Journal of
MOLECULAR STRUCTURE ELSEVIER
Journal of Molecular Structure 347 (1995) 169-186
Remote Characterization of Materials by Vibrational Spectrometry through Optical Fibers Peter R. Griffiths, Ian R. Lewis, Nathan C. Chaffin, Nelson W. Daniel, Jr. and John D. Jegla Department of Chemistry, University of Idaho, Moscow, Idaho 84844-2343, USA
1. I N T R O D U C T I O N In the past five years, several applications have been reported in which spectroscopic measurements have been made with the sample located at distances of up to 200 meters from the spectrometer, with the radiation being transmitted to and from the sample via optical fibers 1'2. By far the most popular of these involve quantitative assays by near infrared (NIR) spectrometry; in such measurements, the radiation is passed to the sample through one silica fiber or a fiber bundle and then returned to the detector through another. Although not as popular as transmission measurements, NIR external reflection spectrometry has also been applied to the quantitative analysis of remote powdered samples. We are presently involved in two research projects in which, rather than a remote quantitative determination, the goal is the identification or characterization of remote materials through their vibrational spectra. In this paper, the nature of each of these projects will be described and the factors influencing the choice of spectroscopic technique will be discussed.
1.1. Three-dimensional Integrated Characterization and Archiving System In the first of these projects, we are one of three groups that are developing a threedimensional integrated characterization and archiving system (3D-ICAS) to classify remote objects at sites such as hazardous or radioactive waste storage facilities. By United States federal law, materials containing hazardous materials, such as asbestos or polychlorinated biphenyls, must be disposed of more carefully than less hazardous materials. In situ identification of such components is necessary before hazardous materials can be handled in the appropriate manner. The 3D-ICAS system consists of three components, a coherent laser radar system by which a 3D image of the site can be obtained, a vibrational spectrometry sensor to identify the major component at a given point, and a combined gas chromatograph and mass spectrometer (GC/MS) to find whether semivolatile materials are present in this point. The coherent laser radar system is analogous to conventional frequency-modulated microwave radar where the FM source produces a continuous beam of radio-frequency radiation which is directed at the target. Since the detector sees energy directly from the source as well as radiation reflected from the target, interference beats are detected as the frequency, ./~, is swept over the interval Af The rate of these beats is a function of the range and the value of Af In the FM laser radar system, coherent 1550-nm radiation from a 0022-2860/95/$09.50 © 1995 Elsevier Science B.V. 169-186
SSDI 0022-2860(95)08544-0
All rights reserved
170 frequency-modulated GaA1As laser diode is transmitted down an optical fiber and collimated; the radiation reflected from a remote object is collected and passed back through an optical fiber to the detector. A second channel of coiled, temperature-regulated optical fiber located within the radar housing provides the reference information to a counter needed to make the necessary range computation. Three-dimensional images with better than l-ram resolution can be obtained at distances up to 100 meters from the control module. Once the three-dimensional image has been obtained, a robotically controlled arm brings the GC/MS and vibrational spectrometry probes to within a few millimeters of each surface of interest. Using the vibrational spectrometry probe, the NIR reflectance spectrum and the Raman spectrum of the material are measured and this material is either characterized by neural network software or identified by spectral searching; these probes will be discussed in more detail later in this paper. To identify the low molecular-weight material near the surface, a high-power pulse is transmitted down an optical fiber to ablate the volatile and semivolatile materials from near the surface. These vapors are swept through a hollow fiber for separation by a small, high-resolution gas chromatograph and characterization of each component by mass spectrometry. The projected device is shown schematically in Figure 1.
1.2. Remote Detection, Location and Characterization of Explosives Another device that involves remote measurements through optical fibers is also being developed at the University of Idaho. The initial purpose of this instrument is to detect the presence of explosive material via fluorescence quenching. Once the source of this material has been determined, the identity of the suspected explosive will be investigated by fiberoptic Raman spectrometry. This instrument has been dubbed a "Bomb Sniffer".
Fast Mapping System
Figure 1. Schematic representation of three-dimensional integrated characterization and archiving system.
