ from the probe at the
surface, then the restricted definition of an incoherent image, equation (4.2), also
holds since the intensity of the probe is a function only of the distance from the
probe location. The conditions under which this might be expected to hold are
discussed by Broeckx et al. (1995). Either way, if the interaction can be adequately
described by a local potential then the images will be incoherent images.
What if the interaction is only adequately described by a nonlocal potential?
Let us define
1 X 0
µ(r⊥ , r0⊥ ) ≡ 2 µh,g e2πih·r⊥ e−2πig·r⊥ . (4.15)
A g,h
This does not conform to the definition of an incoherent image. In particular, the
product ψ ∗ (R, r⊥ , z)ψ(R, r0⊥ , z) allows the phase difference between the wavefunc-
tion at point r⊥ and r0 to affect the signal. We will see implicitly in what follows
that inelastic images formed from nonlocal potentials may often appear like inco-
herent images as per the definitions in section 4.1. Therefore, by common usage in
the STEM field (cf. section 2.8.2), they merit the epithet of “incoherent images”.
However we will also see occasions when this interpretation is less sound, a possi-
bility which serves to emphasize the importance of image simulation as part of the
interpretation process for STEM images.
Equations (4.11) and (4.13) illustrate more clearly what we mean by the dynam-
ical contribution. These equations give the contribution to the detected signal from
electrons scattered out of the wavefunction described by ψ(R, r⊥ , z). By using the
110
imaging described in section 4.2 is a generalization, has been used successfully in cal-
culating channelling contrast as a function of rocking beam orientation for thermal
scattering into HAADF or back-scattered electron detectors, as well as for contrast
derived from atomic ionization events and observed by EELS or EDX (Allen and
Rossouw, 1989; Rossouw, 1995; Allen and Josefsson, 1995, 1996; Rossouw et al.,
1997). This gives much confidence in the validity of the method. The generality
of the method is another strong point in its favour. One noteworthy approxima-
tion made in the approach based on the MDFF is that the scattered electrons do
not undergo further dynamical diffraction, i.e. a single channelling approximation.3
Since a multislice MDFF algorithm has been presented, the question naturally arises
as to how the results of these calculations compare to the frozen phonon model, a
multislice model which, though readily applicable only to HAADF simulations, nat-
urally takes the channelling of the inelastically (thermally) scattered electrons into
account (Loane et al., 1991, 1992; Hillyard et al., 1993). It is interesting and in-
structive therefore to compare three models for HAADF STEM imaging – the Bloch
wave MDFF model, the multislice MDFF model and the frozen phonon model. For
convenience we will provide a very brief summary of the concepts behind the frozen
phonon model, drawing heavily on the discussions by Loane and co-workers (Loane
et al., 1991, 1992; Hillyard et al., 1993).
In the MDFF method we deal with the thermal motion of atoms in an Einstein
model using Debye-Waller factors: the atoms are associated with their equilibrium
position while the thermal motion is accounted for through a time-averaged poten-
tial. However the interaction time of the incident electrons with any particular atom
in the solid is of the order of 10−2 vibrational periods of the atom (Loane et al., 1991).
Therefore any given electron can be said to see a static configuration of atomic posi-
tions, but, over the recording times considered, different electrons will see different
configurations. Thus another way to build up a physically reasonable picture is to
determine the average diffraction pattern for several different static configurations
of atomic positions in accordance with their thermal displacements. The average
taken is incoherent – an average over intensities rather than complex amplitudes.
This constitutes a Monte Carlo style integration over phonon configuration space.
This is the essence of the frozen phonon model.
The positions of atoms in a specimen for a given configuration are obtained by
displacing each atom a distance based on a Gaussian probability distribution arising
3
The expression for incoherent lattice contrast can be generalized to include double channelling
(Josefsson and Allen, 1996) but the calculation then scales as N 8 rather than N 4 , where N is the
order of the A-matrix, and hence is numerically demanding.
112
from a simple harmonic oscillator model. Muller et al. (2001) presented calculations
comparing realistic phonon modes to the Einstein model and found that for the
simulation of HAADF images the difference is small. Therefore we will adopt the
simpler Einstein model.
“A natural consequence of explicitly calculating the scattering from each phonon
configuration is the inclusion of multiple elastic and TDS [thermal diffuse scattering]
scattering to all orders” (Loane et al., 1991). As such, the images display a thermal
background showing Kikuchi band structure. In this sense, it may be argued that
the frozen phonon model is theoretically more realistic, though there is reason to
expect that for HAADF STEM images the integration over the annular detector
is such that the difference between these two models on this score should be small
(Nellist and Pennycook, 1999). In an empirical sense this will be borne out by our
results.
Let us consider the HAADF STEM image resulting from scattering through ZnS
along the [110] zone axis. Indeed we will consider only a line scan, the line scan
through the dumbbells along the [001] direction as shown in figure 2.4 (f), since
quantitative comparisons may more easily be made on such data.
The case considered will be a 122.4 Å thick sample of ZnS (sixteen repeat dis-
tances along the propagation direction). The incident energy is taken as 100 keV.
The aberration-balanced probe, characterized by ∆f = 62 Å, Cs = −0.05 mm,
C5 = 63 mm, and a cutoff of 0.539 Å−1 (i.e. 20 mrad), is used. The annular detector
is taken to span the range 40–80 mrad. Figure 4.6 shows the line scans simulated for
these parameters in the Bloch wave MDFF model (solid line), the multislice MDFF
model (long dashed line) and the frozen phonon model (short dashed line). The
Bloch wave MDFF model uses 205 beams in each of the sub-matrices of the block
diagonalized structure matrix and a 7 × 7 supercell. The multislice MDFF model
used a 512 × 512 grid inside which the unit cell used was tiled 6 × 8 times. For the
frozen phonon model, adequate convergence required a 1024 × 1024 grid (the frozen
phonon model being more susceptible to errors in the high frequency range) inside
which the unit cell used was also tiled 6 × 8 times. In the frozen phonon model the
unit cell was sliced as per a higher order Laue zone simulation – using four distinct
potentials in the 7.65 Å repeat distance – as is the usual practice, although higher
order Laue zone effects are believed to make only a small contribution to HAADF
imaging (Amali and Rez, 1997), a notion borne out by our calculations.
It is readily seen from figure 4.6 that the agreement between the two distinct
codings of the MDFF model is excellent. The comparison of the results of the MDFF
Chapter 4. Incoherent and inelastic imaging
ZnS [110] ADF 40-80 mrad 113
AB probe at 100 keV
0.30
Bloch wave MDFF
0.25 Multislice MDFF
Fractional intensity
Frozen phonon
0.20
0.15
0.10
0.05
Zn S Zn
0.00
0 1 2 3 4 5
Figure 4.6: HAADF STEM image contrast along the line scan [shown in figure 2.4
(f)] for the case of 100 keV electrons incident along the [110] zone axis of ZnS. The
probe is aberration-balanced. The annular detector is taken to span 40–80 mrad. It
is seen that the two MDFF methods, theoretically equivalent but possessing distinct
encodings, give indistinguishable results. It is seen that the frozen phonon model is
in excellent agreement with the results of the MDFF models.
model with the frozen phonon model is also very favourable. Thus we may conclude
that double channelling has small effect for the example considered here.
It should be noted that our use of the fractional intensity allows for comparison
of the MDFF model with the frozen phonon model on an absolute scale – the plots
above have not been re-scaled in any way.
The theoretical basis of the frozen phonon model, inherently taking double chan-
nelling into account, is superior to that of the MDFF model for HAADF and as such
the accuracy of the frozen phonon model will naturally be preferred in many cases.
However there is another concern which should not be wholly neglected in choosing
between the three methods discussed and that is computation time.
Let us first neglect the frozen phonon model and consider only the comparison
between the two MDFF models. For the case of perfect crystals, the block diag-
onalization of the A-matrix makes the Bloch wave method very efficient for the
calculation of STEM images.4 For such cases the Bloch wave method is to be pre-
4
It should be noted that the coupling of reciprocal vectors differing by physical reciprocal lattice
vectors, and the decoupling between sets which do not differ by physical reciprocal lattice vectors,
is a physical effect and thus not particular to the Bloch wave model. A Fourier space encoding
of the multislice could equally well be block diagonalized by separately propagating the different
114
ferred. However, consider now the simulation of diffraction through a crystal defect.
The standard method of treating such structures is to use supercells [the method of
periodic continuation (Fields and Cowley, 1978; Cowley and Fields, 1979)] and so is
naturally suited to the requirements of STEM calculations. For such a situation, the
block diagonalization property is effectively lost, or at least must now be applied to
the supercell which means that the number of non-zero matrix elements increases
accordingly. For the Bloch wave method this is a problem, because determination
of the Bloch states must now be carried out on the full structure matrix and not
on the sub-matrices provided in the periodic case by block diagonalization. But the
multislice calculation of the wavefunction on the supercell is equally amenable to
this case as to the periodic case. For such cases it is therefore the multislice method
that is to be preferred.
Another consideration is that the multislice calculation time scales linearly with
thickness whilst the Bloch wave calculation time is independent of the value of the
thickness. Thus if the periodicity makes the Bloch wave method at all tractable for
simulating the wavefunction then for thick crystals it is to be strongly preferred over
the multislice method.
The frozen phonon model is invariably evaluated by multislice methodology and
as such has the same standing as the multislice MDFF method as compared to the
Bloch wave MDFF method on the points made above. For a single run through
the crystal, the frozen phonon model and multislice MDFF model take the same
amount of computation time. However in order to adequately sample the phonon
configuration space in the Monte Carlo type integration in the frozen phonon model,
it becomes necessary to propagate through the crystal several times for each probe
point considered. Our calculations have used twenty points in the Monte Carlo
integration, and Loane et al. (1992) give a similar number as adequate for simulating
diffraction patterns.5 Thus while the frozen phonon model is theoretically preferred,
a converged frozen phonon calculation may take ten times longer than a multislice
MDFF calculation. Accuracy must not be divorced from the choice of algorithm,
vector sets. Thus in principle one may, say, perform 8 multislice calculations on 64×64 grids rather
than a single calculation on a 512 × 512 grid. However it is our experience that the requirements
of adequate sampling in the multislice require the block diagonal grids to be large and as such the
computational saving afforded is smaller than that of the Bloch wave model.
5
Because the HAADF detector integrates over much of the fine structure in the diffraction
pattern it is often the case that for HAADF images it suffices to use only a few phonon configura-
tions. It should thus be borne in mind that for large detectors adequate results may be obtained
with similar efficiency to the MDFF calculations. We will use the larger number, not to make the
comparison less favourable, but to ensure convergence and for generality (i.e. not being restricted
to large detectors).
Chapter 4. Incoherent and inelastic imaging 115
however when there is reason to believe that the MDFF model will give adequate
results, as in our test case, its relative computational simplicity makes it extremely
competitive.
Table 4.1 shows the computation times for the case shown in figure 4.6, with
the exception that the frozen phonon model times refer to a 512 × 512 grid for fair
comparison with the multislice MDFF model (which computationally corresponds
to a small sacrifice in the accuracy of the frozen phonon model). It also shows
how the different algorithms scale with the number of pixels used in the scan, with
the sample thickness, and, for the frozen phonon case, with the number of times
one traverses the sample in the Monte Carlo style integration, the number of runs.
Note that the Bloch wave method is not precisely linear with pixel number since
the eigenvalue problem must be performed in entirety regardless of the number of
points scanned. Once performed, the remaining time in the Bloch wave calculation
scales linearly with number of points. Note that the Bloch wave computation time
is independent of crystal thickness. The difference in time between the multislice
MDFF model and the frozen phonon model is primarily due to the requirement in
the latter of propagating multiple times through the crystal for each point – 20 times
in our simulation.
To reiterate, accuracy is always the primary consideration. However, in cases
where all models have similar accuracy, the computation times and scaling behaviour
in table 4.1 should be considered carefully in the selection of the algorithm to be
used.
Table 4.1: Comparison of computation times and scaling behaviour between the
models. A dash denotes that a category is not applicable to that algorithm. The
calculation times were obtained using an Intel XEON, 2.0 GHz machine.
(a)
4.0
–5
Multislice
3.0
2.0
1.0
Zn S Zn S Zn
0.0
0 2 4 6 8 10
Probe position (Å)
(b)
0.25
Fractional intensity (×10 )
–5
0.20
0.15
0.10
0.05
0.00
0 2 4 6 8 10
Probe position (Å)
Figure 4.7: (a) Zn L-shell EELS line scan. (b) S K-shell EELS line scan.
detector. The effect of chromatic aberration of the lens is to give different energies
within the beam slightly different defocus values. The range of resultant images
is to be averaged incoherently and this leads to a blurring effect. This can often
be adequately simulated using a coherent envelope function approach (Allen et al.,
2004a).
In STEM this approach is generally not taken, though the approximation is very
similar. The different nature of the transfer function for STEM in an incoherent im-
age formulation as per equation (4.2) is discussed by Nellist and Pennycook (1998b).
With the probe prior to the specimen the range of defocus values occasioned by the
effect of the lens on the spread of energies from the tip should be thought of as
118
producing a set of incoherent probes at the entrance surface. However rather than
averaging the intensities as in the CTEM case, each different wavefunction will now
propagate through the crystal, independently of the others. This extends to the
manner in which these wavefunctions undergo inelastic scattering and so contribute
to the inelastic images we may wish to consider.
Let us re-write the incoherent image expression, equation (4.4), as
X X
I(R, ∆f + F, t) = t B ij (R, ∆f + F, t) Cgi Chj∗ µh,g , (4.17)
i,j g,h
where
B i,j (R, t) ≡ Λi,j (R)Lij (t) , (4.19)
Thus the chromatic aberration term means that Λi,j (R) must be evaluated by car-
rying out the double sum, which represents an increase in computational effort over
the product of single sums implicit in equation (4.5). This factorization structure
may be regained using the factorization approximation of equation (3.56). This
approximation holds best for small momentum vectors, and so the approximation
worsens as the probe-forming aperture size increases. This approximation has been
evaluated in detail by Allen et al. (2004a).
4.7 Conclusion
By far the most significant dividend of the new formulation of the boundary con-
ditions in the Bloch wave model is that it makes immediate the generalization of
the cross-section expression to the STEM geometry, indeed to an arbitrarily-shaped,
Chapter 4. Incoherent and inelastic imaging 119
Chapter 5
©©
H
¼ ª?
¡ RH
@ j
121
122
also through increases in system stability through purpose-built facilities, have al-
lowed for probes capable of sub-Ångstrom resolution and as such column-by-column
imaging, particularly in the HAADF imaging mode, is becoming routine. Since the
interpretation of such images is predominantly visual – bright spots being inter-
preted as atomic columns – more recent theoretical investigations have considered
the extent to which false spots may arise as the result of probe spreading, the sig-
nificant dynamical effect which simplified incoherent imaging models such as that
given by equation (4.2) perforce neglect.
Watanabe et al. (2001a,b) discuss how false peaks arise in open channels due to
the spreading of the probe. The resolution considered in those works is sufficient to
resolve the dumbbell structure in silicon in [011] zone axis orientation, but describes
a pre-Cs -corrected system. As compared to lower resolution systems with broader
probes but smaller incident transverse momentum components, the advantages of
finer, high resolution probes may be offset to some extent by the increase in incident
transverse momentum components. The point that probe spreading may limit the
validity of column-by-column analysis has been made by several authors (Hillyard
et al., 1993; Plamann and Hÿtch, 1999; Ishizuka, 2001; Dwyer and Etheridge, 2003),
and some of that work includes calculations for aberration-corrected probes. Vital
though such investigations of the dynamical propagation of the probe are, the extent
to which the probe spreads in real space inside the crystal should not be automati-
cally interpreted as describing the extent to which other columns influence the signal
that reaches the detector. We refer to the possibility that a probe, initially focused
onto an atomic column, may interact with atoms in neighbouring columns in such
a way as to influence the signal on the detector, as cross-talk. The formulation
for incoherent contrast presented previously allows this issue to be addressed in a
quantitative manner. Ishizuka (2001, 2002) has already explored the quantitative
validity of some qualitative features often attributed to STEM images, such as their
insensitivity to thickness and defocus, and the approximate Z 2 scaling of the column
contrast.
The expression for the mixed dynamic form factors (MDFFs) for thermal diffuse
scattering given in equation (2.39) incorporates the different atom positions through
the site term, the summation of the phase factors over the atom locations. This is
cast in a reciprocal space formulation, but represents the idea that the effective
scattering potential for a cell is the sum of effective scattering potentials for the
individual atoms within that cell. That we can directly associate a component of
the MDFF to a particular atom means that we can meaningfully talk about the con-
124
tribution to the STEM images from individual atoms within the solid. This assumes
the single channelling approximation – that electrons once thermally scattered will
not, for the purposes of obtaining the HAADF image, undergo further scattering.
It also assumes an Einstein model for the thermal scattering, since scattering from
a correlated phonon mode could not be identified with a single atom. Both these
assumptions seem to be supported for HAADF imaging (Muller et al., 2001; Findlay
et al., 2003).
Both the technique and the sorts of results obtained will be illustrated with a
few examples. Multislice simulations for HAADF imaging were performed on ZnS
in [110] zone axis orientation at room temperature. We will compare the results for
three of the different probes, the 100 keV aberration-balanced probe, the 100 keV
aberration-free probe, and a 300 keV aberration-balanced probe. The first two
probes were described in detail in section 2.5. The last, the 300 keV probe, has
parameters ∆f = 20 Å, Cs = −0.037 mm, C5 = 100 mm, and qmax = 1.067 Å−1
(i.e. 21 mrad), which are typical operating parameters for the VG Microscopes’
Chapter 5. Subtleties in qualitative and quantitative image interpretation 125
(a) (b)
× VHAADF
× V HAADF
2.5 1.0
2.0 0.8
(×10 )
1.5
(×10 )
–4
Probe intensty 0.6
–4
Probe intensty
1.0 0.4
0.5 0.2
0.0 0.0
-0.5 2.8 ) 2.8
3.5 (Å -0.2
3.5 (Å
)
200
300
4.2 it ion 200 4.2 io n
Thi
ckn 400 4.9 po
s Th i 300 sit
ess 500
be
ckn
ess
400
500
4.9 po
(Å) o e
Pr (Å) ob
Pr
(c) (d)
10.0 4.0
Fractional intensity (×10 )
–3
full crystal
8.0 column alone
neighbouring 3.0
column
6.0
full crystal less
dumbbell 2.0
4.0
1.0
2.0
0.0 0.0
0 100 200 300 400 500 0 100 200 300 400 500
Thickness (Å) Thickness (Å)
(e) (f)
0.45 0.45
Fractional intensity
Fractional intensity
0.30 0.30
0.15 0.15
0.00 0.00
0 100 200 300 400 500 0 100 200 300 400 500
Thickness (Å) Thickness (Å)
Figure 5.1: HAADF simulations using the 100 keV, aberration-balanced probe on
ZnS along [011]. The product of the probe intensity and the HAADF potential as a
function of thickness along a single, horizontal line intersecting the dumbbells, as per
figure 2.4 (f), with the probe situated on (a) the Zn column and (b) the S column.
