Upload
others
View
5
Download
0
Embed Size (px)
Citation preview
1
Evaluation of XRF Spectra
from basics to advanced systems
Joint ICTP-IAEA School on Novel Experimental Methodologies for Synchrotron Radiation Applications in Nano-science and Environmental Monitoring
Piet Van Espen [email protected]
20 Nov 2014
mailto:[email protected]
2
Content
1. Introduction: some concepts
2. Simple peak integration
3. Method of least squares
4. Fitting of x-ray spectra
5. Improvements to the model
6. Final remarks
3
Some basic concepts
Nature and science: a personal view
Nature = Signal + Noise
Science = Model + Statistics
4 4
Why spectrum evaluation?
element concentrations ⇔ net intensity of fluorescence lines
But: frequent peak overlap presence of a continuum
Especially in energy-dispersive spectra
interference free continuum corrected
5
Spectrum contains • Information: energy and intensity of x-rays
• Amplitude noise: due to Poisson statistics ► fluctuations in the spectrum
• Energy noise: finite resolution of the detector ► nearly Gaussian peaks with a width of ~160 eV
Information content of a spectrum
the signal
the noise
6
Amplitude noise Counting events involves Poisson statistics
Poisson probability density function:
The probability to observe N counts if the true number is µ
Poisson : P(N | µ=3) Normal : P(x | µ=3 σ2 = 3)
Property:
Poisson distribution µ ≅
Normal distribution µ and σ2 = µ
approximation is very good for µ (or N) ≥ 9
7
Resolution of ED-XRF spectrometers
Full Width at Half Maximum (FWHM) of a peak
Mn Kα @ 5.895 keV FWHMDet = 120 eV FWHMElec = 100 eV => FWHMPeak = 156 eV
Intrinsic contribution2 Spectrum analysis
2.35⇥
�� F � E (9)
FWHM2Peak = FWHM2Elec + Det
2Pealk (10)
2
ϵ energy to create e-h pair 3.85 eV F Fano factor ~0.114 E x-ray energy in eV
2 Spectrum analysis
2.35⇥
�� F � E (9)
FWHM2Peak = FWHM2Elec + FWHM
2Det (10)
2
Electronic noise ~100 eV
Energy Noise
8
Cr – Mn – Fe overlap at ~20 eV Cr – Mn – Fe overlap at ~160 eV
Resolution of ED-XRF spectrometers Energy Noise
9
Without amplitude noise (counting statistics) there would be NO PROBLEM
But it is part of the nature We can only measure longer or with a more efficient system
Without energy noise there would be LITTLE PROBLEM
The natural line width of x-rays is only a few eV!!!
The observed peak width is the result of the detection process with a fundamental limitation imposed by the Fano factor
10 10
Information content of a spectrum
If no energy noise or no amplitude noise ► could determine the “information” unambiguously
Need methods to extract information in a optimum way These methods rely on “addition” information (knowledge) to extract the useful information
Not the method itself is important (if implemented correctly) but the correctness of the additional information.
11
2. Simple peak integration
12
Simple peak integration
Estimate
Uncertainty
We have to make assumptions integration limits linear background no interference
As good as is can get if the assumptions (model) are correct!
13
3. Method of Least Squares
14
Need to “estimate” the net peak area with highest possible • correctness (no systematic error) • precision (smallest random error)
Least-squares estimation (fitting): • unbiased • minimum variance
Limiting factors: • counting statistical fluctuations (precision) • accuracy of the fitting model
Method of least squares
15
Method of least squares, straight line
2 Spectrum analysis
2.35p
�⇥ F ⇥ E (9)
FWHM2Peak = FWHM2Elec + FWHM
2Det (10)
SS =X
i
[yi � y(i)]2 =X
i
[yi � b0 � b1xi]2 = min (11)
⇤SS
⇤b0= 0 !
X
i
yi = b0n + b1X
i
xi (12)
⇤SS
⇤b1= 0 !
X
i
xiyi = b0X
xi + b1X
i
x2i (13)
2
2 Spectrum analysis
2.35p
�⇥ F ⇥ E (9)
FWHM2Peak = FWHM2Elec + FWHM
2Det (10)
SS =X
i
[yi � y(i)]2 =X
i
[yi � b0 � b1xi]2 = min (11)
⇤SS
⇤b0= 0 !
X
i
yi = b0n + b1X
i
xi (12)
⇤SS
⇤b1= 0 !
