Multivariate Optical Computing: Application-Specific Optical Sensors And how to create them

Multivariate Optical Computing:
Application-Specific Optical Sensors
And how to create them
Michael L. Myrick
University of South Carolina
Columbia, SC 29208
PG
MM
MCEC
“All we want is a cheap tricorder. Is that too much to ask?”
CPL
Why is all this
complexity necessary?
Absorbance
Absorbance
20
Species 1
Species 2
15
10
5
400
450
500
550
Wavelength
600
650
20
15
10
5
0
400
450
500
550
Wavelength
600
650
A simple 2-component mixture
is trivial to analyze
Add an arbitrary amount of another species, and complexity increases
30
Absorbance
25
20
15
10
5
0
400
450
500
550
Wavelength
600
650
The mathematics of prediction in direct spectroscopy are very
simple. . . but often very expensive to implement.
yest = b • x = b1x1 + b2x2 + b3x3 + b4x4 +. . .
where b = “scaled” regression vector and x = spectrum, or spectral vector
multiplication
Addition
detector
b1•x(λ1)
x(λ1)
yest
…
T (λ1) =b1
b1•x(λ1)
b2•x(λ2)
b3•x(λ3)
b4•x(λ4)
b5•x(λ5)
Our concept was to use this form of computation to produce small,
inexpensive sensors based on multivariate spectroscopy.
1998
Multivariate Optical Computing:
The use of optical discriminators to
directly sense the magnitude of
pectral patterns in light.
These optical discriminators encode,
n one of several ways, spectral
patterns so that multivariate
measurements are simplified.
We are focussing on interference
ilters to serve as optical
discriminators.
March 15, 2001
Detector for
Transmitted Light
MOE
b(λ)
Light from Sample
D
DTT
DR
DR
Detector for
Reflected Light
R(λ)=0.5-b(λ)
T(λ)=0.5+b(λ)
Multivariate optical elements (MOEs) are interference filters
air
BK7
Nb2O5
SiO2
Bandpass Filter Instrument:
intensity vs. BB concentration in the presence of CV interference.
Intensity (arb units)
SEP ˜ 4 µM
2
4
6
8
10
12
14
16
18
Calibrated Concentration of BB (µM)
Anal. Chem. 73 (2001), 1069.
MOC result:
calibrated (T-R) channel difference vs. BB concentration with CV interference
20
18
SEP = 0.68 µM
Predicted conc. (µM)
16
14
12
10
8
6
4
2
2
7
12
17
Actual conc. (µM)
Anal. Chem. 73 (2001), 1069.
Breakdown of key technologies in MOE application:
1. System Design
2. Radiometry
3. Ellipsometry
4. Thin-film synthesis
5. Thin-film deposition
6. Process control
System Design Begins with a Knowledge of Sample Spectroscopy
Dimethyl methyl
phosphonate/ethyl
acetate mixture
transmission spectra
Conventional
multivariate statistics
can be used to
determine whether a
pectrum-based
measurement is
possible.
System Design
Determines Precision
“poor”
J
σ 2MC
2
=
ξ
∆
∑
j
j
2
σ MOE j=1
∆ j =Difference in MOE T and R at wavelen gth j
ξ j =fractional intensity at λ j in calibratio n set
From “Precision in Multivariate Optical Computing”,
F.G. Haibach and M.L. Myrick, Appl. Optics 43 (2004), 2130-2140.
“so-so”
“ideal”
A measurement system is selected that provides sensitivity in the
appropriate band.
light source
lens
Band selection can be
made by a combination of
prefiltering (if any
sample cell
A. light sources
B. reproducible filters
C. detector response
mirrors
detector
MOE
Spectral Radiance of Source
x 10
-3
Spectral Radiance (W/nm)
1
0.8
0.6
0.4
0.2
0
400
450
500
550
Wavelength (nm)
600
650
Filter Bandpass
Transmittance
0.8
0.6
0.4
0.2
0
400
450
500
550
Wavelength (nm)
600
650
Detector Sensitivity
Spectral Sensitivity (V/mW)
4
3.5
3
2.5
2
1.5
1
400
450
500
550
Wavelength (nm)
600
650
Radiometry
The simple picture of all
wavelengths being equally sensed is
incorrect. Optical systems have
efficiencies that vary with
wavelength. For a given
measurement, a “stable” system
design must be selected and
calibrated. For absorbance or
reflectance measurements, a light
source, optical train and detectors
must all be calibrated.