For the probe situated on (c) the Zn column and (d) the S column, the contribution
to the fractional intensity from the previous 2 Å thick slice is shown for the HAADF
signal using the form factor for: the full crystal (solid line), the column beneath
the probe (dashed line), the the neighbouring column in the dumbbell (dash-dotted
line), and the target excluding the dumbbell in question (dotted line). The fractional
intensity is shown in (e) and (f) as a function of sample thickness.
Chapter 5. Subtleties in qualitative and quantitative image interpretation 127
(a) (b)
× V HAADF
V HAADF
3.0 1.0
2.5 0.8
×
2.0
(×10 )
(×10 )
0.6
–4
Probe intensity
–4
Probe intensity
1.5
0.4
1.0
0.5 0.2
0.0 0.0
2.8 2.8 )
-0.5 ) -0.2 (Å
3.5 (Å 3.5
io n
n t
200
300
4.2
it io 200 4.2 si
Thi
ckn 400 4.9 os Thi 300400 4.9 po
ess 500 p ckn e
e ess 500 ob
(Å) ob (Å) Pr
Pr
(c) (d)
15.0 5.0
Fractional intensity (×10 )
–3
full crystal
12.0 column alone 4.0
neighbouring
column
9.0 3.0
full crystal less
dumbbell
6.0 2.0
3.0 1.0
0.0 0.0
0 100 200 300 400 500 0 100 200 300 400 500
Thickness (Å) Thickness (Å)
(e) (f)
0.45 0.45
Fractional intensity
Fractional intensity
0.30 0.30
0.15 0.15
0.00 0.00
0 100 200 300 400 500 0 100 200 300 400 500
Thickness (Å) Thickness (Å)
Figure 5.2: HAADF simulations using the 100 keV, aberration-free probe on ZnS
along [011]. The product of the probe intensity and the HAADF potential as a
function of thickness along a single, horizontal line intersecting the dumbbells, as
per figure 2.4 (f), with the probe situated on (a) the Zn column and (b) the S
column. For the probe situated on (c) the Zn column and (d) the S column, the
contribution to the fractional intensity from the previous 2 Å thick slice is shown
for the HAADF signal using the form factor for: the full crystal (solid line), the
column beneath the probe (dashed line), the neighbouring column in the dumbbell
(dash-dotted line), and the target excluding the dumbbell in question (dotted line).
The fractional intensity is shown in (e) and (f) as a function of sample thickness.
128
(a) (b)
× V HAADF
VHAADF
2.0 0.6
1.5
×
0.4
(×10 )
(×10 )
–4
Probe intensity
–4
Probe intensity
1.0
0.2
0.5
0.0 0.0
2.8 ) 2.8 )
100 3.5 (Å 100 3.5 (Å
200 4.2 o n 200 4.2 io n
T 300 it i Thi 300 it
hick
nes
400 4.9 os p ck nes
400 4.9 os p
s (Å 500 be s( 500 e
) o Å) ob
Pr Pr
(c) (d)
4.0 1.2
Fractional intensity (×10 )
–3
full crystal
column alone
neighbouring
column 0.8
2.0 full crystal less
dumbbell
0.4
0.0
0.0
0 100 200 300 400 500 0 100 200 300 400 500
Thickness (Å) Thickness (Å)
(e) (f)
0.20 0.20
Fractional intensity
Fractional intensity
0.15 0.15
0.10 0.10
0.05 0.05
0.00 0.00
0 100 200 300 400 500 0 100 200 300 400 500
Thickness (Å) Thickness (Å)
Figure 5.3: HAADF simulations using the 300 keV, aberration-balanced probe on
ZnS along [011]. The product of the probe intensity and the HAADF potential as
a function of thickness along a single, horizontal line intersecting the dumbbells, as
per figure 2.4 (f), with the probe situated on (a) the Zn column and (b) the S column
respectively. For the probe situated on (c) the Zn column and (d) the S column, the
contribution to the fractional intensity from the previous 2 Å thick slice is shown
for the HAADF signal using the form factor for: the full crystal (solid line), the the
column beneath the probe (dashed line), the neighbouring column in the dumbbell
(dash-dotted line), and the target excluding the dumbbell in question (dotted line).
The fractional intensity is shown in (e) and (f) as a function of sample thickness.
Chapter 5. Subtleties in qualitative and quantitative image interpretation 129
tatively the amount of cross-talk incurred for a probe focused onto either of neigh-
bouring columns of zinc or sulfur. The different plots represent the contributions per
slice to the HAADF signal for the following four scenarios, in which the HAADF po-
tential is: (i) taken into account for the full crystal; (ii) limited to the column alone;
(iii) limited to the neighbouring column in the dumbbell; and (iv) taken from the
full crystal, but with contributions from the local dumbbell excised. Contributions
(ii)-(iv) naturally add up to give contribution (i), the total HAADF signal. Cross-
talk plays a more significant role when the probe is focused onto a sulfur column.
The integrated total intensity for a given thickness is shown in figures 5.1 (e) and
(f), 5.2 (e) and (f), and 5.3 (e) and (f). The cumulative effect of cross-talk is rela-
tively small for the probe on the zinc column. For the lower energy probes cross-talk
becomes more significant when positioned on a sulfur column. It is interesting to
note that this effect is larger for the finer probes. It would seem that the higher con-
vergence angle of the aberration-free probe leads to more spreading initially. This is
expected geometrically, because the higher convergence angle corresponds to higher
transverse momentum components. It is also interesting to note that for a thick
enough specimen the contributions from the full crystal less the dumbbell exceed
those from the neighbouring column, suggesting that the spreading is significant,
even if the additional contribution to the image is small.
Figure 5.4 applies the same analysis to InP, taken along the [110] zone axis, using
the 100 keV aberration-balanced probe. The increased scattering power of indium
in absolute terms leads to the signal from the probe atop the indium column being
generated primarily near the crystal surface, because the probe is so attenuated
that little significant intensity remains in the elastic channel at greater depths.
The increased scattering power of indium relative to phosphorus means that the
proportion of cross-talk when the probe is situated on the lighter phosphorus column
is greater than when on the sulfur column in ZnS.
We have tended to perform simulations of this nature in the multislice method
because being able to visualize the real space distribution of the probe and thus its
overlap with the HAADF potential is informative. However the shape of the plots
of the cumulative fractional intensity find simple explanation in the Bloch wave
model.1
To this end, let us rewrite equation (4.4) as
X
Idyn (R, t) = I ij (R, t) , (5.1)
i,j
1
I am grateful to Dr Y. Peng for pointing this out.
130
(a) (b)
VHAADF
× VHAADF
8.0 0.8
6.0 0.6
×
(×10 )
(×10 )
–4
Probe intensity
–4
Probe intensity 4.0 0.4
2.0
0.2
0.0
0.0
-2.0 2.8 ) 2.8 )
3.5 (Å 100 3.5 (Å
200 io n 200 4.2 on
4.2 i
Thi 300 sit Thi 300 si t
ck nes
400
500
4.9 po ckn 400
500
4.9 po
e ess e
s (Å ob (Å) ob
) Pr Pr
(c) (d)
15.0 3.0
Fractional intensity (×10 )
–3
full crystal
12.0 column alone
neighbouring 2.0
9.0 column
full crystal less
dumbbell
6.0
1.0
3.0
0.0 0.0
0 100 200 300 400 500 0 100 200 300 400 500
Thickness (Å) Thickness (Å)
(e) (f)
0.3 0.3
Fractional intensity
Fractional intensity
0.2 0.2
0.1 0.1
0.0 0.0
0 100 200 300 400 500 0 100 200 300 400 500
Thickness (Å) Thickness (Å)
Figure 5.4: HAADF simulations using the 100 keV, aberration-balanced probe on
InP along [011]. The product of the probe intensity and the HAADF potential as
a function of thickness along a single, horizontal line intersecting the dumbbells,
as per figure 2.4 (f), with the probe situated on (a) the In column and (b) the P
column. For the probe situated on (c) the In column and (d) the P column, the
contribution to the fractional intensity from the previous 2 Å thick slice is shown
for the HAADF signal using the form factor for: the full crystal (solid line), the
column beneath the probe (dashed line), the neighbouring column in the dumbbell
(dash-dotted line), and the target excluding the dumbbell in question (dotted line).
The fractional intensity is shown in (e) and (f) as a function of sample thickness.
Chapter 5. Subtleties in qualitative and quantitative image interpretation 131
where
X
I ij (R, t) = tB ij (R, t) Cgi Chj∗ µh,g . (5.2)
g,h
Let us suppose that we are interested in the so-called s-state Bloch wave (Buxton
et al., 1978) associated with a particular column, and number the Bloch states such
that this s-state is state 1. We may then write equation (5.2) as
" # " #
X X
Idyn (R, t) = I 11 (R, t) + 2Re I 1j (R, t) + 2Re I ij (R, t) , (5.3)
j6=1 i>j6=1
where we have used the hermiticity of the MDFFs to note that I ij = I ji∗ . The
decomposition is a mathematical trick, but the terms are not without meaning. The
first term is positive definite and a fractional intensity in its own right; if the s-state
was the only state excited then this would be the entire fractional intensity. The
third term is also positive definite, and would be the entire fractional intensity if the
s-state excitation amplitude was zero. It is the middle term, the mixed term which
relies on excitation of both s-states and non-s-states, which is not positive definite.
That it can be negative indicates that the total fractional intensity might be less than
what would be obtained with the s-state in isolation because, when several states
are present, interference can alter the probability of inelastic transition. Figure 5.5
shows the decomposition of the signal from ZnS as investigated with the 300 keV
aberration-balanced probe (cf. figure 5.3) into these three terms.2
Figures 5.5 (a) and (b) show again the total HAADF signals with the probe
situated above the zinc and sulfur columns respectively, as well as their decompo-
sition into s-state, mixed s-state/non-s-state and non-s-state contributions as per
the terms in equation (5.3) with the s-states corresponding to the column on which
the probe is placed. In both cases the pure s-state contribution is very similar in
magnitude to the total signal, suggesting that this Bloch state dominates the signal.
The discrepancies, more pronounced for sulfur than for zinc, are attributable to the
mixed terms for thicknesses to around 100 Å, beyond which it is the non-s-state
terms which give the greater contribution. Figures 5.5 (c) and (d) show the per
slice contribution of the signals from figures 5.5 (a) and (b). The s-state contri-
butions here are smooth, decaying slowly with increasing thickness as a result of
2
Note: The notation of equations (5.1), (5.2) and (5.3) suggests that the s-state is a single
Bloch state. However Bloch states have the periodicity of the physical crystal. The column s-state
is rather the superposition of the s-state Bloch waves for different q values, treated as a branch
cluster. It should therefore be understood that a sum over q values is contained in each term in
equation (5.3).
132
(a) (b)
0.10
0.15
Fractional intensity
Fractional intensity
0.08
0.00 0.00
0 100 200 300 400 500 0 100 200 300 400 500
Thickness (Å) Thickness (Å)
(c) (d)
4.0 1.2
Fractional intensity (×10 )
–3
3.0 0.9
2.0 0.6
1.0 0.3
0.0 0.0
-1.0 -0.3
0 100 200 300 400 500 0 100 200 300 400 500
Thickness (Å) Thickness (Å)
Figure 5.5: HAADF simulations using the 300 keV, aberration-balanced probe show-
ing a decomposition of the total intensity into separate contributions. (a) The in-
tegrated total intensity for a given thickness with the probe atop the Zn column.
The s-state is that pertaining to this column. (b) The integrated total intensity for
a given thickness with the probe atop the S column. The s-state is that pertaining
to this column. The contributions per slice to the HAADF signals given in (a) and
(b) are shown in (c) and (d) respectively.
the absorptive attenuation of the state. The oscillatory structure in the per slice
plot for the total thickness is almost all attributable to the equivalent fluctuations
in the mixed s-state/non-s-state term. This results from the thickness dependent
interference between the different Bloch states. This allows the contribution from
the mixed term to become negative. It also leads to the reduction of this term for
large thicknesses where the result of de-phasing is to give no net contribution.
It should be stressed that while for the probe directly above the columns we have
demonstrated that the s-state model gives a good feel for the general magnitude and
thickness dependence of the detected signal, the same may not be said when the
Chapter 5. Subtleties in qualitative and quantitative image interpretation 133
probe is somewhat off the column. For such cases the measured signal strength tends
to exceed that attributable to s-state components, even though this signal is smaller
than that with the probe on the column. The s-state model is indeed a useful tool,
but should be used with care when more quantitative insights are sought.
have been inconsistent, some predicting significant delocalization and others strong
localization (Kohl and Rose, 1985; Holbrook and Bird, 1995; Muller and Silcox,
1995; Rafferty and Pennycook, 1999). In the local approximation, the effective
scattering potential may again be plotted in real space and the delocalization may
be assessed in terms of the width of the peaks. However, it has been shown that
EELS simulations generally need to be done using the fully nonlocal model (Allen
and Josefsson, 1995, 1996; Oxley and Allen, 1998, 2000, 2001) and it is not easy to
carry that definition of delocalization over to nonlocal potentials.
Since we are interested in STEM images we will discuss delocalization, not of
the interaction potential, but rather of the resultant STEM image, by which we
simply mean the spread of significant signal about the pertinent column location.
The image, being a real space quantity, can easily be visualized. More importantly,
it is what is measured experimentally. The disadvantage is that the STEM image
is not purely a property of the atom but also depends on the probe. Some of the
effects of channelling through the crystal can be separated by simulating the STEM
image of a single atom in isolation. This situation of itself has limited experimental
relevance, but will provide some useful general insights prior to investigating the
realistic crystal case in depth.
Consider equation (4.16) restricted to a single atom (i.e. with the integration
over depth performed on the assumption that the quantities in the integrand do not
significantly vary with z):
Z Z
I(R) = ψ ∗ (R, r⊥ )µ(r⊥ , r0⊥ )ψ(R, r0⊥ )dr⊥ dr0⊥ . (5.4)
We may obtain a quantity specific to the potential by assuming the probe wavefunc-
tion to be a δ-function: ψ(R, r⊥ ) = δ(R − r⊥ ). In this case [cf. equation (F.3)]
1 X
I(R) = µ(R, R) = µh,g e2πi(h−g)·R (5.5)
A2 h,g
Z Z
' µh,g e2πi(h−g)·R dhdg ,
where the final step follows from treating an isolated atom. This might be considered
as the limiting case of ultra-fine probes and the delocalization of this STEM image
may be regarded as the limit applied by the nature of the interaction.3
3
It is not guaranteed that, for any given definition of image width, I(R) defined by equation
(5.5) will be less delocalized than that of equation (5.4) because the form of the nonlocal interaction,
the integral over contributions of the product of complex wavefunction amplitudes evaluated at
Chapter 5. Subtleties in qualitative and quantitative image interpretation 135
(a) (b)
1.0 1.0
O O
Normalized intensity
Normalized intensity
Si Si
0.8 0.8
Ca Ca
0.6 0.6
0.4 0.4
0.2 0.2
0.0 0.0
2 3 4 5 6 7 8 2 3 4 5 6 7 8
(c) (d)
1.0 1.0
O O
Normalized intensity
Normalized intensity
Si Si
0.8 0.8
Ca Ca
0.6 0.6
0.4 0.4
0.2 0.2
0.0 0.0
2 3 4 5 6 7 8 2 3 4 5 6 7 8
Figure 5.6: K-shell EELS images of single atoms of O, Si and Ca with (a) δ-function
probe, (b) aberration-free probe with qmax = 0.676 Å−1 (i.e. 25 mrad), (c) aberration-
balanced probe with qmax = 0.539 Å−1 (i.e. 20 mrad), and (d) aberrated probe with
qmax = 0.38 Å−1 (i.e. 14 mrad). The images have been normalized to a common
intensity to facilitate comparison of their spatial distribution.
Several general trends may be noted. The images for more tightly bound shells,
the K-shells for heavier elements, are more localized. The light oxygen atoms lead
differing positions, may allow interferometric cancellation to significantly alter the shape of I(R).
136
to signals which are significantly less localized than silicon and calcium. These last
two are more similar, even though the separation in atomic number between silicon
and calcium is identically that between oxygen and silicon. It is noted that the
finer probes do indeed tend to give more localized images, as is the experimental
hope implicit in the quest to improve resolution by improving the probe-forming
optics. However note also that in approaching the limit of the δ-function probe
we see volcano-structure developing in the images of single atoms. This feature in
single atom images has been explained previously by Kohl and Rose (1985). Briefly,
in the single atom case it arises when the detector semi-angle is relatively small
compared with the probe forming semi-angle such that with increasing distance
between the probe and the atom site the proportion of signal forward scattered into
the detector initially increases faster than the total signal decreases, providing the
maximum detector signal when the probe is slightly off the atom position. Put
another way, for the probe above the atom much large angle inelastic scattering
misses the detector, so much so that it rates as a deficit relative to the case where
the probe is displaced slightly from the column.