X
i
xiyi = b0X
xi + b1X
i
x2i (13)
2
Set of 2 equations in 2 unknowns b0 and b1 Normal equations
Direct analytical solution
Data: {xi,yi}, i=1, 2, …, N
Model: y(i) = b0 + b1xi
Fitting the model: estimating b0 and b1
Criterion:
noise
16
4. Fitting X-ray Spectra
17
Least-squares estimate of x-ray spectrum parameters
Peak described by a Gaussian
Minimum: No direct analytical solution Search χ2 for minimum
�
2
=
1
⌫
n2X
i=n1
1
wi[y(i)� yi]2 (1)
y(i) = y
Cont
(i) +
X
j
Aj
"X
k
RijP (i, Eij)
#(2)
P = G(i, Ejk) =Gain
�jk
p2⇡
exp
"�1
2
✓E(i)� Ejk
�jk
◆2
#(3)
E(i) = Zero + Gain⇥ i (4)
�jk =
"✓Noise
2
p2 ln 2
◆2
+ ✏ Fano Ejk
#1/2
(5)
T (i, Ejk) =Gain
2�� exp
h� 1
2�2
iexp
E(i)� Ejk
��
�erfc
E(i)� Ejk
�
p2
+
1p2�
�
S(i, Ejk) =Gain
2Ejkerfc
E(i)� Ejk
�
p2
�
P (i, Ejk) = G(i, Ejk) + fSS(i, Ejk) + fT T (i, Ejk) (6)
fS(Ejk) = µDet(Ejk) (a0 + a1Ejk)
fTK↵(Ejk) = b0 + b1µDet(Ejk)
fTK�(Ejk) = c0 + c1µDet(Ejk)
�(Ejk) = d0 + d1Ejk
Gain
�jk
p2⇡
K
E(i)� Ejk
�jk
p2
,
↵L
2�jk
p2
!
K(x, y) = Re
⇥exp(�z2)erfc(�iz)
⇤z = x + iy (7)
1
y(i) = b + A
1
�
p2⇡
exp
(xi � xp)2
2�
2
�
1
position
width
area
continuum
Criterion, agreement between model and data
linear parameter
non-linear
18
We can still apply the concept of least squares minimising the square of the differences between the model and the data
The sum of squares is a function of the values of the parameters and for a given set of values should be minimum
In this case SS describes a 4 dimensional hyper-surface in a 5-dimension space
h
We can only “see” in 3-dimensions but mathematically we can search in a higher dimensional space to locate the minimum
Starting from some initial values we can modify the parameter values until the minimum is reached.
�
2 = �2(b, h, x0, w) =1⌫
Pi
1yi
[yi � y(xi, b, h, x0, w)]2
w
19
General form of such a search algorithm
1. Select starting values for all parameters bj and calculate the ch-square
2. Obtain (calculate, guess...) a change (increment or decrement) Δbj such that one moves towards the minimum:
3. Replace the old parameter values with the new ones b ← b + Δb
4. repeat step 2 until the “true” minimum is found
Iterative process
AXIL = Analysis of X-ray spectra by Iterative Least-squares
� = �(b)
�(b+�b) < �(b)
20
Analytically important parameters: net peak areas
In general y(i) is non-linear → Marquardt – Leverberg algorithm Gradient search ↔ linearisation Reliable error estimated But unstable
Statistical optimal estimate: using correct weight (Poisson statistics wi = yi)
�
2
=
1
⌫
n2X
i=n1
1
wi[y(i)� yi]2 (1)
y(i) = y
Cont
(i) +
X
j
Aj
"X
k
RijP (i, Eij)
#(2)
P = G(i, Ejk) =Gain
�jk
p2⇡
exp
"�1
2
✓E(i)� Ejk
�jk
◆2
#(3)
E(i) = Zero + Gain⇥ i (4)
�jk =
"✓Noise
2
p2 ln 2
◆2
+ ✏ Fano Ejk
#1/2
(5)
T (i, Ejk) =Gain
2�� exp
h� 1
2�2
iexp
E(i)� Ejk
��
�erfc
E(i)� Ejk
�
p2
+
1p2�
�
S(i, Ejk) =Gain
2Ejkerfc
E(i)� Ejk
�
p2
�
P (i, Ejk) = G(i, Ejk) + fSS(i, Ejk) + fT T (i, Ejk) (6)
fS(Ejk) = µDet(Ejk) (a0 + a1Ejk)
fTK↵(Ejk) = b0 + b1µDet(Ejk)
fTK�(Ejk) = c0 + c1µDet(Ejk)
�(Ejk) = d0 + d1Ejk
Gain
�jk
p2⇡
K
E(i)� Ejk
�jk
p2
,
↵L
2�jk
p2
!