Total
Photon
flux
Detector
efficiency
Average
Optical Bandpass spectral
efficiency efficiency radiant flux
Ellipsometry
Ellipsometry is used to provide “best
estimates” of the optical constants of
the real materials that will be used to
make a MOE. Every deposition
chamber is slightly different and
changes gradually over time; monthly
recalibration is not a bad idea at all.
Models in the UV-Vis region are easy.
Models in the NIR are slightly more
difficult.
Models in the MIR are an art form.
Deposition and nucleation conditions
can change the exact values for a given
material and can defeat the best models.
Lattice vibrations
Electronic transitions
Refr. Index
Thin Film Design
3.0
1.0
Thickness (100s of nm – 10s of µms)
Paul Gemperline and Olusola Soyemi
developed an iterative solving procedure to
generate structures for MOEs using the
SEP as a figure of merit that has become a
basis for our current algorithms.
A design represents a road map for
deposition of a MOE, not a full description
of what the finished product will look like.
Thin Film Deposition
In our laboratory we use reactive magnetron
sputtering to make thin films. For 350-1400
nm MOEs we use Nb2O5 and SiO2 films.
For longer wavelengths, higher-index
materials are available. Oxides deposit at
about 100-200 nm/hour. Silicon deposits at
about 10X this rate.
Most commercial film production is done
with e-beam evaporation, which is usually
more than 10X faster than sputtering.
Process Control
Because of nucleation and growth
properties of films, their optical
properties vary with thickness. Small
changes in temperature, oxygen content,
deposition rate, etc. can result in
deviations of the growing film from the
trajectory expected of the design.
Nb
Si
Nb
Si
Reflectance is a Complex Quantity
The Bootstrap Method
In the past several months, we have developed
a new method of controlling the deposition of
thin films in which we abandon complete
modeling of the film stack by deriving the
complex reflectance at each wavelength using
real experimental measurements.
The consequence of “bootstrapping” is that the
imperfections in already-deposited layers are
eliminated so that redesign can be performed
using “true” reflectance values rather than
idealized values.
Bootstrapping has not been automated (yet).
Regression Vector of IMOE @ 45º
To date, all our development work has
been done with visible-absorbing
samples because most of our experience
has been with thin films for visible
optics.
0.8
Regression Vector (unscaled) @ 45º
0.6
Calculated
Measured
0.4
0.2
0
Ordinarily, the extension to NIR/MIR
would be very difficult because the
materials are less dependable.
-0.2
-0.4
-0.6
-0.8
-1
400
450
500
550
600
650
However, Bootstrapping does not
require “dependable” films!
Wavelength, nm
We believe that NIR and MIR sensors using multivariate optical
computing are within reach once we’ve finalized our automation code.
So What Can MOEs be Used For?
UV-Vis absorbance
Point Sensing
Fluorescence/Phosphoresce
nce
Imaging
Classification
Raman Scattering
Spectral Data Compression
Reflectance
Adulteration Testing
NIR absorbance
Miniature optical sensors
MIR absorbance
In-vivo or In-process sensors
Thermal Emission
T -Is o
40
20
T-R
0
D una
-2 0
-4 0
-6 0
-8 0
- 10 0 x1 0
3
P h ae
0
100
200
300
T+R
400
5 0 0 x1 0
3
Multivariate Optical Computing is not magic:
it is another way of performing an end measurement.
MOE Advantages
MOE Disadvantages
Low cost
Small
Low power
Rapid
Low-light
Mass producible
“just give me the result”
Not a spectroscopic research tool
Fixed spectral interpretation
We haven’t made a tricorder yet, but we’re making progress. . .
Acknowledgements
Paul Gemperline, ECU
Norman Schmidt, GSU
Olusola Soyemi
Karl Booksh
Fred Haibach
Maria Schiza
Raju Karunamuni
Matt Nelson
Hong Li
Ashley Greer
David Perkins
Ryan Priore
Ruth Wang
Joshua Farr
Burt V. Bronk
Funding: ONR, NSF, AFRL*, Detect-X,
Foster Miller Inc.