Note that for the purposes of exploring resolution we have been interested in
a comparison of the width of the signals from different atoms. To emphasize this
feature, the plots in figure 5.6 have all been normalized to a maximum signal of
unity. However the count rate drops off rapidly with increasing binding strength
of the inner shell. Thus for the aberration-free probe, figure 5.6 (b), the maximum
signal strength from the oxygen atom is nine times that from the silicon atom and
sixty-one times that from the calcium atom.
A systematic and quantitative study of the delocalization of single atom STEM
images has been carried out by Cosgriff et al. (2005), who considered not only
the variations and trends for different atomic numbers but also for different probe
apertures. The tools developed here, and the systematic explorations of Cosgriff
et al. (2005), provide a solid basis from which to explore the localization of EELS
images in crystals. To that task we now turn our attention.
STEM images of single atoms. The width of the images shown was generally smaller
than typical inter-column spacings, giving some cause for optimism. However it
should be stressed that the results of the previous section hold for a single atom.
In a crystal, the dynamical nature of propagation may serve to alter any probe in
such a way as to change the delocalization of the images. We will explore this idea
through a specific case study involving comparison with experimental data.
Line scans along [001] direction of SrTiO3 viewed along the [100] zone axis were
taken using the 100 keV VG Microscopes’ HB501UX microscope at Oak Ridge Na-
tional Laboratory, equipped with a Nion aberration corrector (Dellby et al., 2001).
The probe used is characterized by the parameters ∆f = 62 Å, Cs = −0.05 mm,
C5 = 63 mm, and qmax = 0.539 Å−1 (i.e. 20 mrad). The EELS scan is based on
titanium L-shell ionization, with a detector semi-angle of 20 mrad, and an energy
window of 40 eV above the ionization threshold. Line scans are taken along the
[011] direction in which strontium columns alternate with mixed titanium and oxy-
gen columns. The simultaneously acquired HAADF scan used an annular detector
range of 56–202 mrad.
Figure 5.7 shows the experimental line scans for HAADF and titanium L-shell
EELS in SrTiO3 , scaled by their average value for convenience of display. The
HAADF signal clearly alternates between strong and weak peaks at a spacing con-
sistent with the known structure of SrTiO3 (cubic, at room temperature, with side
length 3.9 Å). The bright peaks are taken to be the strontium columns, the weaker
peaks then are the mixed titanium and oxygen columns. Simulations bear this out;
in no parameter set investigated was the ratio between the columns reversed.
2.5
HAADF
Ti L-shell EELS
2.0
N/NAVE
1.5
1.0
0.5
2.75 Å
0.0
0 2 4 6 8 10 12 14
Probe position (Å)
Figure 5.7: Experimental data plots showing the HAADF and Ti L-shell EELS
results from SrTiO3 . Experimental parameters as given in the text.
138
Since the two signals were recorded simultaneously, we may take the peaks of
the HAADF signal to calibrate the probe position as reference for the EELS signal.
We see that the EELS signal does indeed give peaks corresponding to the mixed
titanium and oxygen column positions.
The thickness of the crystalline region is gauged to be approximately 200 Å. It
is likely that amorphous surface layers are present in addition to this, as ion milling
was used in specimen preparation. The fluctuation in the signals, and the distinct
variation between the two adjacent cells, may be attributed to physical instability in
the microscope and the presence of amorphous surface layers. Yet in spite of this it
is fair to say that a clear discrimination between columns is evident in these results.
However it is instructive and important to explore whether and how this obser-
vation is supported in the theory and simulation. Two mechanisms have already
been encountered which may serve to confuse the issue as regards the validity of
column-by-column resolution for EELS imaging. Firstly, some analyses which have
predicted significant delocalization of the interaction cast doubt upon the feasibil-
ity of column-by-column spectroscopy since a strongly delocalized interaction may
lead to overlap of signals from adjacent columns such that individual columns are
unresolvable. Secondly, it has been demonstrated, for example by Hillyard et al.
(1993), Plamann and Hÿtch (1999), and Dwyer and Etheridge (2003), that the
probe spreads considerably for thicker specimens, with the implication that column-
by-column spectroscopy may be questionable in those circumstances.
As regards the first point, the results of the previous section suggest the signals
are localized to a scale smaller than typical inter-column spacings. That approach
has subsequently been expanded and systematized by Cosgriff et al. (2005), and they
find similar conclusions over a wide range of elements. As regards the second point,
in the context of HAADF we have shown in section 5.1 that, within the validity of
the MDFF model, dynamical probe spreading does not strongly affect the reliable
interpretation of the images, because much of the signal derives from the top portion
of the crystal before the probe has spread significantly. Let us now further explore
these two issues for the case at hand.
Let us consider three probes. Probe 1 will contain no aberrations, and has
qmax = 1.0 Å−1 (i.e. 37 mrad). Probe 2 will be the aberration-balanced probe,
characterized by Cs = −0.05 mm, C5 = 63 mm, ∆f = 62 Å, and qmax = 0.539 Å−1
(i.e. 20 mrad), which set corresponds to the experiment of this case study. Probe 3
will be the aberrated probe, characterized by Cs = 0.5 mm and Scherzer conditions
(∆f = −497 Å and qmax = 0.378 Å−1 , or 14 mrad). The scaled intensity profiles
Chapter 5. Subtleties in qualitative and quantitative image interpretation 139
are shown in figure 5.8 (a), and it is seen that they are representative of a moderate
range of probe widths.
(a) (b)
1.0
1.0 probe 1
probe 2
Normalized intensity
probe 3 0.8
0.8
Probe intensity
0.6
0.6
0.4 0.4
0.2 0.2
0.0 0.0
0.0 0.4 0.8 1.2 1.6 2.0 4 5 6 7 8 9 10
Figure 5.8: (a) Intensity profiles of the three probes under consideration. Probe 1 is
an aberration-free probe with qmax = 1.0 Å−1 (i.e. 37 mrad), probe 2 is an aberration
balanced-probe with qmax = 0.539 Å−1 (i.e. 20 mrad), and probe 3 has Cs = 0.5 mm
and Scherzer conditions (∆f = −497 Å and qmax = 0.378 Å−1 , or 14 mrad). (b) Ti
L-shell EELS image of a single, isolated titanium atom using the probes of (a).
Figure 5.8 (b) shows, for the different probes, a titanium L-shell EELS line scan
for a single titanium atom. The detector semi-angle is 20 mrad and the energy
window is 40 eV above the ionization threshold. As in the previous section, it is
seen that the full width at half maximum of the image gets progressively broader
as we move from the first and finest probe to the third and broadest. Thus the
delocalization of the STEM image is affected by the properties of the probe. Again
we note that for the finest probe the image contains some structure.
In SrTiO3 viewed along the [100] zone axis, the mixed titanium and oxygen
columns are 1.95 Å distant from the nearest oxygen columns and 2.76 Å distant
from the nearest strontium columns. By contrast, the half-width at half maximum
of the single atom STEM images in figure 5.8 (b) are around 1.0 Å, significantly
smaller than the inter-column separations. This strongly suggests that there is no in-
principle reason why a titanium L-shell EELS image for the quoted detector should
not admit column-by-column resolution.
140
probe 1
probe 2
3 probe 3
experiment
N/NAVE
0 TiO Sr TiO Sr
2 4 6 8 10 12 14
Probe position (Å)
Figure 5.9: Simulated data plots showing the Ti L-shell EELS results in the full
nonlocal calculation for a perfect crystal of thickness 200 Å using the three probes
of figure 5.8 (a). The experimental data is also plotted. The parameters used in the
experiment and the simulation are described in text.
The behaviour observed in the single atom case, specifically that finer probes
lead to finer features in the STEM image, has carried over to the crystalline case.
Decreasing the width of the probe here leads to a decrease in the width of the column
peaks, and an enhancement of the contrast.
Such a result is not unexpected. It has been shown previously (Hillyard et al.,
1993; Rossouw et al., 2003) that the majority of the dynamical signal comes from
the initial portion of the crystal, where probe spreading and absorption have not yet
begun to dominate the behaviour of the wavefunction. We saw similar behaviour
Chapter 5. Subtleties in qualitative and quantitative image interpretation 141
signal to that depth. Figures 5.10, 5.11, and 5.12 show just such plots for the three
probes of interest.
(a) (b)
intensi–6ty
7.5 160
(×10 )
Thickness (Å)
6.0
4.5 120
per depth
Fractional
3.0
1.5 80
0.0
40 0 40
Th 80 1
ick 120 2
Å)
ne
ss 160
3 n(
4 itio
(Å 5 p os 0 1 2 3 4 5
) be
Pro Probe position (Å)
(c) (d)
ional
2.0 160
)
e fract–4
Thickness (Å)
0
1.6
(× 1
1.2 120
intensity
0.8
Cumulativ
0.4 80
0.0
40 0 40
Th 80 1
ick 120 2
ne Å)
ss 160
3 n(
4 itio
(Å 5 p os 0 1 2 3 4 5
) be
Pro Probe position (Å)
Figure 5.10: Ti L-shell EELS using probe 1, the signal strength per depth is shown
in (a) mesh and (b) grey scale plots. The cumulative signal, the sum of the per
depth contributions, is also shown in (c) mesh and (d) grey scale plots.
We noted previously in discussing HAADF signals that the layers within the ini-
tial portion of the crystal where the probe has had little chance to spread produce
greater signals than those for latter portions of the crystal. The signal per depth
plots show that this behaviour occurs in EELS imaging, and again oscillatory be-
haviour is seen in the signal per depth for increasing depths. Again this is somewhat
ironed out in the cumulative or integrated signal.
The per depth contribution from the first slice is effectively the single atom
result discussed previously [cf. figure 5.8 (b)]. The cumulative signal for 200 Å is,
of course, the total signal plotted in figure 5.9. The transition in form between the
single atom result and the full crystal result is accounted for in the form of the per
Chapter 5. Subtleties in qualitative and quantitative image interpretation 143
(a) (b)
ty
10.0 160
–6
)
intensi
0
8.0
Thickness (Å)
(× 1
6.0 120
per depth
Fractional 4.0
2.0 80
0.0
40 0 40
Th 80 1
ick 120 2
Å)
ne
ss 160
3
o n(
(Å
4 siti
) 5 po 0 1 2 3 4 5
be
Pro Probe position (Å)
(c) (d)
ional
2.5 160
)
e fract–4
Thickness (Å)
10
2.0
(×
1.5 120
intensity
1.0
Cumulativ
0.5 80
0.0
40 0 40
Th 80 1
ick 120 2
ne 3 ( Å)
ss 160 ion
(Å
4
osi t 0 1 2 3 4 5
) 5 p
be
Pro Probe position (Å)
Figure 5.11: Ti L-shell EELS using probe 2, the signal strength per depth is shown
in (a) mesh and (b) grey scale plots. The cumulative signal, the sum of the per
depth contributions, is also shown in (c) mesh and (d) grey scale plots.
depth contributions for larger depths. For example, the slight volcano-like feature
present in the single atom STEM image formed with probe 1 does not persist to the
full crystal, and the per depth plot shows that this is because this feature does not
persist to greater depth. Conversely, probes 2 and 3 did not produce a volcano-like
feature in the single atom image, but such features are present in the full crystal
results. The per depth images show that this is because there are depths such as
40 Å and 120 Å at which the signal strength at the column drops off substantially,
especially in relation to the signal at the same depth when the probe on the surface
is displaced slightly from the column location.
The detailed form of the plots in figures 5.10, 5.11 and 5.12 is the result of many
simultaneous and inter-related physical processes: the dynamical channelling, the
thermal absorption, and the nonlocal interaction potential. It is these complications
144
(a) (b)
ty
7.5 160
–6
)
intensi
10
Thickness (Å)
6.0
(×
120
4.5
per depth
Fractional
3.0
80
1.5
0.0
40 0 40
Th 80 1
ick 120 2
ne 3 (Å )
ss 160 on
(Å
4 siti 0 1 2 3 4 5
) 5
e po
P rob Probe position (Å)
(c) (d)
ional
2.0 160
)
e fract–4
Thickness (Å)
1.6
10
1.2 120
intensity (×
0.8
Cumulativ
0.4 80
0.0
40 0 40
Th 80 1
ick 120 2
ne 160 3 Å)
ss n(
(Å 5
4
s itio 0 1 2 3 4 5
) po
be Probe position (Å)
Pro
Figure 5.12: Ti L-shell EELS using probe 3, the signal strength per depth is shown
in (a) mesh and (b) grey scale plots. The cumulative signal, the sum of the per
depth contributions, is also shown in (c) mesh and (d) grey scale plots.
which we sought to avoid in considering single atom STEM images as a first estimate
for the resolution attainable in EELS. While figures 5.10, 5.11 and 5.12 do describe
the cumulative signal, going from the limit of a single atom to the limit of the
moderately sized crystal, they do not greatly illumine the cause of these changes.
One cause must, of course, be the dynamical channelling of the probe. In our
discussions about HAADF, where a local approximation was valid and so a direct
correspondence between signal strength and probe density could be assumed, we
related explicitly these oscillations to variations in the distribution of the dynamical
wavefunction. We have cautioned that this interpretation should not be carried
directly over to the EELS regime because not only the intensity but the phase of
the wavefunction enters into the nonlocal expression for the fractional intensity for
inelastic scattering. We may highlight this point by a useful simulation trick: we will
Chapter 5. Subtleties in qualitative and quantitative image interpretation 145
consider again the per depth contributions with probe 2 incident upon the specimen,
but now within the phase object approximation.
The phase object approximation consists of neglecting the propagation term: the
Laplacian in the Schrödinger equation, the diagonal elements of the structure matrix
in the Bloch wave formulation, or the free space propagator in the multislice. Thus
the intensity of the wavefunction cannot change by propagation, though it may by
absorption, from that incident upon the surface; the primary modification is in the
phase. If EELS could adequately be approximated by a local model then, assuming
the phase object approximation for probe propagation, we would anticipate little
variation in signal with depth into the specimen.
The result of this simulation is shown in figure 5.13, and the variation with thick-
ness is evident – the phase of the wavefunction plays a significant role in determining
the fractional intensity due to inner-shell ionization. The comparison between the
phase object approximation results of figure 5.13 and the full propagation results
of figure 5.11 is informative. Naturally they both agree as the thickness tends to
zero, since in the formal model propagation effects have had no chance to manifest
themselves. The agreement persists a little way, for about the first 30 Å, where the
phase object approximation has some validity. Even beyond this point the broad
general features of the two are similar: the images become volcano-like as the signal
strength on the column ceases to grow as rapidly as that at slight displacements
from the column. The differences include the oscillatory picking up of signal on
the column, evident in figure 5.11 (a), and that the total signal strength is lower
in the phase object approximation since an electron density initially displaced from
a column cannot be drawn towards that column without dynamical propagation.
Though the phase object approximation would never be applied to as great a thick-
ness as used for figure 5.13, this theoretical trick has demonstrated the significance of
the wavefunction phase in contributing to the nonlocal expression for the fractional
intensity.
(a) (b)
ty
10.0 160
–6
)
intensi
0
Thickness (Å)
8.0
(×1
6.0 120
h
Fractional
ept 4.0
2.0
per d
80
0.0
40 0 40
Th 80 1
ick 120 2
ne 3 ( Å)
ss 160 on
(Å
4
ositi 0 1 2 3 4 5
) 5 p
be
Pro Probe position (Å)
(c) (d)
nal
1.2 160
)
e fractio
–4
1.0
Thickness (Å)
1 0
0.8
(×
120
0.6
intensity
Cumulativ
0.4
0.2 80
0.0
40 0 40
Th 80 1
ick 120 2 Å)
ne
s 160
3 n(
s( 4 s itio 0 1 2 3 4 5
Å) 5 po
be
Pro Probe position (Å)
Figure 5.13: Ti L-shell EELS using probe 2 and the phase object approximation
for all propagation through the crystal, the signal strength per depth is shown in
(a) mesh and (b) grey scale plots. The cumulative signal, the sum of the per depth
contributions, is also shown in (c) mesh and (d) grey scale plots.
we found for single atoms combined with channelling along atomic columns. In this
section we will explore another method of combining these two aspects by considering
a topical experimental issue: the spectroscopic imaging of single atoms within a bulk
crystal.
The STEM imaging of single, isolated atoms presented in section 5.2 through
simulation was intended as a theoretical tool, a measure of the width of interaction
as a guide to the resolution attainable by STEM EELS techniques. However single
atom imaging has been experimentally demonstrated in STEM. Crewe et al. (1970)
provided one of the earliest demonstrations of STEM imaging in HAADF mode by
directly imaging individual uranium and thorium atoms on a thin carbon film.
More recent work in the HAADF mode has successfully imaged single atoms both
Chapter 5. Subtleties in qualitative and quantitative image interpretation 147
on and within a range of materials (Nellist and Pennycook, 1996; Voyles et al., 2002;
Lupini and Pennycook, 2003) with resolutions approaching 1 Å. The Z-contrast na-
ture of HAADF makes it most suited to identifying heavy atoms in relatively light
surrounds. It is not, by itself, a reliable identifier of atomic species and, depending
on the nature of the specimen, it may be insensitive to some impurity elements –
say if they had very similar atomic weights to adjacent atoms. Chemical identifica-
tion is more readily achievable by spectroscopic analysis. We have already empha-
sized the suitability of STEM to simultaneous HAADF imaging and spectroscopy,
but few results have been reported that utilize these techniques in conjunction to
provide the needed spectroscopic identification. Suenaga et al. (2000) used phase
contrast CTEM imaging in conjunction with STEM EELS to locate and identify
single gadolinium atoms. Their spatial resolution, or rather probe size, was of the
order of 6 Å, adequate for the sample they had prepared but larger than the inter-
column spacing in most materials. Scanning probe microscopies have demonstrated
the location and spectroscopic identification of single atoms on surfaces, but cannot
probe individual atoms within the bulk environment.4
Single atom analysis could prove to be a powerful tool in many areas of science
in which isolated atoms control macroscopic phenomena. Examples include the
origin of ductility or embrittlement in structural alloys, the nature of trap states at
semiconductor interfaces, and active sites in heterogeneous catalysts (Varela et al.,
2004).
Varela et al. (2004) presented results demonstrating the location and spectro-
scopic identification of a single lanthanum atom within a bulk CaTiO3 crystal. In
this section we will explore that result in some detail, presenting simulations which
both support the direct interpretation of the experimental results and further them
through semi-quantitative comparison between theory and experiment.