171 Since most explosive materials contain nitro groups, the purpose of the first component of this instrument is to identify the presence of nitro-containing molecules present at low levels in the atmosphere. The key component of this device is an optical fiber at the end of which perylene groups have been immobilized. Perylene groups are strong fluorophores that are efficiently and selectively quenched by nitro compounds. When air is drawn over the end of the fiber, all organic molecules of low or medium polarity that are present dissolve in the perylene layer but only those compounds containing nitro groups quench the fluorescence. By monitoring the fluorescence signal at the return fiber, the presence of nitro-compounds at sub parts-per-billion levels in the atmosphere can be detected. The magnitude of the signal loss as the position of the fluorescence probe is changed gives a strong indication of the source of the putative explosive. Once the presence of the suspected explosive has been detected, its location and identity can be confirmed by fiber-optic Raman spectrometry, as will be described in this paper.
2. S P E C T R O S C O P I C AND O P T I C A L C O N S I D E R A T I O N S To determine the best approach for identifying materials through optical fibers, a variety of spectroscopic and optical parameters must be considered. These include signal-tonoise ratio, spectral contrast, effect of sample morphology, attenuation of source radiation by the optical fiber. The possibility of several mechanisms contributing to a given spectrum should also be questioned. Examples of interfering spectroscopic phenomena include the effect of specular reflection on diffuse reflection measurements and the degradation of Raman spectra by fluorescence or thermal emission. In addition, because the instrumentation for both applications described above will be used in the field, its size, weight and ruggedness must also be taken into account. Three vibrational spectroscopic techniques could be used for the remote characterization of solid samples through optical fibers: midand near-infrared external reflection spectrometries and Raman spectrometry. Let us consider the effect of various factors on each of these spectroscopic techniques.
2.1. Optical Fibers By far the best material to use as a fiber for remote vibrational spectrometry is fused silica, since it has excellent transmission characteristics to wavelengths as long as about 2200 nm. Fused silica can be used for Raman spectrometry with any commonly used laser and most NIR measurements. Several other types of fibers can be used for measurements at longer wavelengths. Heavy metal fluoride fibers can be used from visible wavelengths to about 5 gin; thus fundamental X-H stretching modes are accessible with fluoride fibers but the region below 2000 cm 1 is completely absorbed. Either chalcogenide or silver halide fibers are suitable for this region but neither these materials or fluoride fibers can be used at lengths exceeding about 20 meters. Furthermore, the transmission of silver halide fibers degrades rapidly on exposure to radioactive environments. From the standpoint of the fiber materials, therefore, both Raman and NIR external reflection spectrometry appear more promising than mid-infrared measurements.
172
2.2 Reflection Spectrometry The benefits of NIR spectrometry are reinforced when mid- and near-infrared reflection spectrometries of neat materials are compared. Provided that the effect of specular reflection is minimal and the scattering coefficient is constant, diffuse reflection (DR) spectra are readily changed to a form where band intensities are directly proportional to the absorptivity at each wavelength by conversion to the Kubelka-Munk function. However, although the scattering constant is reasonably constant 3, the assumption that the specular reflection spectrum can be ignored is far from true for strong fundamental modes, i.e., those where the imaginary refractive index, K, exceeds about 0.2). Strong features in the specular reflection spectra of organic compounds are dispersive in shape and have the effect of significantly distorting the diffuse reflection spectra of neat materials. This is readily seen by comparing the relative band intensities and band shapes in the Kubelka-Munk spectra of neat finely-powdered caffeine and a 1% dispersion of the same material in KBr, see Figure 2. This situation is even worse for many inorganic materials containing oxyanions, for which K can often exceed 1. In the region of such strong bands, the specular reflection will dominate the diffuse reflection, see Figure 3. The smoother is the surface of large bulk samples or the larger the particle size of powdered samples, the greater will be the contribution of specular reflection to a mid-infrared diffuse reflection spectrum. Thus even though the spectral contrast of mid-infrared DR spectra can be excellent (as is evident in Figures 2.b and 3.b), the distortion of the measured spectra because of contributions of both specular and diffuse reflection mechanisms and the effect of the sample morphology on the relative contribution of each mechanism would make it difficult to identify the materials by conventional spectral searching techniques.