K(x, y) = Re
⇥exp(�z2)erfc(�iz)
⇤z = x + iy (7)
1
21
10 peaks ⇒ > 30 parameters !!!! WANT WORK in practice!!!!
But we can do better
⇒ Add additional information to the model
We known the energies of the x-ray lines (in most cases) Where they are depends on the energy calibration (the same applies to the width of the peaks: resolution calibration)
y(i) = b + A
1
�
p2⇡
exp
(xi � xp)2
2�
2
�
G(i, E) =
Gain
�(E)
p2⇡
exp
(Ei � E)2
2�
2(E)
�
1
Gaussian peak shape
Energy relation:
Resolution relation:
y(i) = b + A
1
�
p2⇡
exp
(xi � xp)2
2�
2
�
G(i, E) =
Gain
�(E)
p2⇡
exp
(Ei � E(i))2
2�
2(E)
�
E(i) = Zero + Gain⇥ i
�(E) =
"✓Noise
2
p2 ln 2
◆2+ ✏FanoE
#1/2
1
y(i) = b + A
1
�
p2⇡
exp
(xi � xp)2
2�
2
�
G(i, E) =
Gain
�(E)
p2⇡
exp
(Ei � E(i))2
2�
2(E)
�
E(i) = Zero + Gain⇥ i
�(E) =
"✓Noise
2
p2 ln 2
◆2+ ✏FanoE
#1/2
1
Only 4 non-linear parameters For 10 peaks only 14 parameters
Need parsimony!!!
22
Already better, but we know more
We know (to some extend) the ratio between lines of an element
y(i) = b + A
1
�
p2⇡
exp
(xi � xp)2
2�
2
�
G(i, E) =
Gain
�(E)
p2⇡
exp
(Ei � E(i))2
2�
2(E)
�
E(i) = Zero + Gain⇥ i
�(E) =
"✓Noise
2
p2 ln 2
◆2+ ✏FanoE
#1/2
IK↵2
IK↵1,
IK�
IK↵. . .
1
We can group lines together (“peakgroup”) with one “area” and fixed intensity ratios
Continuum function
Area Line ratio
Peak shape
for j elements (or peak groups)
y(i) = b + A
1
�
p2⇡
exp
(xi � xp)2
2�
2
�
G(i, E) =
Gain
�(E)
p2⇡
exp
(Ei � E(i))2
2�
2
(E)
�
E(i) = Zero + Gain⇥ i
�(E) =
"✓Noise
2
p2 ln 2
◆2
+ ✏FanoE
#1/2
IK↵2
IK↵1,
IK�
IK↵. . .
y(i) = y
Cont
(i) +
X
j
Aj
(X
k
RjkP (i, Ejk)
)
1
10 elements ⇒ 10 Area’s + 4 calibration parameters
23
Further refinements: escape peaks Known position (energy) intensity (escape probability)
24
Sum peaks Known position relative intensity
25
And more Different background models polynomial exponential polynomial Bremsstrahlung background filter background
Parameter constraining
and more...
26
27
Highly flexible method • Fit individual lines, multiplets, elements… • Different parametric and non-parametric continuum models • Include escape and sum peaks
Quality criteria • Chi-square of fit • uncertainty estimate of parameters
Statistically correct • unbiased, minimum variance estimate of the parameters
“Resolving power” is only limited by the noise (counting statistic) BUT
THE MODEL MUST BE ACCURATE
28
5. Improvements to the model
29
Incorrectness of the model
Not all peaks follow the energy calibration relation - incoherent (Compton) scatter peaks - spurious peaks (diffraction, γ-rays) - even the relation might not be linear
Not all peaks follow the resolution calibration relation - incoherent scatter peaks (are wider) - spurious peaks
Peaks are certainly not perfect Gaussians - shelf (step) due to detector effects (incomplete charge collection) - tailing due to radiative effects and detector effects - deviation due to natural line width (Lorentzian)
30
Incorrect fitting model biased results
Solution Adapt the model (fitting region, which lines to include...)