The experiment was carried out using a VG Microscopes’ HB501UX field emis-
sion STEM operating at 100 kV equipped with a Nion aberration corrector (Dellby
et al., 2001), on which a (lateral) spatial resolution of around 1.1 Å is achieved rou-
tinely. The sample consisted of a superlattice of alternating CaTiO3 and Lax Ca1−x TiO3
layers. The sample was constructed in laminar fashion via pulsed laser deposition,
resulting in a highly ordered structure with well-controlled concentrations of lan-
thanum. Further experimental parameters may be found in Varela et al. (2004).
Reference layers of La0.04 Ca0.96 TiO3 (i.e. x = 0.04 ± 0.002) of approximately
4
Single atom defect identification has also been accomplished in CTEM. For example, Jia et al.
(2003) compare intensities measured in a transmission electron microscope using negative Cs with
image simulation to identify oxygen vacancies in SrTiO3 .
148
23 Å thickness were grown. Layers with lower lanthanum concentrations were grown
with fewer pulses and were consequently thinner. Thus a layer formed from a single
pulse, padded out with pure CaTiO3 , corresponds to a lanthanum concentration of 1
atom per 60 nm2 projected along the axis of deposition. The STEM experiment was
performed from an orthogonal projection, and it is estimated that for a sample 10 nm
thick there is about a 5% probability of finding more than a single lanthanum atom
in a column containing at least one lanthanum atom. Figure 5.14 (a) is a HAADF
image showing a La0.002 Ca0.998 TiO3 layer adjacent to a La0.04 Ca0.96 TiO3 layer. The
Z-contrast nature of the image allows these regions to be distinguished because the
higher concentration of lanthanum in the latter leads to a slightly higher average
signal, evident in the figure in the slightly brighter region between the dashed lines.
The single bright spot in the La0.002 Ca0.998 TiO3 region, indicated by the thinner
arrow, is believed to be a single lanthanum atom. Figure 5.14 (b) shows line scans
across this bright spot in order to better quantify the difference to adjacent sites.
[110]
3200
2800
1 nm 2400
2000
0.0 1.0 2.0 3.0
nm
Figure 5.14: (a) HAADF image showing an individual La atom, marked by the thin
arrow. The reference La0.04 Ca0.96 TiO3 layer is marked with a thick arrow and the
interfaces are marked with dotted lines. (b) Averaged intensity profiles in the [100]
and [110] directions (top and bottom respectively). The image obtained with probe
current of approximately 100 pA over 10 s. (Varela et al., 2004)
Identifying such impurity atoms with HAADF can be difficult because the os-
cillatory nature of the evolving wavefunction, as seen in section 5.1 in relation to
the channelling discussion, means that the contribution to the HAADF image from
Chapter 5. Subtleties in qualitative and quantitative image interpretation 149
any given depth oscillates depending on the concentration of the probe at the col-
umn. However this problem is more likely to manifest in missing impurities which
are present, not falsely identifying dopants where none exist: the increased signal
evident in the line scans of 5.14 (b) is significant. Given the low probability of find-
ing two lanthanum atoms in a single column, and given that this level of contrast
discrepancy is seen on a number of other sites within material with this concentra-
tion of lanthanum, we may be fairly confident on visual grounds alone that these
are single impurity atoms. Similarly we identify that the lanthanum impurities sit
on calcium columns, identifiable as being slightly less bright than the combined ti-
tanium and oxygen columns. Having identified the lateral site of an impurity, let
us confirm that it is indeed lanthanum that is seen. Since the experimental work
was not carried out by the author, the reader is directed to Varela et al. (2004) for
a more detailed description of the EELS detection procedure. We will cover only
what is necessary to understand figure 5.15 and the parameters needed to perform
the associated simulations.
To obtain energy loss spectra, an EELS detector with entrance aperture of 8 mrad
was used. To obtain a signal associated with a given column in the HAADF image,
the image was magnified above the columns of interest, and an EELS spectrum
was recorded by scanning over the enlarged region. This ensures that the probe
remains upon the column because the HAADF image can be used for reference to
correct any drift. Such scans were performed on the calcium column believed to
contain a lanthanum impurity and on nearby columns. Figure 5.15 (a) shows the
HAADF image with white squares denoting the columns scanned. Column 3 is
the calcium column with the suspected dopant, columns 4 and 6 are pure oxygen
columns (virtually invisible in the HAADF image due to their low atomic number),
column 2 is a mixed titanium and oxygen column, and columns 1 and 5 are other
calcium columns. The resultant EELS spectra are shown in Figure 5.15 (b), and
the characteristic M4,5 lines of lanthanum, at 832 and 849 eV, are clearly visible
in spectrum 3, confirming that the probe situated on this column interacts with at
least one lanthanum atom. The total number of M4,5 excitations for this spectrum is
of the order of 104 , which is low but unsurprising given the small number of counts
associated with the excitations of a single atom. Similar treatment applied to other
suspected lanthanum sites yielded similar signals, again supporting the notion that
only a single impurity atom is being detected.
(a) (b)
0.5 nm
Intensity
5
4
1 6 3
2
3 4 5
2
1
820
820 850
850 880
880
E loss (eV)
Figure 5.15: (a) HAADF image with (b) EELS traces showing spectroscopic identi-
fication of a single La atom at atomic spatial resolution, with the same probe used
for figure 5.14. The M4,5 lines of La are clearly visible in spectrum 3 obtained from
the column identified by HAADF as likely to contain a La impurity (total collection
time 30 s). Other spectra from neighbouring columns show significantly reduced or
undetectable lanthanum signals. These spectra were obtained with collection times
of 20 s, and are shown normalized to the pre-edge intensity and displaced vertically
for clarity. (Varela et al., 2004)
the dynamical channelling of the probe, the spreading inside the crystal as the probe
evolves, or perhaps a combination of both these things. Our simulations allow these
possibilities to be investigated, and we will see that it is predominantly the probe
spreading interpretation that is supported. Good agreement will not constitute a
proof of this interpretation, but will constitute a strong piece of additional evidence
to support it.
To simulate this situation one further vital piece of information is needed: the
depth of the impurity atom. This is not given directly from the experimental results
obtained. But it may be obtainable indirectly. If our supposition of a single on-
column impurity is correct, then its depth must be such that the recorded signals
on the nearest neighbour columns agree with the measured ratios. The signal on
the oxygen column should be about 20% that on the calcium column; the signal on
the mixed titanium and oxygen column should be about 10% that on the calcium
column. Already this sounds encouraging because the lighter column, which has less
binding strength and is therefore expected to admit more probe spreading, gives the
larger contribution.
It should be presaged now that the fair agreement we will find depends on in-
cluding much of the experimental detail: realistic probe parameters, a realistic (i.e.
distorted perovskite) crystal structure, the finite scan region, probe instability, and
so forth. However such a range of considerations makes for lengthy calculations and
so in the first instance, while trying to estimate a reasonable depth for the dopant,
a much simplified model will be used. We will regard CaTiO3 as possessing a simple
cubic perovskite structure with lattice parameter a = 3.795 Å, and the probe as
being aberration-free with a probe-forming aperture of 25 mrad, which corresponds
to qmax = 0.676 Å−1 at 100 keV. Figure 5.16 (a) and (b) show line scans from the
calcium column across a nearest neighbour oxygen and a mixed titanium and oxygen
column respectively for a range of impurity depths.
It is interesting to note that if we take 4.0×10−7 as an estimate for the fractional
intensity on the calcium column then we should be able to predict the number
of counts. A 100 pA current for 30 s corresponds to 1.9 × 1010 electrons. By
the definition of the fractional intensity we therefore predict a signal of around
1.9 × 1010 × 4.0−7 = 7500 counts which is indeed on the order of the 104 counts
recorded. While this sort of quantitative analysis is strongly desirable, we will not
estimate the depth of the dopant by matching this number, but rather by looking
at the ratios of signals on the different columns.
The required 20% contribution on the oxygen column relative to the calcium
152
(a) (b)
7.6 7.6
Fractional intensity
Fractional intensity
34.2 Å
5.7 68.3 Å 5.7
102.5 Å
(×10 )
(×10 )
–7
–7
3.8 136.6 Å 3.8
1.9 1.9
0.0 0.0
Ca O Ca TiO
0.0 0.5 1.0 1.5 2.0 2.5 3.0 0.0 0.5 1.0 1.5 2.0 2.5 3.0
Figure 5.16: La M -shell EELS line scans from Ca column across to (a) the O column,
and (b) the TiO column, for a range of dopant depths. The column locations are
indicated by the the atom names shown below the plots.
column is obtained for an impurity depth around 100 Å. By 140 Å this fraction
Fractional intensity (×10 )
–7
9.5
has been exceeded; the signal for the probe near the oxygen
Ca column has risen,
the probe having had more 7.6 chance to spread off the oxygenO
column and out to
TiO
adjacent (and more strongly
5.7 binding) columns, while the signal from the probe atop
the calcium column has decreased, due to a combination of thermal absorption and
3.8
probe spreading off the column. Yet, whilst the peak on the calcium column spreads
out between the calcium and1.9 the mixed titanium and oxygen columns for increasing
probe depth, at no point0.0 in the range of dopant depths tested does the signal on
the mixed titanium and oxygen 0 column
20 40as 60a fraction
80 100of 120
that140
on the calcium column
increase significantly from zero, certainly not towards 10% even given the ±5%
Impurity depth (Å)
tolerance. We see this still more clearly in figure 5.17, which shows the signals from
just these three columns as a function of the depth of the dopant.
The signal on the calcium column rises out to around 50 Å as the probe is drawn
more tightly onto the column, but then decreases as absorption and probe spread-
ing reduce the electron density in the elastic beams in proximity to this column.
The signal on the oxygen column rises steadily, though is seen to be flattening out
towards 150 Å. The signal on the mixed titanium and oxygen column does not rise
appreciably over the range displayed.
Can we estimate an approximate depth for the impurity? Recall that the exper-
iment involved scanning over a finite region, a square cell approximately 1 Å in side
length centred at the atom. In addition there may be some instability in the posi-
Chapter 5. Subtleties in qualitative and quantitative image interpretation 153
3.8
1.9
0.0
0 20 40 60 80 100 120 140
Figure 5.17: La M -shell EELS signal as a function of impurity depth within the
Ca column with the probe on the Ca column, a nearest neighbour O column and a
nearest neighbour TiO column.
tion of the probe or equivalently some finite size to the source. These effects tend
to average the signal over a broader region than the single points or lines treated
in figures 5.16 and 5.17. Given that there is generally a peak about the calcium
column, averaging about the centre will tend to reduce the signal strength. Given
the form of figure 5.16 (a) it may raise the signal on the oxygen column for small
impurity depths but may lower it for greater depths. Given the form of figure 5.16
(b) it can only serve to raise the signal on the mixed titanium and oxygen column,
which would improve the agreement. Expecting this to improve the agreement from
the mixed titanium and oxygen column but to have less of an effect on the ratio be-
tween the calcium and oxygen column signals we will select from figure 5.17 a depth
which gets the ratio with the oxygen column about right. We will choose 120 Å, or
the nearest integer number of cells, as fitting this requirement. This choice is some-
what arbitrary, a range of depths either side would also have sufficed. We are not
seeking to represent exactly the experimental situation here, since there are factors
for which we will not account (such as the ion-milling preparation which may leave
amorphous surface layers), but rather to provide a motivation for plausibility. Nor
are we seeking to determine the depth resolution obtainable by comparison of ratios
alone, though the variation in calcium and oxygen column signals evident in figure
5.17 is quite suggestive about the possibilities of obtaining such a result from more
quantitative analyses.
It is not particularly enlightening to present here a range of simulations exploring
154
Table 5.1: Probe parameters. All quantities are given in units of nanometers.
We have selected an impurity depth of 120 Å, but the impurity is inside the
sample and so the crystal thickness is greater than this. Within the approximation
5
Though not, it should be admitted, recorded in the same experiment as the experimental data.
Again, we seek a motivation of plausibility, not an exact simulation of the experiment.
Chapter 5. Subtleties in qualitative and quantitative image interpretation 155
(a) (b)
(c) (d)
Figure 5.18: (a) HAADF image with aberration-free probe. (b) HAADF image
with aberration-balanced probe and 0.4 Å Gaussian broadening. (c) EELS image
corresponding to probe of (a). (d) EELS image corresponding to probe of (b).
No accounting for finite source size and position instability of the probe relative
to the specimen has yet been made. To a good first approximation these factors
can be simulated by convolving the STEM image with a Gaussian (Nellist and
Rodenburg, 1994). Figure 5.18 (b) shows the result of figure 5.18 (a) convolved
with a Gaussian of half-width 0.4 Å, which is a reasonable number and provides
much better agreement with the spot sizes in the experimental data. Figures 5.18
(c) and 5.18 (d) show the zone-axis EELS images corresponding to the HAADF
images of figures 5.18 (a) and 5.18 (b). The asymmetry of the structure is evident,
particularly where oxygen columns distorted towards the calcium column provide a
stronger signal than those distorted away from it.
6
Since the detector aperture is very small, 8 mrad, it is likely that this approximation is begin-
ning to break down. Nevertheless we will persist with it in this example.
156
The final experimental feature to be added is the integration over the finite,
square scan region. We will do this for both figures 5.18 (c) and 5.18 (d), that is to
say with and without the blurring effect. The results are summarized in figure 5.19.
2.3
3.0
21 38
23 37
3.0 3.1
100
4.0 3.8
27 11
28 13
4.5
5.5
Figure 5.19: Signal strength in raster scan regions as a percentage of that on central
Ca column. Top numbers exclude the blurring effect. Bottom numbers include the
blurring effect. The central column is the Ca column, the columns indicated by long
dashed squares are the O columns and the columns indicated by the short dashed
squares are the TiO columns.
It is evident from figure 5.19 that the blurring effect has the desired result of in-
creasing the signal on the surrounding columns relative to that of the centre column.
This is understood by comparing figure 5.18 (d) to figure 5.18 (c), the latter extend-
ing the signal which is predominantly situated on the calcium column further out
and so it may contribute to the integration region on adjacent columns. A diversity
of values are obtained on the different oxygen columns – because of the asymmetry
of structure and probe. Note that the oxygen column signal strength is on average
higher than that of the experiment, and the mixed titanium and oxygen column sig-
nal strength is on average lower. However for the blurred result the top left oxygen
column is comfortably within the 20% ± 5% range recorded in the experiment. The
proportional signal on the mixed titanium and oxygen columns are more consistent
and still rather on the small side. However the bottom most mixed titanium and
oxygen column has a signal nominally within the 10% ± 5% range recorded in the
experiment.
Thus, for the case including blurring, there are columns which agree with the
Chapter 5. Subtleties in qualitative and quantitative image interpretation 157
experimental results to within the error bars. While the parameters used may not
correspond precisely to those characterizing the experiment, the simulations have
elucidated the mechanism by which the signals are obtained. The simulations sup-
port the localization of the ionization interaction and demonstrate that the dy-
namical propagation, in transferring some electron intensity to adjacent columns, is
responsible for the signals obtained.
As a final note, it may be objected that, since we suggested that some of the
reduction of signal on the calcium column with increasing depth was due to ab-
sorption, our model may be inadequate to quantitatively analyse the single atom
case because it does not account for the contribution to the signal from electrons
which have undergone thermal scattering. We will revisit some of these calculations
in the next chapter after describing a model which does account for the thermal
scattering, and it will be found that though some quantitative difference exists it
is fairly small and does not contradict any of the interpretations made about the
mechanisms by which signals are obtained with the probe above columns adjacent
to that containing the impurity sought.
~ 16 Å
Fractional intensity (×10 )
–7
~ 33 Å
3 ~ 49 Å
~ 65 Å
~ 81 Å
2
0
0 20 40 60 80
|∆f| (Å)
Figure 5.20: K-shell EELS signals from a single P atom in Si, in the [100] orientation,
as a function of defocus. The different plots correspond to different depths of the
P atom, as given in the legend. The 100 keV probe is aberration-free with qmax =
1.0 Å−1 (i.e. 37 mrad), while the EELS detector has a semi-angle of 30 mrad and a
60 eV collection window above the ionization threshold.
Chapter 5. Subtleties in qualitative and quantitative image interpretation 159
5.6 Conclusion
The simulation theory developed in previous chapters has been used to explore issues
relating to the validity of direct interpretation of STEM imaging.
160
three-dimensional imaging were touched upon, though here the theoretical predic-
tions out-strip what is possible with the current generation of microscopes.
Through these studies we can make important contributions to topical questions,
such as the importance of cross-talk and the feasibility of column-by-column EELS
analysis, as well as demonstrating tools which may be of long term use in furthering
these questions and others. Though in some cases we have simply demonstrated
what is already the consensus amongst the STEM community, the importance of
testing and supporting such ideas by comparing simulation with experiment cannot
be overstated. Moreover we have seen that further information might be gleaned if
the comparison between theoretical and experimental results was to become more
quantitative.
162
¤£¡
ª@? ¡¢
R
R?
@ ª
¡
Chapter 6
©©
H
¼ ª?
¡ RH
@ j
The previous chapter used the simulation theory of earlier chapters to explore top-
ical questions relating to both theoretical and experimental aspects of scanning
transmission electron microscopy (STEM): the significance of cross-talk effects, the
prospect of column-by-column spectroscopic analysis and the location and chemical
identification of a single atom within the bulk. Experimental results were compared
with theoretical simulations to support the assertion that column-by-column spec-
troscopy is indeed possible. In the case of electron energy loss spectroscopy (EELS)
of a single atom in the bulk, we were able to support the direct interpretations made
from the images. Moreover, by making quantitative comparisons we were able to
extract still more information: the depth of the impurity. Such dividends rely on
being able to make meaningful quantitative comparisons between simulation and
experimental results.
In this chapter we will consider a final point made conspicuous by its evasion in
earlier chapters: the issue of the thermally scattered electrons. We have emphasized
repeatedly that the most significant outcome of this work is the adaptation of the
163
164
known that bright spots can arise in open columns due to the spreading of the probe
(Watanabe et al., 2001a,b; Yamazaki et al., 2002), that the dynamical channelling
can reduce the relative brightness of different columns from the value predicted by a
simple Z-contrast theory (Ishizuka, 2001, 2002), and that the dynamical behaviour
of the probe can strongly affect the signal from single dopant atoms at certain depths
(Voyles et al., 2004). In the EELS regime, Varela et al. (2004) obtained an M4,5 sig-
nature of lanthanum from the bulk of a CaTiO3 crystal for which there was good
reason to attribute this signal to a single impurity atom, but smaller signals were
also recorded on adjacent atomic columns. We discussed this result in section 5.4,
and simulations presented there were able to explain the contribution on adjacent
columns: they arose from the spreading of the probe as it propagated through the
crystal. Allen et al. (2003a) presented simulations of titanium L-shell EELS line
scans in SrTiO3 which suggested that column resolution was obtainable. It was also
suggested that, for very fine probes, a significant signal could arise when the probe
was situated on strontium columns. This was not because of elastic channelling but
because the strength of the scatterer led to an increase in the number of thermally
scattered electrons. These electrons, which according to the model used in that
paper are less localized to the site of the probe, were able to interact with titanium
atoms within the sample leading to additional, detectable ionization events. We dis-
cussed the experiment of that paper in section 5.3, but presented simulations using
only the dynamical contribution and thus this additional feature was not seen.
For investigating single atom impurities, we suggested that the signals on adja-
cent columns due to probe spreading, which at first glance seem to be a hindrance
to direct interpretation, may provide a method of extracting information about the
depth of the impurity. The anomalous peaks noted by Allen et al. (2003a) seem
less tractable, but it has since been shown that the model used for treating the
contribution from thermally scattered electrons – the diffuse contribution to which
the anomalous signal was attributed – is inadequate for the STEM regime.
As we have seen, the MDFF model, presented as an efficient method for cal-
culating inelastic STEM images, together with a new formulation of the boundary
conditions in the Bloch wave model allow for a generalization of a cross-section
expression for inelastic scattering which proved to accurately simulate channelling
maps obtained in the conventional transmission electron microscopy (CTEM) ge-
ometry (Oxley et al., 1999). In that context the diffuse contribution was essential in
obtaining excellent agreement with experiment. The diffuse term, modelled on the
assumption that thermal scattering produces a background of mutually incoherent
166
plane waves, had thus been tested against experiment in the CTEM regime, as well
as being well justified theoretically. In STEM, where the probe and hence the wave-
function are spatially localized to within a small volume, the assumption that the
diffuse background can be modelled by a set of mutually incoherent plane waves is
rather poor: it does not account for the expected variation in density of inelastically
scattered electrons with increasing distance from the probe.
This subtlety was lightly passed over in the initial adaptation of the MDFF
method to STEM (Allen et al., 2003b; Rossouw et al., 2003). For the case of large
probes it was supposed that the diffuse background was evenly spread over the basic
repeat unit of the crystal, a circumstance which would justify the form used for the
diffuse background as a first approximation. It was felt that a diffuse term would
be necessary given how important it had been in the CTEM geometry, though the
specimen thickness, and hence the proportion of thermally scattered electrons, is
typically much higher in many of the pertinent CTEM experiments. However that
form for the diffuse term proves to be inadequate when the size of the probe is
notably smaller than the repeating unit of the crystal.
It should be emphasized that it is only the form of the diffuse term in the MDFF
model as applied to STEM imaging that is being called into question. The dynam-
ical term, which in most circumstances hitherto considered is by far the dominant
contribution, is sound. This has been seen in the excellent quantitative compari-
son demonstrated between MDFF simulations and frozen phonon calculations for
HAADF images [cf. section 4.4, and Findlay et al. (2003)], and is further implicit
in the fair comparison between MDFF simulations and experimental EELS images
previously presented (cf. chapter 5).
The prevalent method for dealing with the detailed spatial distribution of both
elastically and thermally scattered electrons is the frozen phonon method (Loane
et al., 1991, 1992; Hillyard et al., 1993; Kirkland, 1998; Dwyer and Etheridge, 2003).
The frozen phonon method is able to simulate HAADF images directly, though in
the crystalline case is not as efficient as the MDFF Bloch wave method.1 Using the
frozen phonon model, Dwyer and Etheridge (2003) provided a method for estimating
EELS signal strengths by integrating the electron density within an appropriately
1
The significant factor in the calculation time for frozen phonon images is the averaging over a
range of configurations. To produce well-converged STEM diffraction patterns, it may be necessary
to average over around twenty configurations. (It will be seen later that to correctly obtain fine
detail in real space may require averaging over a significantly larger number of configurations.)
However HAADF images integrate over much of the fine structure in such patterns, and for HAADF
image simulation one may use fewer than four or so different configurations and still obtain good
results.
Chapter 6. Thermal scattering 167
chosen area about the atomic columns. With that notable exception, the majority
of EELS simulations are based on MDFF-style calculations (Dinges et al., 1995;
Allen et al., 2003b; Findlay et al., 2003; Rossouw et al., 2003). Thus, in order to
adequately account for the spatial distribution of inelastically scattered electrons
when simulating EELS, we present a synthesis of the frozen phonon model and the
MDFF method. We also present synthesis between the MDFF method and a variant
on the scattering function model of Anstis and co-workers (Anstis et al., 1996; Anstis,
1999), in a form which relates it to the models of Rose and co-workers (Dinges et al.,
1995; Hartel et al., 1996; Dinges and Rose, 1997; Müller et al., 1998). These methods
will be explored through re-simulation of some results of the previous chapter along
with some additional cases which highlight significant points. The relative merits
and limitations of the syntheses will be discussed in comparison to the previous
approach.
68 Å
Probe on Ca
–6 –2 –11 –2 –6 –3
0.54×10 0.15×10 0.62×10 0.15×10 –0.69×10 0.14×10
136 Å
Probe on Ca
–6 –2 –8 –2 –6 –3
0.77×10 0.20×10 0.18×10 0.18×10 0.20×10 0.26×10
68 Å
Probe on O
–6 –2 –9 –2 –5 –4
0.13×10 .12×10 0.40×10 .12×10 –0.51×10 0.56×10
136 Å
Probe on O
–6 –3 –10 –3 –5 –4
0.25×10 0.67×10 0.25×10 0.65×10 –0.27×10 0.31×10
Figure 6.1: Real space intensity images in CaTiO3 viewed along the [001] zone axis,
simulated using a 100 keV, aberration-free probe with aperture semi-angle 25 mrad
(as per section 5.4). The depths chosen and the column over which the probe is
located is listed on the left. The viewing region contains 2 × 2 unit cells of the form
shown in figure 2.4 (c), but the images are centred about the initial probe location.
The first column was simulated using the frozen phonon model (having averaged
over 100 configurations) and the second using a model which includes absorption.
The third column plots the difference between the first and second column images.
Minimum and maximum values are given below each image.
170
that they do not quite possess the symmetry expected for an axially symmetric probe
above positions of two-fold rotational (and reflection) symmetry. This is an artifact
of the calculation, indicating that insufficient phonon configurations have been taken
into account. Nevertheless, it is a persistent artifact: the images were averaged over
100 different frozen phonon configurations. For HAADF images it is well known
that twenty configurations is generally quite sufficient. Silcox and coworkers cite this
order number for HAADF (Loane et al., 1991; Hillyard et al., 1993). Kirkland (1998)
cites a similar number for well-converged diffraction patterns. Dwyer and Etheridge
(2003) use a slightly higher figure, and Muller et al. (2001) used 44 configurations
for simulations of diffraction patterns using realistic phonon dispersion curves. The
distortions in figure 6.1 are at the 10% level: not significant to the total intensity, but
defining to the difference images. This result could have been anticipated if we tried
a configurational average on the projected potentials – well over a hundred different
randomly generated phonon configurations must be averaged before one obtains a
result similar to that generated using Debye-Waller factors. The significant features
of the above images are clear enough that we have chosen to comment on this effect
rather than increasing the number of configurations in order to suppress it. In the
following section we will attempt to quantify the spread of the thermally scattered
electrons by integrating the intensity within a chosen radius about a chosen column.
The results of those calculations have essentially converged by 100 configurations,
indicating that the general distribution of the electrons in figure 6.1 is reliable.
In section 6.2.1 it was seen that the spatial distribution of thermally scattered elec-
trons, as determined by the difference in images simulated via the frozen phonon
model and the absorptive model, tended to be located about the scattering columns
on which the probe has some significant intensity in the elastic wavefunction. In the
few examples shown, the maximum value of the intensity of the difference tended to
be small relative to the intensity in the elastic wavefunction. This suggests that the
elastic wavefunction as simulated by the absorptive model may be a fair approxima-
tion to that simulated by frozen phonon methods. That the spatial distribution of
thermally scattered electrons often seems localized to the column which produced
them suggests that the largest differences in simulating EELS images with and with-
out accounting for the spatial distribution of thermally scattered electrons will occur
Chapter 6. Thermal scattering 171
(a) (b)
Fractional intensity (×10 )
10.0 1.0
–7
Frozen phonon
Integrated intensity
8.0 0.8 Absorptive
6.0 0.6
4.0 0.4
2.0 0.2
0.0 0.0
0.0 0.5 1.0 1.5 2.0 2.5 3.0 0 20 40 60 80 100 120 140
Probe position (Å) Thickness (Å)
(c) (d)
0.20 0.25
Ca Ca
Integrated Intensity
Integrated intensity
O 0.20 O
0.15
TiO TiO
0.15
0.10
0.10
0.05
0.05
0.00 0.00
0 20 40 60 80 100 120 140 0 20 40 60 80 100 120 140
r = 2.0 Å
r = 1.0 Å
3.795 Å
Ca
O
TiO
Figure 6.2: (a) Single atom STEM image for a La atom. (b) Integrated intensity
within a 1.0 Å radius of the Ca column with the probe centred on the Ca column.
(c) Difference in integrated intensity within a 1.0 Å radius of the Ca column between
the frozen phonon and absorptive models, with the probe centred on the columns
indicated. (d) Same scenario as for (c) using a 2.0 Å radius of integration. (e)
Projected structure showing the two integration radii. Imaging parameters are as
per figure 6.1. The frozen phonon simulation averaged over 100 configurations.
Chapter 6. Thermal scattering 173
(a) (b)
7.0
ce
Intensity differen
5.0 0.4
4.0 0.3
0.2
3.0
0.1
2.0
0.0
Int
1.0 eg 3 160
ra 2
0.0
t io 120
nr 1 80
ad 40 s (Å)
nes
0.0 0.5 1.0 1.5 2.0 2.5 3.0 ius
(Å ick
Probe position (Å) ) Th
(c) (d)
0.35 0.020
ce
ce
0.30
Intensity differen
Intensity differen
0.016
0.25
0.20 0.012
0.15 0.008
0.10
0.004
0.05
0.00 0.000
Int 3 0.8
eg Int
rat 2 160 eg 0.6 160
ion 120 ra 120
tio
rad 1 80 (Å) n r 0.4 80 (Å)
ius 40
kn ess ad 0.2 40
kn e
ss
c ius hic
(Å
) Thi (Å T
)
(e)
r = 4.0 Å
r = 2.0 Å
5.5 Å
r = 1.0 Å
O
TiO Sr
Figure 6.3: (a) Ti L-shell EELS image for a single atom. (b) Difference in integrated
intensity about the Sr column between the frozen phonon and absorptive models,
with the probe centred on the Sr column. (c) Difference in integrated intensity about
the TiO column between the frozen phonon and absorptive models, with the probe
centred on the TiO column. (d) Difference in integrated intensity about the TiO
column between the frozen phonon and absorptive models, with the probe centred
on the Sr column. (e) Projected structure indicating significant integration radii.
The 100 keV probe is aberration-free with qmax = 1.0 Å−1 (i.e. 37 mrad). The frozen
phonon simulations were averaged over 100 configurations.
Chapter 6. Thermal scattering 175
Figure 6.3 (c) shows the contribution about the mixed titanium and oxygen
column with the probe initially situated upon that column. This plot suggests that
an integrated intensity corresponding to about 10% of the electrons present in a
given slice arises quickly within 1.0 Å of the mixed titanium and oxygen column,
and this level of contribution is then maintained consistently to a significant depth.
It also suggests that twice as many electrons again reside in the thermal background
between about 1.0 Å and 4.0 Å of the column. This, given the relative area of these
regions, corresponds to a lower average electron density than at the column location.
However it does represent a significant migration of thermally scattered electrons
away from the column on which they were generated.
Figure 6.3 (d) shows the contribution about the mixed titanium and oxygen
column with the probe initially situated upon the strontium column. This is the
sort of plot to consider in a first attempt to predict whether electrons thermally
scattered from the strontium column will be in sufficiently close proximity to a
mixed titanium and oxygen column to be potentially able to cause titanium L-shell
ionization events. It should be noted that this represents the integration about
one mixed titanium and oxygen column, however symmetry gives that four such
columns are equally spaced about the strontium column. As such, the scale on this
plot should be mentally multiplied by four to predict the total possible contribution
to the titanium EELS signal. This plot displays more oscillatory behaviour than the
others. It we take 0.8 Å as a reasonable integration radius for EELS [gauged from
figure 6.3 (a)], then we conclude that there is a reasonable range of thicknesses for
which the integrated intensity contribution is around 3–4%, which is about half the
value at the same radius for the probe situated atop the mixed titanium and oxygen
column.
Recall that our examination of the distribution of thermally scattered electrons was
motivated by the idea that the contribution to the EELS images from thermally
scattered electrons could depend strongly on their spatial distribution. However
there is one significant impediment to simply making this deduction from the images
we have already simulated. We have made several references to the nonlocal nature
of the method for simulating EELS images. It is worth re-emphasizing here why an
increase in EELS signal when thermally scattered electrons are included does not
176
[cf. equation (4.16)]. In principle, the nonlocal kernel A(r, r0 ) can be calculated for
any given inelastic scattering event and detector geometry, and from this, and a
knowledge of the elastic wavefunction, the above equation can be used to calculate
the image.
It proves that for many inelastic signals of experimental interest (such as annular
dark field images and energy dispersive X-ray spectrometry) a local approximation,
in which [A(r, r0 ) − A∗ (r0 , r)] ∝ V 0 (r)δ(r − r0 ), suffices. For such effective interaction
potentials, equation (6.1) simplifies to
Z
I local
(R) ∝ |Ψ(R, r)|2 V 0 (r)dr (6.2)
[cf. equation (4.14)]. Interactions of this form lend themselves much more readily
to analysis. The effective interaction potential V 0 (r) can be plotted in real space,
and it follows from the form of the above equation that obtaining a signal requires
that there be appreciable local electron density |Ψ(R, r)|2 for at least some point r
in which V 0 (r) 6= 0. Thus for a local effective scattering potential, an increase in the
signal strength when the effect of thermally scattered electrons is taken into account
would have a direct interpretation in terms of electron density in the vicinity of the
non-zero regions of V 0 (r) . However the effective interaction potential for most EELS
Chapter 6. Thermal scattering 177
experiments retains its nonlocal character. Equation (6.1), or its reciprocal space
equivalent, must be used instead and this does not have a direct interpretation in
terms of local electron density.
where Idyn and Idiff denote the dynamical and diffuse terms respectively. The dynam-
ical contribution has been covered in detail in chapter 4 and we will not revisit the
details here. The diffuse term has a particularly simple structure: the term in square
brackets gives the proportion of electrons lost due to absorption. Comparison with
the normalization condition of equation (2.35) shows that in the absence of absorp-
R
tion the diffuse term is zero. In the presence of absorption V |ψ(R, r⊥ , z)|2 dr will
be less than unity, based on the extent to which absorption has removed electrons
from the elastic beams. The diffuse signal is the proportion of absorbed electrons
178
scaled by the mean probability of causing the inelastic scattering event described
by the µ-matrix elements. Such an approach must ultimately fail when localized
probes are used; in the extreme case of a spectroscopic signal deriving from a local-
ized impurity we do not expect an EELS signal to result when the probe is far away
from the impurity, irrespective of the amount of absorption due to thermal scatter-
ing. We will now describe methods by which we can more accurately account for
the contribution to the EELS images from thermally scattered electrons. In making
subsequent comparisons we will refer to calculations using only the first term in
equation (6.3) as dynamical calculations and to calculations using both terms as
dyn/diff calculations.
The elements µh,g (z) are the MDFFs at a given depth. Prior to this point we have
not explicitly included a depth dependence in the MDFFs, however we implicitly
did so to simulate the case of a single atom impurity in the bulk. The z-dependence
Chapter 6. Thermal scattering 179
indicates that, at least in principle, we might use different projected matrix elements
at different slices.
To adapt this expression to the frozen phonon method we define a total fractional
intensity which is an average over the fractional intensities for a variety of configu-
rations. We will use the index n to denote different configurations: ψn will denote
the wavefunction resulting from purely elastic scattering within a crystal in the nth
configuration and µnh,g (z) the MDFFs appropriate to that configuration. Then
N N Z
" #
1 X n 1 X t X n
Idyn (R, t) = I (R, t) = 2 µ (z)Ψ∗n (R, h, z)Ψn (R, g, z) dz ,
N n=1 dyn A N n=1 0 g,h h,g
(6.5)
in which Ψn is simply the Fourier transform of ψn . N denotes the number of different
configurations and must be large enough to ensure a converged calculation.
Various attempts, most notably by Wang (1998), exist to explain how the frozen
phonon model, which is conceptually a semi-classical model, is related to a rigorous
quantum mechanical theory. We will only offer the more intuitive semi-classical
reasoning here. The view is that electrons traverse the crystal much more rapidly
than the oscillation period of the atoms within the crystal, and so any given electron
“sees” only a frozen distribution of atoms. Thus the recorded experimental image is
built up of the images of several electrons which, for sufficiently large recording time,
all see different configurations of atoms. The MDFF model can be applied to a series
of wavefunctions resulting from the different configurations with the interpretation
that this gives a measure of the probability of ionization resulting from the single
electron interacting with the static arrangement of atoms.3
The incoherence between thermally scattered and elastically scattered electrons
is not incorporated into each individual wavefunction; no distinction is made in the
frozen phonon model at the level of individual wavefunctions. This false coherence
between the two types of scattering mechanisms is what the average over several
configurations is intended to remove (Ishizuka, 2002). This too is the conceptual
justification for the present work; the division between the two sorts of scattering
is not explicitly provided, but the average over configurations should yield a final
result which approximates the incoherent sum between the different contributions.