for each particular case
Very inconvenient when analyzing many spectra
⇒
especially for trace elements
31
Step
y(i) = y
Cont
(i) +
X
j
Aj
"X
k
RijP (i, Eij)
#(1)
P = G(i, Ejk) =Gain
�jk
p2⇡
exp
"�1
2
✓E(i)� Ejk
�jk
◆2
#(2)
E(i) = Zero + Gain⇥ i (3)
�jk =
"✓Noise
2
p2 ln 2
◆2
+ ✏ Fano Ejk
#1/2
(4)
T (i, Ejk) =Gain
2�� exp
h� 1
2�2
iexp
E(i)� Ejk
��
�erfc
E(i)� Ejk
�
p2
+
1p2�
�
S(i, Ejk) =Gain
2Ejkerfc
E(i)� Ejk
�
p2
�
1
Peak
y(i) = y
Cont
(i) +
X
j
Aj
"X
k
RijP (i, Eij)
#(1)
P = G(i, Ejk) =Gain
�jk
p2⇡
exp
"�1
2
✓E(i)� Ejk
�jk
◆2
#(2)
E(i) = Zero + Gain⇥ i (3)
�jk =
"✓Noise
2
p2 ln 2
◆2
+ ✏ Fano Ejk
#1/2
(4)
T (i, Ejk) =Gain
2�� exp
h� 1
2�2
iexp
E(i)� Ejk
��
�erfc
E(i)� Ejk
�
p2
+
1p2�
�
S(i, Ejk) =Gain
2Ejkerfc
E(i)� Ejk
�
p2
�
P (i, Ejk) = G(i, Ejk) + fSS(i, Ejk) + fT T (i, Ejk) (5)
1
Adding 1 non-linear and 2 linear parameters
for each peak!!!
Tail
y(i) = y
Cont
(i) +
X
j
Aj
"X
k
RijP (i, Eij)
#(1)
P = G(i, Ejk) =Gain
�jk
p2⇡
exp
"�1
2
✓E(i)� Ejk
�jk
◆2
#(2)
E(i) = Zero + Gain⇥ i (3)
�jk =
"✓Noise
2
p2 ln 2
◆2
+ ✏ Fano Ejk
#1/2
(4)
T (i, Ejk) =Gain
2�� exp
h� 1
2�2
iexp
E(i)� Ejk
��
�erfc
E(i)� Ejk
�
p2
+
1p2�
�
S(i, Ejk) =Gain
2Ejkerfc
E(i)� Ejk
�
p2
�
P (i, Ejk) = G(i, Ejk) + fSS(i, Ejk) + fT T (i, Ejk) (5)
1
Improve the peak profile
Where is my parsimony gone!!!
Improvements
32
Step parameterisation
Step is a fundamental aspect of the detector (charge loss by photo-electrons near the surface of the detector)
Step fraction fS is related to the MAC of the detector crystal
Tail fraction parameterisation
Tail fraction fT is related to the MAC of the detector and the type of radiation (Kα and Kβ)
The tail has a component due to the detector and a radiative component
33
Tail width parameterisation
similar magnitude over the entire energy range
y(i) = y
Cont
(i) +
X
j
Aj
"X
k
RijP (i, Eij)
#(1)
P = G(i, Ejk) =Gain
�jk
p2⇡
exp
"�1
2
✓E(i)� Ejk
�jk
◆2
#(2)
E(i) = Zero + Gain⇥ i (3)
�jk =
"✓Noise
2
p2 ln 2
◆2
+ ✏ Fano Ejk
#1/2
(4)
T (i, Ejk) =Gain
2�� exp
h� 1
2�2
iexp
E(i)� Ejk
��
�erfc
E(i)� Ejk
�
p2
+
1p2�
�
S(i, Ejk) =Gain
2Ejkerfc
E(i)� Ejk
�
p2
�
P (i, Ejk) = G(i, Ejk) + fSS(i, Ejk) + fT T (i, Ejk) (5)
fS(Ejk) = µDet(Ejk) (a0 + a1Ejk)
fTK↵(Ejk) = b0 + b1µDet(Ejk)
fTK�(Ejk) = c0 + c1µDet(Ejk)
�(Ejk) = d0 + d1Ejk
1
(compare to Zero, Gain, Noise and Fano)
⇒ Fitting parameters a0, a1, b0, b1, c0, c1, d0, d1
34
improvementFit of a NIST SRM 1106 Brass spectrum (SpecTrace 5000, Rh tube)
35
To account for peak shift and peak broadening
Need to make the peak profile still a bit more complicated
y(i)
=
b +A
1
�
p2⇡
exp
(xi � xp)2
2�
2
�
G(i
, E) =
Gai
n
�(E
)
p2⇡
exp
(Ei � E(i))2
2�
2
(E)
�
E(i
) =
Zer
o +G
ain⇥ i
�(E
) =
"✓N
oise
2
p2 l
n 2
◆2
+✏F
anoE
#1/2
IK↵2
IK↵1,
IK�
IK↵. .