Thus far we have presented the equations necessary to simulate the images using
a direct synthesis between the frozen phonon model and the MDFF model. Frozen
phonon calculations do not use Debye-Waller factors for the elastic potential and
3
Fanidis et al. (1992, 1993) formalize a similar concept through the use of an ensemble average.
180
it would therefore be consistent not to use them in the MDFF elements either.
However to make the calculations tractable it is advantageous to make an approx-
imation. It will be appreciated that in the frozen phonon model, where for plane
wave incidence it is possible to scatter to angles other than the Bragg angles, it is
necessary to use many more Fourier coefficients of potential to obtain the same res-
olution. Put another way, by using a supercell which is internally non-periodic (due
to the “static” displacements) the reciprocal representation of the potential will fill
the array, rather than entries occurring only at discrete points as they would for an
internally periodic supercell. This breaking of periodicity results from the probable
asymmetry of any given frozen configuration of atoms. Formally, the MDFF ele-
ments should also possess this property. But this is problematic because calculating
the µnh,g (z) may be computationally challenging.4 This lack of periodicity also re-
moves the block-diagonal property whereby µnh,g (z) is only non-zero if h − g is a
physical reciprocal lattice vector [cf. equation (4.6)]. These two aspects complicate
the practical implementation of a full calculation using equation (6.5). To regain
these simplifying properties we propose that the MDFFs used should be those cor-
responding to the periodic structure and that they therefore include Debye-Waller
factors; we propose that the MDFFs used be exactly those used in the previous work
(Allen et al., 2003b).
To justify this approximation we reason as follows. The wave function will be
calculated separately for each different configuration as per a standard frozen phonon
calculation. As such, the distribution of electron density should, after the average
over several configurations, account for the spatial distribution of both the elastically
and thermally scattered electrons. It is this more realistic distribution that we seek
to include by this synthesis. Let us neglect correlations between the displacements
of the atoms: we will use an Einstein model.5 If we now constrain our attention
to a single slice within the crystal then the change to the wavefunction incident
upon this slice due to the elastic scattering within this slice is small. Therefore,
in calculating the contribution to the EELS signal from this slice, it is sufficient to
use the wavefunction from the previous slice prior to modification by the present
slice. That this is a good approximation is borne out by numerical testing; the
4
For K-shell EELS one might use a hydrogenic model, which would be more tractable, however
we seek to keep the current treatment as general as possible.
5
It has been shown that the inclusion of correlations changes the fine detail distribution of
electrons in the diffraction pattern (Muller et al., 2001) and presumably therefore has some effect
in the real space electron density distribution (Dwyer and Etheridge, 2003). However this will be
a secondary effect to that of obtaining a reasonably realistic distribution of thermally scattered
electrons on the scale of the repeating unit of the crystal.
Chapter 6. Thermal scattering 181
approximation is essentially that which says the multislice method gives the same
image whether the potential is projected on the front or back of each slice. The
consequence of this assumption is that the EELS signal arising from a given slice
depends only on the µ elements of that slice and on the wavefunction which reaches
that slice. Because we have neglected correlations, this wavefunction is completely
independent of the position we might choose to assign to the atoms in the present
slice. This dissociation at each and every slice between the wavefunction at that
slice and the MDFFs which can be used for that slice allows the configurational
average over these quantities to be performed separately. Thus equation (6.5) may,
to a good approximation, be rewritten as
N Z
( " M
# )
1 X t X 1 X m
Idyn (R, t) ' µ (z) Ψ∗n (R, h, z)Ψn (R, g, z) dz
A2 N n=1 0 g,h
M m=1 h,g
N Z t
( )
1 X X
= 2
µh,g (z)Ψ∗n (R, h, z)Ψn (R, g, z) dz , (6.6)
AN n=1 0 g,h
where µh,g without the superscript denotes the usual MDFFs, those including the
Debye-Waller factors which arise from performing the average over configurations
in the Einstein model.
By way of example let us consider a 100 Å thick slab of silver viewed along
the [001] zone axis from which EELS line scans along [100] will be taken using a
100 keV, aberration-balanced probe characterized by ∆f = 62 Å, Cs = −0.05 mm,
C5 = 63 mm, and qmax = 0.539 Å−1 (i.e. 20 mrad). The EELS signals will be
integrated over a detector semi-angle of 20 mrad and energy window of 40 eV above
ionization threshold. Figures 6.4 (a) and (b) show line scans for the silver M -shell,
with edge at 395 eV, and L-shell, with edge at 3.5 keV, respectively. The former is
more accessible experimentally than the latter. Plots are shown for the dynamical
model including absorption, the dynamical model without absorption and the frozen
phonon MDFF synthesis.
Comparing the dynamical results including absorption with those using the
frozen phonon model, we see the sort of differences anticipated: the signal using
the frozen phonon model is slightly higher due to the contribution of the thermally
scattered electrons, and more so about column locations where the degree of ab-
sorption is greater. However in both these models it is the shape of the M -shell plot
which is most striking, because the peaks are not on the columns, as they are for
the L-shell, but rather are between them.
182
(a) (b)
(c) (d)
slice
e ity per slic
8.0 2.5
sity per
7.0
2.0
6.0
(×10 )
(×10 )
5.0
Fractional inten–8
–6
1.5
Fractional intens
4.0
3.0 1.0
2.0 0.5
1.0
0.0 0.0
20 0 20 0
40 1 40 1
Th 60 2 (Å) Th 60 ( Å)
ick ion ick 2
iti on
ne
ss
80 3 osit ne
ss
80 3 os
(Å) p (Å ep
be ) ob
Pro Pr
(e) (f)
ity per slice
8.0 3.5
7.5 3.0
7.0 2.5
(×10 )
6.5
(×10 )
–6
Fractional intens
Fractional inten–8
2.0
6.0
5.5 1.5
5.0 1.0
4.5 0.5
4.0 0.0
20 0 20 0
(Å)
40 1 ) 40 1
Th 60 (Å Th
ic 2 on ick 60 2 ion
kn
es 80 3 ositi ne
ss
80 3 osit
s( p p
Å) be (Å be
Pro ) Pro
(g) (h)
0.006 0.0016
0.0012
Intensity
Intensity
0.004
0.0008
0.002
0.0004
20 20
Th 40 -4 Th 40 -4
ick 60 -2 i ck 60 -2
ne 80 0 ne 0
s s( 100 2 ) ss 80100 2 )
Å) 4 X (Å (Å
) 4 X (Å
Figure 6.4: EELS line scans for Ag (a) M -shell and (b) L-shell using the dynamical
model, the dynamical model without absorption, and the frozen phonon model.
(c) and (d) show the contributions per slice for the M -shell and L-shell results
respectively using the dynamical model. (e) and (f) show the contributions per slice
for the M -shell and L-shell results respectively using the dynamical model without
absorption. (g) and (h) show intensity plots across the wavefunction as a function
of depth for the probe on the Ag column and on the gap between the columns
respectively.
Chapter 6. Thermal scattering 183
Kirkland, 1998). For high-angle annular dark field imaging, which integrates over
much of the fine detail, this number may be reduced. How many configurations
are required for adequate convergence in simulating EELS images? Appreciably
under-converged calculations reveal themselves through some irregularity which is
inconsistent with expectations of symmetry and smoothness, though the converse –
that symmetry and smoothness indicate adequate convergence – is not guaranteed.
That said, the inelastic images depend most strongly on the low order Fourier coef-
ficients of the wavefunction, and as such the results are not exquisitely sensitive to
the finest features of the electron density.
Figure 6.5 indicates the degree of convergence obtained as the number of config-
urations in the average increases. It is seen that irregularities persist in the result for
ten configurations, but have been essentially removed by fifteen to twenty. Figure
6.5 (d) shows both frozen phonon and dynamical calculations, demonstrating the
importance of accounting for the thermal scattering: for a 100 Å thick silver sam-
ple almost a third of the L-shell signal strength is attributable to ionization events
caused by electrons which have, in the dynamical model, been absorbed.
(a) (b)
Fractional intensity
Fractional intensity
3.0 3.0
(×10 )
(×10 )
–7
–7 2.0 2.0
1.0 1.0
Ag Ag Ag
0.0 0.0
0 1 2 3 4 0 1 2 3 4
Probe position (Å) Probe position (Å)
(c) (d)
Fractional intensity
Fractional intensity
3.0 3.0
(×10 )
(×10 )
–7
–7
2.0 2.0
Figure 6.5: Ag L-shell EELS line scans (detector semi-angle 20 mrad and energy
window 40 eV above ionization threshold) using the aberration-balanced probe char-
acterized by ∆f = 62 Å, Cs = −0.05 mm, C5 = 63 mm, and qmax = 0.539 Å−1 (i.e. 20
mrad). The zone axis is [001], the scan is along [100]. The accelerating voltage is 100
kV. The crystal thickness is 100 Å. The average is taken over (a) five configurations,
(b) ten configurations, (c) fifteen configurations, and (d) twenty configurations. The
final image includes a dynamical model calculation for comparison.
of thermally scattered electrons may be significant when the probe is situated above
the calcium column.
In the simplest treatment the CaTiO3 structure may be regarded as a cubic (cf.
distorted) perovskite, with a single lanthanum atom substituting on a calcium site
at some depth into the crystal. As such, the z-dependence contained in equation
(6.4) and equation (6.6) gives that the fractional intensity calculation requires only
the form of the wavefunction in the vicinity of the impurity. We will perform two
types of simulation. The first pre-supposes a fixed depth for the impurity atom
and calculates an EELS line scan. The second, which we call a through-thickness
scan, plots the signal as a function of the depth of the impurity atom for a single
186
position of the probe. The former, for a suitable value of impurity depth, is the more
informative in interpreting the results of a given experiment. The latter is useful
in estimating the depth range in which to place the impurity atom to obtain fair
agreement between simulation and experiment. We saw such plots previously, in
figures 5.16 and 5.17. Figure 6.6 shows such plots for the frozen phonon model, with
the plots based on the absorptive model reproduced here for ease of comparison.
The zone axis is taken to be the [001] direction. Figures 6.6 (a) and 6.6 (b)
show line scans along the [100] direction, in which calcium and oxygen columns
alternate. The signal strength on the calcium column dominates that on the oxygen
column for small values of the impurity depth. For larger values of the impurity
depth this discrepancy decreases. Because this occurs in the absorptive code we
have already noted it to be primarily a channelling effect, not a result of thermal
scattering. For small values of the impurity depth the wavefunction has not spread
very much. Therefore the contribution with the probe on the calcium column,
in which the impurity sits, is large while the contribution with the probe on the
oxygen column is very small. This changes for larger values of the impurity depth
because the wavefunction resulting from a probe atop the calcium column has had
the opportunity to spread off the column, and so diminish the signal, while a probe
initially on the oxygen column, in spreading off this column, gives rise to a significant
electron density nearer to the impurity site. Comparing figure 6.6 (a) with 6.6 (b),
it is seen that the additional contribution from the thermal background picks up
the signal strength for the case of the probe initially on the calcium column but has
little effect for other positions.
Similar behaviour is seen in comparing figure 6.6 (c) with 6.6 (d), in which the
scan direction of [110] is that in which calcium and mixed titanium and oxygen
columns alternate. It will be noted that the contribution from the mixed titanium
and oxygen column is negligible at all depths considered here. In the absorptive
model this suggests that dynamical channelling of the elastic wavefunction does not
transfer significant electron density towards the calcium column. That the absence
of signal persists in the frozen phonon MDFF model suggests that the distribution
of thermal electrons does not give rise to any significant contribution either.
Figures 6.6 (e) and 6.6 (f) show through-thickness scans. The utility of such
plots is in estimating a reasonable depth for the impurity atom. In the experiment,
a 20% ± 5% relative signal (i.e. signal strength relative to that recorded on the
calcium column) was obtained on the oxygen column and a 10% ± 5% relative signal
was obtained on the mixed titanium and oxygen column. The simulations here
Chapter 6. Thermal scattering 187
–7
9.5 9.5
34.2 Å 34.2 Å
7.6 7.6
68.3 Å 68.3 Å
5.7 102.5 Å 5.7 102.5 Å
136.6 Å 136.6 Å
3.8 3.8
1.9 1.9
0.0 0.0
Ca O Ca O
-1.9 -1.9
0.0 0.5 1.0 1.5 2.0 2.5 3.0 0.0 0.5 1.0 1.5 2.0 2.5 3.0
–7
9.5 9.5
34.2 Å 34.2 Å
7.6 7.6
68.3 Å 68.3 Å
5.7 102.5 Å 5.7 102.5 Å
136.6 Å 136.6 Å
3.8 3.8
1.9 1.9
0.0 0.0
Ca TiO Ca TiO
-1.9 -1.9
0.0 0.5 1.0 1.5 2.0 2.5 3.0 0.0 0.5 1.0 1.5 2.0 2.5 3.0
–7
9.5 9.5
Ca Ca
7.6 O 7.6 O
TiO TiO
5.7 5.7
3.8 3.8
1.9 1.9
0.0 0.0
0 20 40 60 80 100 120 140 0 20 40 60 80 100 120 140
Figure 6.6: Plots derived from the La M -shell EELS signal from a single atom
impurity within CaTiO3 , using a 100 keV, aberration-free probe with semi-angle
25 mrad. The EELS detector semi-angle angle is 8 mrad. Figures (a), (c) and (e)
are simulated using the dynamical model, figures (b), (d) and (f) using the frozen
phonon method. Figures (a) and (b) are line scans along [100], with the Ca column
at the origin and the O column at 2.0 Å. Figures (c) and (d) are line scans along
[110], with the Ca column at the origin and the TiO column at 2.8 Å. The impurity
depths are given in the legend. Figures (e) and (f) are through-thickness scans with
the probe above the Ca, O and TiO columns respectively.
188
suggest that no depth gives the appropriate signal from the mixed titanium and
oxygen column; properly accounting for this required taking the finite scan region,
position instability, and the distorted structure into account. However the signal
from the oxygen column was primarily due to the probe spreading, and so we may
loosely estimate the depth of the impurity by comparing the through-thickness scans
for the calcium and oxygen columns. Using figure 6.6 (e) it may be seen that in
the absorptive model the relative signal strength is approximately correct for an
impurity depth in the range 80–120 Å. However in figure 6.6 (f), the frozen phonon
model, the appropriate impurity depth range is closer to 100–140 Å.
Thus the previous results may slightly underestimate the depth of the dopant
atom, since the effect of the frozen phonon model is to increase the signal from
the calcium column without a significant corresponding increase from the adjacent
columns – a greater depth is required for the calcium signal to drop and the oxygen
signal to rise. Note, however, that the discussion in section 5.4 about the mechanism
giving rise to the contributions on the different columns is sound.
model. The reason for this is two-fold. Firstly it relates directly to the model of Hall
and Hirsch (1965) upon which the majority of absorptive potential calculations are
based [though Anstis (1996) has presented modified scattering factors for the more
elaborate phase grating approach]. Secondly it will be seen that the resultant first
order terms are precisely those of the generalized multislice formulation of Rose and
co-workers (Dinges et al., 1995; Hartel et al., 1996; Dinges and Rose, 1997; Müller
et al., 1998) allowing a further useful comparison of models.
Hall and Hirsch (1965) used reasoning similar to that underpinning the later
frozen phonon model to obtain the absorption potential in the Einstein model.
They derived, in a kinematic approximation, an expression for the intensity of the
diffracted wave field arising from a lattice and performed the average over phonon
modes analytically. Their approach identifies a contribution to the scattered inten-
sity which derives from the Debye-Waller factor averaged projected potential, but
groups the remainder together to be treated as an absorption. In a phase grating
formulation Anstis (1999) took this approach further to obtain a description of the
total intensity distribution as the sum of scattering functions. In appendix G we
retain the kinematic approximation of Hall and Hirsch but use the concept of Anstis
and co-workers to express the result in terms of mutually incoherent scattering func-
tions. These functions all have the form of a product of an incident wave field and
an effective scattering potential. The key result is that the scattered intensity from
a single atom may be expressed as
∞ X
X n
hI(s)i = |F {Vn,k (r⊥ )ψ(r⊥ )}|2 , (6.7)
n=0 k=0
with elastic scattering factor fh , Debye-Waller factor M (h) = 2π 2 hu2p ih2 for pro-
jected mean square displacement hu2p i, R0 the mean atom position, and
s
[4π 2 hu2p i]n
An,k = . (6.9)
(n − k)!k!
190
Equation (6.7) shows that the intensity is made up of an incoherent sum of terms
which individually have the form of scattered wavefunctions. Note that the potential
in the zeroth order term, V0,0 , is the familiar thermally smeared elastic potential.
Note too that the first order terms are equivalent to the first two terms in the work
of Rose and co-workers (Dinges et al., 1995; Hartel et al., 1996; Dinges and Rose,
1997; Müller et al., 1998). A guide to this equivalence is provided at the end of
appendix G.
Using equation (6.7), the first order wavefunctions describing thermal scattering
from a single atom may be written as
Note that we have substituted a function φ(r⊥ ) where the derivation of equation
(6.7) should strictly have the elastic wavefunction incident upon the atom. This is
because, as we will presently show, to a good approximation, other wavefunctions
can be used for the purpose of increasing the efficiency of the calculation. Figure
6.7 (a) shows V1,1 (r⊥ ) for a silver atom. The entire crystal consists of many atoms
and with each of them we associate such a set of potentials. This is consistent
with the Einstein model where each atom vibrates independently. Thus incoher-
ence is included in two senses: each atom is treated independently and so thermal
scattering from different atoms is incoherent, and, by using two wavefunctions per
atom, the thermal scattering from a single atom is the incoherent sum of different
contributions. The elastic wave field incident upon each atom in the sample will ini-
tiate mutually incoherent scattered wavefunctions like those above. We will assume
that these wavefunctions will not undergo absorption thereafter. This is compara-
ble to the single elastic-to-inelastic scattering approximation made in most coupled
channels work (Coene and Van Dyck, 1990).
Rose and co-workers introduce random phases such that the incoherence between
thermally scattered wavefunctions for an individual atom, and between different
atoms, can be built in. We will not take that approach but rather treat each atom
independently at the outset. Anstis et al. (1996) note that the number of scattering
functions required for full convergence can be large. As it stands both these prop-
erties reduce the efficiency of the algorithm. However some simplifications can be
made.
Strictly, the wavefunction φ(r⊥ ) causing the excitation in equation (6.10) should
be the dynamical wavefunction ψ(r⊥ ). However this requires a calculation of the
propagation of ψa (r⊥ ) and ψb (r⊥ ) from every depth, since the evolution of the wave-
Chapter 6. Thermal scattering 191
function ψ(r⊥ ) causes it to change at each depth and thus produce different scat-
tering functions ψa (r⊥ ) and ψb (r⊥ ) at each depth. However the product of equa-
tion (6.10) gives a localized function because the scattering potentials V1,1 (r⊥ ) and
V1,0 (r⊥ ) are localized. Thus not all of the dynamical wavefunction is required, just
the portion in proximity to the scatterer. This suggests an interesting possibility:
if one could use the same function φ(r⊥ ) for all depths, which loosely assumes that
the form of the dynamical wavefunction close to the columns does not change signif-
icantly with depth, then one would only need to propagate ψa (r⊥ ) and ψb (r⊥ ) once
through the crystal for each distinct atom in the unit cell, the contribution from
atoms at all other depths being obtainable then from various stages of this single
propagation. For example, the evolution of scattering function ψa (r⊥ ) initiated two
thirds of the way through the crystal until it reaches the exit face will be identical
to the evolution of that same scattering function from the top surface to a third
of the way into the crystal. We will consider two possible forms for this general
approximation to φ(r⊥ ): that of a plane wave and that of the incident STEM probe.
The former assumes the dynamical wavefunction is effectively constant within the
region where V1,1 (r⊥ ) and V1,0 (r⊥ ) are significant, the latter accounts for a drop
in intensity away from the atom location and may be more appropriate for tightly
localized probes.
These approximations would give poor results if it were not for one further
modification which moreover allows us to limit the number of scattering functions
required. Equation (6.7) implies an absolute scale: the product of the effective
scattering potential and the incident wavefunction gives the scattered wavefunction.
However proceeding in this fashion using only two scattering functions does not
conserve electron flux, and the consequences are more pronounced if φ(r⊥ ) is treated
approximately since its normalization is not given. In order to use just two scattering
functions, the first order terms, and yet get a better estimate for the magnitude of
the contribution, we renormalize the electron density in these scattered functions
such that is equal to that absorbed from the elastic wavefunction. It is this step
that puts the true dynamically evolving wavefunction into the calculation, even if
a plane wave is used as the effective exciting function φ(r⊥ ), through the degree of
absorption to which the true wavefunction is subject.
Figure 6.7 (b) shows the fraction of the incident flux absorbed in one repeat
distance as a function of depth with the 100 keV, aberration-balanced probe, char-
acterized by ∆f = 62 Å, Cs = −0.05 mm, C5 = 63 mm, and qmax = 0.539 Å−1
(i.e. 20 mrad), as used in figure 6.5, positioned on an atomic column in [100] silver.
192
Figure 6.7 (c) shows the total EELS signal produced by the wavefunctions initiated
by a single atom at a given depth. The plots that do not include renormalization
to the proportion of electrons absorbed are seen to give a significantly lower signal
strength than those that do. We will presently show that it is the renormalized
signals which give good agreement with the frozen phonon model. Provided the
renormalization is performed, the correction obtained when higher order terms in
equation (6.7) are included is negligible. That the first order potential terms are
antisymmetric suggests that for very fine probes it will be necessary to take higher
order terms into account, however none of the probes and simulations described
herein require this.
If the effective probe is a plane wave, or at least the variation of the wavefunction
is sufficiently slow across the effective scattering potential that it may be approx-
imated as such, then the scattered wavefunctions arising from equivalent atoms in
equivalent cells are identical and it suffices to calculate the contribution only once for
each different atom in the unit cell for the entire crystal. When this approximation
is justified we can use this model to calculate STEM EELS images which account for
the contribution from thermally scattered electrons very efficiently because the ther-
mally scattered components need only to be run once (cf. the frozen phonon model
where the average over configurations must be performed for each and every probe
position). This is in the spirit of the “β-approximation” of Anderson et al. (1997)
which improved the efficiency of calculating HAADF images by an approximation
to the approach of Dinges et al. (1995).
The validity of these approximations is best tested through comparison with
more detailed calculations. Such a comparison is presented in figure 6.8. It is seen
that both approximations are in good agreement with the frozen phonon model.
Taken as a correction to the dynamical model, it certainly provides an excellent
indication of where the contribution from thermally scattered electrons should not
be neglected.
Consider three 100 keV probes. The first is aberration-free, with qmax = 1.0 Å−1
(i.e. 37 mrad). The second is the aberration-balanced probe, with ∆f = 62 Å,
Cs = −0.05 mm, C5 = 63 mm, and qmax = 0.539 Å−1 (i.e. 20 mrad), pertaining
to the EELS experiment described by Allen et al. (2003a). The third is aberrated,
Chapter 6. Thermal scattering 193
(a) (b)
Absorbed fraction
40 0.06
Potential (eV)
20 0.04
0
0.02
-20 3
-40 2 0.00
Å)
0 20 40 60 80
x(
2 1
y (Å) 1 Depth (Å)
(c)
Fractional intensity (×10 )
6.2
–9
1.6
0.0
0 20 40 60 80
Figure 6.7: (a) V1,1 (r⊥ ) for a silver atom [V1,0 (r⊥ ) has the same form, rotated by 90
degrees]. (b) The number of electrons absorbed in each repeat distance as a function
of depth, rated as a fraction of incident flux with the aberration-balanced probe used
in figure 6.5 situated above a silver column. (c) The total EELS signal produced
by the thermal channel wavefunctions initiated at the depth shown. The effective
wavefunctions φPW and φSTEM denote that the effective probe is a plane wave and
the STEM probe at the surface, respectively, while ψ(z) denotes that the effective
probe is the elastic wavefunction at the depth of the atom. “Renorm.” denotes that
the results have been renormalized to force conservation of electrons.
3.0
Fractional intensity
(×10 )
–7
2.0
Frozen phonon
1.0 Dynamical
Scattering function, φSTEM
Scattering function, φPW
0.0
0 1 2 3 4
Probe position (Å)
Figure 6.8: Ag L-shell EELS line scans simulated with the aberration-balanced probe
characterized by ∆f = 62 Å, Cs = −0.05 mm, C5 = 63 mm, and qmax = 0.539 Å−1
(i.e. 20 mrad). The frozen phonon and dynamical simulations are provided for
reference. The two approximations, φSTEM (r⊥ ) being the STEM probe and φPW (r⊥ )
being a plane wave, are seen to be in good agreement with the results of the frozen
phonon model.
indeed with the aberration-free probe situated on the strontium column almost half
of the incident flux has been absorbed by 50 Å. It is unsurprising then that a model
which treats the contribution to the EELS signal from thermally scattered electrons
as proportional to the fraction of absorbed electrons will give significant signal for
a probe situated on the strontium column, even if it is titanium L-shell ionization
being measured. It is also easy to see why a realistic account of thermal scattering
is needed: the thermal channels may contain a significant portion of the electron
density.
We will simulate EELS scans based on titanium L-shell ionization, with a de-
tector semi-angle of 20 mrad, and an energy window of 40 eV above the ionization
threshold. The zone axis is [001]. The thickness is 200 Å. The line scans are taken
along the [110] direction in which strontium columns alternate with mixed titanium
and oxygen columns. Figure 6.9 shows the fractional intensity calculated in the
four models being discussed. The scattering function calculation, in which φ(r⊥ ) is
approximated by the STEM probe wavefunction incident upon the entrance surface,
is again seen to be in good agreement with the frozen phonon model. Using a plane
wave for φ(r⊥ ) does not alter this result on the strontium columns, but gives a signal
on the mixed titanium and oxygen column more similar to that of the absorptive
model; the agreement between the two approximations in this case is less good than
it was for silver (figure 6.8).
The difference between the scattering function and dynamical models is purely
due to the additional electron density included within the scattering functions. As
such we will regard the dynamical simulation as the reference against which we judge
what additional effect is introduced when the distribution of thermally scattered
electrons are taken into account.
There is a notable difference in figure 6.9 between the frozen phonon simulation
and the dynamical simulation at the location of the mixed titanium and oxygen
column due to ionization events caused by electrons in the thermal background.
Given that for the aberration-free probe about 50% of the electron density consists
of thermally scattered electrons by 100 Å, the contribution from thermally scattered
electrons is relatively small. This is because the thermal scattering functions do
not contribute to the inelastic signal coherently. Note too that the difference is
less pronounced the broader the probe, correlating well with the smaller degree of
absorption occasioned by these probes (cf. table 6.1).
The proportion of electrons absorbed when the probe is on the strontium column
is even greater than when it is on the mixed titanium and oxygen column, and yet in
absolute terms the difference between the frozen phonon and dynamical simulations
is very small around the strontium column. Thus while the proportion of intensity
absorbed is large, the distribution therein does not give a significant contribution to
the titanium L-shell EELS signal.
The dyn/diff model, that is to say the calculation using both terms in equation
(6.3) and so treating the diffuse contribution based purely on the proportion of
electrons absorbed, is seen to over-estimate the thermal contribution to the mixed
titanium and oxygen column and greatly over-estimate the thermal contribution to
196
(a)
2.4
2.0
1.6
1.2
0.8
0.4
0.0
-0.4
0 1 2 3 4 5
2.0
1.5
1.0
Frozen phonon
0.5 Dynamical
Dyn/diff
0.0 Scattering function
0 1 2 3 4 5
Figure 6.9: Ti L-shell EELS images in SrTiO3 using (a) the aberration-free probe,
(b) the aberration-balanced probe, and (c) the aberrated probe. The solid line is
calculated from equation (6.6), the result of the frozen phonon MDFF synthesis.
The dashed line is calculated in the dynamical model [i.e. using only the first term
in equation (6.3)], while the dash-dot line is calculated in the dyn/diff model [i.e.
using both terms in equation (6.3)]. The triangles are calculated from the scattering
function model with φ(r⊥ ) approximated by the STEM probe wavefunction incident
upon the entrance surface.
Chapter 6. Thermal scattering 197
the strontium column relative to the frozen phonon model, though the discrepancy
decreases for broader probes. That said, the difference between simulations with and
without accounting for the distribution of thermally scattered electrons is notable
about the mixed titanium and oxygen column: thermal scattering is important and
there will be a need in quantitative work for modelling this detail.
In section 6.2.2 we compared the spatial distribution of electrons predicted in
the frozen phonon model to that in the absorptive model. In figure 6.3 we compared
the integrated electron density within the vicinity of columns of interest for probe
1 incident upon SrTiO3 . Figure 6.3 (c) showed the difference in density for the
two models at the mixed titanium and oxygen column with the probe above this
column, and the contribution within a radius of 1.0 Å for the full range of depths was
around 5–15%. Figure 6.3 (d) makes a similar comparison at the mixed titanium
and oxygen column but with the probe atop an adjacent strontium column. To
make fair comparison we should multiply the result by four, since there are four
mixed titanium and oxygen columns equidistant from the strontium column. We
may then estimate that the contribution for the full range in this case is around
1–5%. These numbers are not hugely different, and yet the difference between the
frozen phonon EELS signal and the absorptive EELS signal in figure 6.9 (a) is much
greater, in absolute terms, at the mixed titanium and oxygen column than it is
at the strontium column. This is further evidence that the electron density alone
should only be treated as a guide to estimating the signal strength: other factors
influence the detailed structure of an EELS image simulated via the full nonlocal
model.
6.4 Conclusion
It is the author’s experience that, for moderate probe sizes, thin crystals, or crystals
composed entirely of fairly light atoms, the degree of absorption due to thermal
scattering is fairly small. In such cases the refinement offered by the techniques we
have been describing is minimal and the extra computational complexity of the new
approach is unwarranted.
For cases where there is a significant degree of absorption, which may, for ex-
ample, result from strongly scattering samples or from very finely focused probes,
the new methods represent a notable improvement on the previous approach. The
new methods are also more important for simulating EELS images arising from the
presence of only a few impurity atoms because in such a circumstance the signal is
198
Chapter 7
©©
¼ ªH
¡ ?
RH
@ j
This thesis has been concerned with simulating coherent and incoherent images
as obtained in scanning transmission electron microscopy (STEM), and using the
tools developed to explore some topical questions about the reliability of direct
interpretation of images and the prospects for column, and indeed atom, location
and chemical identification.
By applying boundary conditions for the incident probe as a complete entity we
have derived excitation amplitudes of each Bloch state as a function of the probe
position and of the coherent aberrations in the lens. It has been shown that the
structure matrix, or A-matrix, the eigen-structure of which embodies the solution of
the Schrödinger equation, may be block diagonalized in the case of perfect crystals.
This leads to a very efficient calculation for the wavefunction, but more importantly
gives considerable insight into the physics of the coupling between the probe and
the electron wavefunction within the crystal. It has been seen that plane wave
components in the incident probe which differ by what have been called physical
reciprocal lattice vectors can profitably be treated together, while those which do
not differ by such vectors do not interact within the crystal.
This conceptualization led to a proof of the equivalence with the prevalent
method for the boundary conditions (the coherent superposition of phase-linked
plane waves), and simple derivations of some well-known properties of STEM images,
including properties of coherent bright-field images and the reciprocity relationship
with conventional transmission electron microscopy (CTEM). Using coherent images
199
200
in the diffraction plane we described a means of solving the phase problem to obtain
the scattering matrix and so effect a quantitative inversion procedure to determine
the projected potential which is able to accommodate multiple scattering and does
not require that the thickness is known.
The key result of the thesis is the cross-section expression for inelastic or inco-
herent imaging, making use of the mixed dynamic form factor (MDFF) concept.
Following readily from the new formulation for the boundary conditions, it was de-
rived in a Bloch wave model by a direct generalization of the cross-section expression
previously used in the CTEM geometry. Recast into the form of a fractional inten-
sity, thus providing more directly the link to the quantities measured experimentally,
this allows for the simulation of a variety of inelastic signals routinely recorded in
STEM, such as high-angle annular dark field (HAADF) images and energy disper-
sive x-ray spectroscopy and electron energy loss spectroscopy (EELS) images derived
from inner-shell ionization.
The dividends of the block diagonalization used in determining the wavefunction
in crystals carries over to an efficient implementation of the inelastic imaging expres-
sion in the Bloch wave model. However not all materials of interest are crystalline.
To handle a broader range of scientifically and technologically relevant materials,
the expression was recast into a form amenable to evaluation in the multislice tech-
nique. The numerical agreement between the Bloch wave and multislice methods
for calculation of incoherent lattice contrast in the MDFF method was presented.
The excellent agreement between these approaches ensures that the particular com-
putational technique may be selected based on the problem in question. For fairly
simple structures, where the dependence on parameters such as thickness, spherical
aberration and defocus are of interest, the Bloch wave method is most apt; it allows
ready simulation of all requisite images while only solving for the Bloch states once.
Conversely, for more complex structures, particularly structures with defects, the
multislice method may be preferred. In such cases it is computationally simpler; the
Bloch wave method becomes cumbersome when block diagonalization is no longer
possible and the number of beams involved is large.
Having established the underlying theory and simulation methods we discussed
how they might be used to look at issues relating to the imaging of structures. In the
context of HAADF imaging we investigated the effect of cross-talk, whereby signals
derived from the probe above one column may be contaminated by nearby columns
in a manner hindering direct interpretation, and found this effect to be small for
moderate thicknesses and fine probes. For the particular case of ZnS, a perceptible
Chapter 7. Summary and Conclusions 201
contribution to the HAADF signal occurs when the probe is situated on a sulfur
column from not only the neighbouring zinc column, but also from more distant
columns. Whilst the effect is small here, it does show that the quantitative inter-
pretation of experimental incoherent lattice images on a column-by-column basis
should be approached with some caution.
In the context of EELS we presented single atom STEM imaging as a theoretical
tool for estimating the delocalization of the interaction by assessing the localization
of the image. We demonstrated that it is currently the width of the probe, rather
than the nature of the ionization interaction, which limits the width of single atom
features in EELS, even for the lighter atoms. This motivates the drive for aberration
correction, implying that finer probes will indeed yield finer features, and supports
the idea that column-by-column spectroscopy is not precluded by the nature of the
ionization interaction.
Through a combination of simulation and experimental results we demonstrated
that the probe channelling did not prevent column resolution in EELS either, though
simulation was a very useful adjunct to experiment to fully interpret the structure
of the results. We investigated the location and chemical identification of a single
atom in with the bulk. Through quantitative comparison between simulation theory
and experimental data we were able not only to support the interpretation of the
experimental result but also to advance it by obtaining an estimate for the depth
of the impurity atom. The prospect for such three-dimensional investigations was
discussed.
To that point our simulations had been based on the contribution to the inelastic
images from electrons scattered out of the elastic beams. However for fine STEM
probes, which by intention are tightly focused on atomic columns, there may be
significant attenuation of the elastic beams by thermal scattering and it is therefore
of importance in quantitative EELS studies to gauge and understand the contri-
bution to inelastic signals from electrons which had previous undergone inelastic
thermal scattering. To this end we provided a means of combining the expression
for the fractional intensity due to inelastic scattering utilizing MDFFs with both
the frozen phonon model and the scattering function model, two techniques in the
literature which provide a realistic spatial distribution of the thermally scattered
electrons. It was concluded that for the finest currently available probes focused on
thick crystals or crystals containing heavy elements that the additional contribution
from thermally scattered electrons is important in quantitative work, though the
qualitative shape of the images in the samples explored was not greatly affected.