.
y(i)
=
y
Cont
(i) +
X
j
Aj
(X
k
RjkP (i, Ejk)
)
G(i
, E) =
Gai
n
��(E
)
p2⇡
exp
(Ei � E(i) + �E)2
2��
2
(E)
�
�Eini ���E �E �Eini + ��E
�ini ��� � �ini + ��
1
γ peak broadening parameter (normally 1.00)
δE peak shift parameter (normally 0.00)
These parameters are constrained to vary within a certain range
y(i) = b + A
1
�
p2⇡
exp
(xi � xp)2
2�
2
�
G(i, E) =
Gain
�(E)
p2⇡
exp
(Ei � E(i))2
2�
2
(E)
�
E(i) = Zero + Gain⇥ i
�(E) =
"✓Noise
2
p2 ln 2
◆2
+ ✏FanoE
#1/2
IK↵2
IK↵1,
IK�
IK↵. . .
y(i) = y
Cont
(i) +
X
j
Aj
(X
k
RjkP (i, Ejk)
)
G(i, E) =
Gain
��(E)
p2⇡
exp
(Ei � E(i) + �E)2
2��
2
(E)
�
�Eini ���E �E �Eini + ��E
�ini ��� � �ini + ��
1
36
The last step
Replace Gaussian with the convolution of a Gaussian with a Lorentzian = Voigt profile
y(i) = y
Cont
(i) +
X
j
Aj
"X
k
RijP (i, Eij)
#(1)
P = G(i, Ejk) =Gain
�jk
p2⇡
exp
"�1
2
✓E(i)� Ejk
�jk
◆2
#(2)
E(i) = Zero + Gain⇥ i (3)
�jk =
"✓Noise
2
p2 ln 2
◆2
+ ✏ Fano Ejk
#1/2
(4)
T (i, Ejk) =Gain
2�� exp
h� 1
2�2
iexp
E(i)� Ejk
��
�erfc
E(i)� Ejk
�
p2
+
1p2�
�
S(i, Ejk) =Gain
2Ejkerfc
E(i)� Ejk
�
p2
�
P (i, Ejk) = G(i, Ejk) + fSS(i, Ejk) + fT T (i, Ejk) (5)
fS(Ejk) = µDet(Ejk) (a0 + a1Ejk)
fTK↵(Ejk) = b0 + b1µDet(Ejk)
fTK�(Ejk) = c0 + c1µDet(Ejk)
�(Ejk) = d0 + d1Ejk
Gain
�jk
p2⇡
K
E(i)� Ejk
�jk
p2
,
↵L
2�jk
p2
!
1
y(i) = y
Cont
(i) +
X
j
Aj
"X
k
RijP (i, Eij)
#(1)
P = G(i, Ejk) =Gain
�jk
p2⇡
exp
"�1
2
✓E(i)� Ejk
�jk
◆2
#(2)
E(i) = Zero + Gain⇥ i (3)
�jk =
"✓Noise
2
p2 ln 2
◆2
+ ✏ Fano Ejk
#1/2
(4)
T (i, Ejk) =Gain
2�� exp
h� 1
2�2
iexp
E(i)� Ejk
��
�erfc
E(i)� Ejk
�
p2
+
1p2�
�
S(i, Ejk) =Gain
2Ejkerfc
E(i)� Ejk
�
p2
�
P (i, Ejk) = G(i, Ejk) + fSS(i, Ejk) + fT T (i, Ejk) (5)
fS(Ejk) = µDet(Ejk) (a0 + a1Ejk)
fTK↵(Ejk) = b0 + b1µDet(Ejk)
fTK�(Ejk) = c0 + c1µDet(Ejk)
�(Ejk) = d0 + d1Ejk
Gain
�jk
p2⇡
K
E(i)� Ejk
�jk
p2
,
↵L
2�jk
p2
!
K(x, y) = Re
⇥exp(�z2)erfc(�iz)
⇤z = x + iy (6)
1
with K(...) the complex error function
For high Z elements the natural line width becomes substantial relative to the detector resolution
37
Natural line width at high Z elements becomes important e.g. W K ~ 50 eV
Gaussians Voigts
38
Original fit of a geological standard (JG1)
39
Improved fit of the geological standard (JG1)
Mo secondary target No background
40
NIST SRM 1155
41
NIST SRM 1247
42
More Details
Handbook of X-ray Spectrometry R. Van Grieken, A. Markowicz Marcel Decker, N.Y. 2002 ISBN: 0-8247-0600-5
Chapter 4: Spectrum Evaluation
43
Some final remarks: The future
Non-linear least-squares works
if you have a good parsimonious modelif you have TIME
X-ray fluorescence imaging: 256 x 256 image = 65536 x-ray spectra
@ 1 s / spectrum= 65536 seconds = 1092 minutes = 18 hours !!!!
Need to explore new methods Linear models? Multivariate models?
44
6. Final remarks
45
Thanks for your attention