202
for novel imaging modes in which inelastically scattered electrons can be collected
to form images in the image plane for a range of positions of the incident STEM
probe (Möbus and Nufer, 2003).
Though many useful scientific investigations have been and will continue to be
done with STEM at a qualitative level, the common thread running through all the
future prospects described above is the need for quantitative comparison with theory
in order to extract the maximum amount of useful information from experiments.
The theoretical tools developed in this thesis, and the examples used to demonstrate
their consequences, are primarily a contribution towards making the quantitative
comparison between theory and experiment in STEM routine.
204
Bibliography
205
206
D. M. Bird and Q. A. King. Absorptive form factors for high-energy electron diffrac-
tion. Acta Cryst., A46:202–208, 1990.
A. Bleloch and A. Lupini. Imaging at the picoscale. Materials today, 7:42–48, 2004.
M. Born and E. Wolf. Principles of optics. Cambridge University Press, 7th edition,
1999.
J. M. Cowley and P. M. Fields. Dynamical theory for electron scattering from crystal
defects and disorder. Acta Cryst., A35:28–37, 1979.
P. Geuens and D. Van Dyck. The S-state model: a work horse for HRTEM. Ultra-
microscopy, 93:179–198, 2002.
Bibliography 211
P. Hartel, H. Rose, and C. Dinges. Conditions and reasons for incoherent imaging
in STEM. Ultramicroscopy, 63:93–114, 1996.
S. Hillyard and J. Silcox. Detector geometry, thermal diffuse scattering and strain
effects in ADF STEM imaging. Ultramicroscopy, 58:6–17, 1995.
C. J. Humphreys. The scattering of fast electron by crystals. Rep. Prog. Phys., 42:
1825–1887, 1979a.
K. Ishizuka. A practical approach for STEM image simulation based on the FFT
multislice method. Ultramicroscopy, 90:71–83, 2002.
R. F. Loane, P. Xu, and J. Silcox. Incoherent imaging of zone axis crystals with
ADF STEM. Ultramicroscopy, 40:121–138, 1992.
D. A. Muller, Y. Tzou, R. Raj, and J. Silcox. Mapping sp2 and sp3 states of carbon
at sub-nanometre spatial resolution. Nature, 366:725–727, 1993.
M. P. Oxley and L. J. Allen. Atomic scattering factors for K-shell and L-shell
ionization by fast electrons. Acta Cryst., A56:470–490, 2000.
M. P. Oxley and L. J. Allen. Atomic scattering factors for K-shell electron energy-
loss spectroscopy. Acta Cryst., A57:713–728, 2001.
T. Plamann and M. J. Hÿtch. Tests on the validity of the atomic column approxi-
mation for STEM probe propagation. Ultramicroscopy, 78:153–161, 1999.
P. G. Self and M. A. O’Keefe. Calculation of diffraction patterns and images for fast
electrons. In P. R. Buseck, J. M. Cowley, and L. Eyring, editors, High resolution
transission electron microscopy and associated techniques, pages 244–307. Oxford
University Press, 1988.
D. Shindo and K. Hiraga. High resolution electron microscopy for materials science.
Springer, Tokyo, 1998.
J. Silcox, P. Xu, and R. F. Loane. Resolution limits in annular dark field STEM.
Ultramicroscopy, 47:173–186, 1992.
S. Van Aert, A. J. den Dekker, D. Van Dyck, and A. van den Bos. Optimal experi-
mental design of STEM measurement of atom column positions. Ultramicroscopy,
90:273–289, 2002.
D. Van Dyck and J. H. Chen. Towards an exit wave in closed analytical form. Acta
Cryst., A55:212–215, 1999.
A. P. Young and P. Rez. Resonance errors and partial coherence in the inelastic
scattering of fast electrons by crystal excitations. J. Phys. C: Solid State Phys.,
8:L1–L7, 1975.
Appendix A
©©
¼ ªH
¡ ?
RH
@ j
List of acronyms
225
226
¤£¡
ª@? ¡¢
R
R?
@ ª
¡
Appendix B
©©
¼ ªH
¡ RH
@
? j
We adopt the convention whereby the free space plane wave propagating in the
direction of k is denoted by
and the inverse Fourier transform, i.e. from reciprocal space to real space, is given
by Z
−1
F [F (q)] ≡ f (r) = F (q) exp (+2πiq · r)dq . (B.3)
where g is a reciprocal lattice vector and fi is the atomic scattering factor for the
atom i at site ri .
227
228
1
Rt
where the projected potential φ(r⊥ ) = t 0
φ(r⊥ , z)dz is positive and attracts elec-
trons, and σ is a constant.
The free space propagator may be written as
In practice the effects of the astigmatic terms are directly visible and may be cor-
rected manually. Therefore we often neglect these. Making the identifications
∆f = C10 , Cs = C30 , C5 = C50 , we thus write
1 1 1
χ(q, λ) = ∆f λ2 q 2 + Cs λ4 q 4 + C5 λ6 q 6 , (B.9)
2 4 6
Appendix C
©©
¼ ªH
¡ ?
RH
@ j
Recall that the expression for the fractional intensity from inelastic scattering is
X X
I(R, t) = t B ij (R, t) Cgi Chj∗ µh,g , (C.1)
i,j g,h
where
exp[2πi(λi − λj∗ )t] − 1
B ij (R, t) = αi (R)αj∗ (R) . (C.2)
2πi(λi − λj∗ )t
We now recast this expression into matrix form. For clarity we will adopt the
notation Cg,i ≡ Cgi ∈ C, the matrix of eigenvectors. We also define Bi,j ≡ B ij (R, t) ∈
B and µh,g ∈ U , where B and U are n2 N ×n2 N matrices (based on an n×n supercell
using N physical reciprocal lattice vectors). In this notation the first subscript labels
the rows and the second subscript the columns of each matrix. We will also use the
superscript T to represent the transpose of these matrices and their related elements.
T
For example Cg,i = Ci,g .
229
230
We reason as follows
X X X X X
∗ ∗
Bi,j Cg,i Ch,j µh,g = Bi,j Ch,j µh,g Cg,i
i,j h,g i,j h g
X X
∗
= Bi,j Ch,j (UC)h,i
i,j h
X X
∗T
= Bi,j Cj,h (UC)h,i
i,j h
XX ¡ ¢
= Bi,j C ∗T UC j,i
i j
X¡ ¢
= BC ∗T UC i,i
i
¡ ¢
= Tr BC ∗T UC . (C.3)
where C(ql ) and λ(ql ) [required for the evaluation of B(ql )] are calculated using
equation (3.15).
The matrix product C † (ql )U(ql )C(ql ) is independent of probe position R, the R
dependence being confined to the matrix B(ql ). This means that C † (ql )U(ql )C(ql )
need only be calculated once, and only B(ql ) needs be recalculated as a function of
position. Equivalently, the summations over g and h occurring in equation (C.1)
need only be calculated once, with signals for subsequent probe positions requiring
only summations over i and j.
¤£¡
ª@? ¡¢
R
R?
@ ª
¡
Appendix D
©©
¼ ªH
¡ ?
RH
@ j
Incoherent imaging in a
multislice formulation –
nonlocal expression
Equation (4.4) can be recast into a form amenable to multislice evaluation. The
derivation follows. Equation (4.5) may be written as
Z t
ij i 1 j∗ i j∗ z
B (R, t) = α (R)α (R) e2πiλ z e−2πiλ dz . (D.1)
t 0
X Z
i 1 t 2πiλi z −2πiλj∗ z X i j∗
j∗
I(R, t) = t α (R)α (R) e e dz Cg Ch µh,g
i,j
t 0 g,h
X Z t X X
i j∗
= αi (R)αj∗ (R) e2πiλ z e−2πiλ z dz Cgi Chj∗ δg,g0 δh,h0 µh0 ,g0 .
i,j 0 g,h g0 ,h0
(D.2)
In the last line the Kronecker delta has been introduced with the aim of introducing
integrals over the plane of the crystal through the identity
Z
1 0
δg,g0 = e2πi(g−g )·r⊥ dr⊥ , (D.3)
A A
231
232
where A is the area of the supercell. Substituting equation (D.3) into equation (D.2)
Z tX X X X
i j∗ z
I(R, t) = αi (R)ei2πλ z Cgi αj∗ (R)e−2πiλ Chj∗ ×
0 i g j h
X 1Z 1
Z
2πi(g−g0 )·r⊥ 0 0
e dr⊥ e2πi(h −h)·r⊥ dr0⊥ µh0 ,g0 dz
g0 ,h0
A A A A
Z ( Z X " #
1 X t i
X 0
= 2
αi (R)e2πiλ z Cgi e2πig·r⊥ e−2πig ·r⊥ dr⊥ ×
A g0 ,h0 0 A i g
Z "X X
# )
j∗ 0 0 0
αj∗ (R)e−2πiλ z Chj∗ e−2πih·r⊥ e2πih ·r⊥ dr0⊥ µh0 ,g0 dz
A j h
Z t X"Z Z #
1 0 0 0
= ψ(R, r⊥ , z)e−2πig ·r⊥ dr⊥ ψ ∗ (R, r0⊥ , z)e2πih ·r⊥ dr0⊥ µh0 ,g0 dz ,
A2 0 g0 ,h0 A A
(D.4)
where in the last line we have identified the Bloch wave expression for the wave-
function as given in equation (3.1) and equation (3.2). Identify the integrals over
r⊥ , which are a Fourier transforms, by
Z
Ψ(R, h, z) ≡ ψ(R, r⊥ , z) exp(−2πih · r⊥ )dr⊥ . (D.5)
A
This expression may readily be evaluated via the multislice technique by performing
the integration over z as a sum over slices
M
" #
1 X X
I(R, t) = 2 Ψ(R, h, zm )Ψ∗ (R, g, zm )µh,g ∆zm . (D.7)
A m=1 g,h
By construction, the results of this algorithm should be equal to those resulting from
the Bloch wave implementation; the expressions are equivalent.
¤£¡
ª@? ¡¢
R
R?
@ ª
¡
Appendix E
©©
H
¼ ª?
¡ RH
@ j
Incoherent imaging in a
multislice formulation – the
local approximation
The local approximation is obtained by the substitution µh,g ⇒ µh−g,0 (Allen and
Josefsson, 1995). Implementing this in equation (D.7) and then using equation
(D.5):
M
" #
1 X X
I(R, t) ' Ψ(R, h, zm )Ψ∗ (R, g, zm )µh−g,0 ∆zm
A2 m=1 g,h
M
( "Z
1 X X
= ψ(R, r⊥ , zm )e−2πig·r⊥ dr⊥
A2 m=1 g,h A
Z # )
0
× ψ ∗ (R, r0⊥ , zm )e2πih·r⊥ dr0⊥ µh−g,0 ∆zm
A
M Z Z
X
1
= ψ(R, r⊥ , zm )ψ ∗ (R, r0⊥ , zm ) ×
A2 m=1 A A
à !
X
2πih·r0⊥
µh−g,0 e−2πig·r⊥ e dr⊥ dr0⊥ ∆zm . (E.1)
g,h
233
234
X X · Z ¸
−2πig·r⊥ ih·r0⊥ 2πih·r0⊥ 2πig·(r0⊥ −r⊥ )
µh−g,0 e e = µh,0 e A e dg
g,h h
X 0
= A µh,0 e2πih·r⊥ δ (r0⊥ − r⊥ ) . (E.4)
h
M Z Z
A X
I(R, t) = ψ(R, r⊥ , zm )ψ ∗ (R, r0⊥ , zm )
A2 m=1 A A
" #
X 0
× µh,0 e2πih·r⊥ δ (r0⊥ − r⊥ ) ∆zm dr⊥ dr0⊥
h
M
"Z Z
1 XX
= µh,0 ψ(R, r⊥ , zm )ψ ∗ (R, r0⊥ , zm )
A m=1 h A A
#
0
×e2πih·r⊥ δ (r0⊥ − r⊥ ) ∆zm dr⊥ dr0⊥
M ·Z ¸
1 XX 2 2πih·r⊥
= µh,0 |ψ(R, r⊥ , zm )| e dr⊥ ∆zm
A m=1 h A
Z "X M
#
1X
= µh,0 |ψ(R, r⊥ , zm )| ∆zm e2πih·r⊥ dr⊥ .
2
(E.5)
A h A m=1
This gives the dynamical contribution to the STEM image for a local effective po-
tential. Equation (4.13) reduces to this when the integral over z is approximated by
the sum over slices.
¤£¡
ª@? ¡¢
R
R?
@ ª
¡
Appendix F
©©
¼ ªH
¡ ?
RH
@ j
As per equation (4.11), the fractional intensity for inelastic scattering may be writ-
ten, in a form amenable to multislice evaluation, as
Z t "X #
1
I(R, t) = 2 µh,g Ψ∗ (R, h, z)Ψ(R, g, z) dz , (F.1)
A 0 g,h
1 X 2πih·r⊥ −2πig·r0⊥
µ(r⊥ , r0⊥ ) ≡ µ h,g e e . (F.3)
A2 h,g
235
236
This normalization has been chosen because, as may readily be shown with use of
equation (D.3), it gives that
Z Z
0
µh,g ≡ µ(r⊥ , r0⊥ )e−2πih·r⊥ e2πig·r⊥ dr⊥ dr0⊥ . (F.4)
Appendix G
©©
¼ ªH
¡ ?
RH
@ j
We reconsider the derivation by Hall and Hirsch (1965) of the effect on the scattering
of electrons due to the thermal vibration of the atoms. We consider a single, isolated
atom, the position of which is described by an average position R0 and the time
dependent displacement u(t):
where k0 is the wavevector and g denotes a reciprocal space vector in the plane
perpendicular to the optical axis, though for a single atom the mesh on which
g lies is not strictly discrete. We seek to express the intensity scattered to the
vector s = k − k0 . In what follows we will make use of the flat Ewald sphere
approximation to assert that s will lie in the zero order Laue zone. Using the
kinematic approximation, and following the reasoning of Hall and Hirsch (1965),
237
238
Taking the thermal average over positions of the atom, assuming an Einstein model,
we write
X
hI(s)i = Cg Ch∗ fs−g fs−h
∗
hexp[−2πi(h − g) · R]i . (G.4)
g,h
Using the harmonic oscillator or Einstein model for the single atom, the thermal
average obeys the rule hexp(A)i = exp( 12 hA2 i) (Willis and Pryor, 1975). Thus
n o
2 2
hexp[−2πi(h − g) · u(t)]i = exp − 2π h[(h − g) · u(t)] i
≡ exp[−2π 2 (h − g)2 hu2p i] , (G.6)
where hu2p i denotes the projected mean square vibrational amplitude. The last term
in the previous equation may be expanded as
where we have introduced the Debye-Waller factor M (g) = 2π 2 g2 hu2p i. Consider the
Taylor expansion
∞
X ∞
X
cn (p · q)n cn
exp(cp · q) = = (px qx + py qy )n
n=0
n! n=0
n!
∞ X
X n n
c
= pk pn−k q k q n−k , (G.8)
n=0 k=0
(n − k)!k! x y x y
where in the final step we have used the binomial theorem. Now consider the last
exponential term in equation (G.7). Identifying c with 4π 2 hu2p (t)i, p with s − g, and
Chapter G. Derivation of the scattering functions 239
q with s − h we may combine equations (G.6), (G.7) and (G.8), motivated by the
approach of Anstis and co-workers (Anstis et al., 1996; Anstis, 1999), to re-write
equation (G.5) as
X n on o
hI(s)i = Cg Ch∗ fs−g exp[−M (s − g)] fs−h
∗
exp[−M (s − h)]
g,h
∞
XXn 2
[4π hu2p i]n
× (s − g)kx (s − g)n−k
y (s − h)kx (s − h)n−k
y exp[−2πi(h − g) · R0 ]
n=0 k=0
(n − k)!k!
X∞ X n
[4π 2 hu2p i]n
=
n=0 k=0
(n − k)!k!
( )
X
× Cg fs−g exp[−M (s − g)](s − g)kx (s − g)yn−k exp[−2πi(s − g) · R0 ]
g
( )∗
X
× Ch fs−h exp[−M (s − h)](s − h)kx (s − h)yn−k exp[−2πi(s − h) · R0 ]
h
∞ X
X n
≡ |ψn,k (s)|2 , (G.9)
n=0 k=0
where
s
[4π 2 hu2p i]n X
ψn,k (s) = Cg fs−g exp[−M (s − g)](s − g)kx (s − g)yn−k
(n − k)!k! g
× exp[−2πi(s − g) · R0 ]
X
≡ An,k Cg fs−g exp[−M (s − g)](s − g)kx (s − g)yn−k
g
× exp[−2πi(s − g) · R0 ] ,
(G.10)
is simply the expression for the thermally smeared elastic potential of an atom. Thus
the first term in the series of equation (G.14) is simply the scattering distribution
due to elastic scattering of a single atom in the first Born approximation.
Using an analytic parameterization for the electron scattering factors of the form
X £ ¤
f (g) = Aν exp −Bν g2 /4 , (G.17)
ν
such as that of Peng et al. (1996), and converting the discrete sum in equation
(G.16) to an integral as appropriate for the treatment of a single atom, it may be
shown that
X · ¸
2 Aν 4π 2 r⊥
2
V1,1 (r⊥ ) = i16π A1,1 rx exp − ,
ν
[Bν + 4M ]2 Bν + 4M
X · ¸
2 Aν 4π 2 r⊥
2
V1,0 (r⊥ ) = i16π A1,0 ry exp − . (G.18)
ν
[Bν + 4M ]2 Bν + 4M
These equations are equivalent to the effective transmission functions for thermal
scattering introduced by Dinges et al. (1995). We have introduced this parameteri-
Chapter G. Derivation of the scattering functions 241
zation to show this equivalence, but simulations in this thesis were based upon the
parameterization of Waasmaier and Kirfel (1995).
242
¤£¡
ª@? ¡¢
R
R?
@ ª
¡
Appendix H
©©
H
¼ ª?
¡ RH
@ j
List of publications
243
244
• S. D. Findlay
Quantitative structure retrieval using scanning transmission electron microscopy
Acta Cryst. A61 (2005) 397-404.