&
17])0
16b0
15100
14'00
13'00
1200
wavenumber (era-l) Figure 2: Mid-infrared diffuse reflection spectra of caffeine in the region of two strong absorption bands; (upper trace) neat analyte, (lower trace) diluted to 1% in KBr.
173
J
]
20100
15100 wavenumber (era-l)
10100
Figure 3: Mid-infrared diffuse reflection spectra of porcelain in the region of the strongest absorption band; (upper trace) neat analyte, (lower trace) diluted to 1% in KBr.
Since the molecular vibrations leading to absorption in the near infrared region are all either overtones or combinations of C-H, O-H or N-H stretching fundamentals, the values of K for these bands are well over an order of magnitude less than those of the fundamental modes from which these bands are derived. As a result, the contribution of specular reflection to NIR external reflection spectra is minimal. Furthermore, the weakness of bands in the NIR spectra of most compounds is advantageous, as it is very rare that any band in an NIR DR spectrum of a neat organic material exhibits saturated absorption and yet the spectral contrast is usually quite acceptable. Unfortunately, few bands in the spectra of inorganic materials that absorb strongly in the mid infrared may be seen in their NIR spectrum. The reason may be found in the fact that most strong fundamentals in the spectra of inorganic materials absorb below 1500 cm J. For example, if we consider the intense 1100-cm -~ band of porcelain, its first and second harmonics both absorb in the mid-infrared spectrum and it is not until the third harmonic does an overtone absorb in the NIR region. Since the anharmonicity coefficient of these fundamentals is quite low, their third harmonic is too weak to lead to observable bands in the NIR DR spectrum. Only if these materials are hydrated or their surface is hydroxylated will inorganic materials give rise to reasonably strong absorption bands in the NIR spectrum. While this effect may make the identification of anhydrous inorganics impossible by NIR reflection spectrometry, it is very beneficial for the detection of asbestos, as we will see later.
174
2.3 Spectral Information Content Most spectroscopists would assert that it is far easier to interpret mid-infrared or Raman spectra than NIR spectra, since most of the stronger bands are caused by fundamental vibrational modes. However, in an instrument to be used for applications of the type described here, automated approaches to identification or classification (such as spectral searching or artificial neural networks, respectively) are applied. In these cases, the timehonored "Colthup-chart" approaches to spectral interpretation are not required. Thus it makes little difference whether the spectral features are derived from fundamentals or from overtones and combinations, provided that the spectral contrast is sufficiently high. From this argument, it may be believed that NIR spectra would have the same spectral information content as mid-infrared or Raman spectra, but this is not necessarily so. As noted in the previous paragraph, most features in NIR spectra are derived from X-H stretching modes. Thus if the key fundamental modes needed to identify a material absorb below about 2000 cm ~, they will be unlikely to contribute substantially to the NIR spectrum unless they are "carried" into this region by combination with an appropriate X-H stretching mode. 2.4 Raman Spectrometry Until about five years ago, Raman spectrometry would not usually have been considered for remote analysis of materials because of its low sensitivity and speed of analysis. The development of Raman spectrometers in which the sensitivity has been dramatically increased through the use of either multiplex (Fourier 4 or Hadamard 5 transform) or multichannel (array detection6'7'8) techniques has changed this situation dramatically. Since most Raman spectra require detection in the visible or short-wave NIR region (where silica fibers are readily applicable), Raman spectrometry now represents an attractive option for the remote characterization of materials through optical fibers. The largest drawback to the use of Raman spectrometry with visible excitation for the investigation of samples in the field is the effect of fluorescence. Since fluorescence is largely eliminated when radiation of wavelength greater than 1000 nm is used, and because Nd3+:YAG lasers emitting at 1064 nm have enough power to offset the v 4 dependence of Raman band intensities, the application of Fourier transform (FT) Raman spectrometers equipped with a Nd3+:YAG laser and either a Ge or an InGaAs detector would seem to be very attractive. However, the use of high-powered lasers is often limited when studying "real-world" samples as some inorganic materials fluoresce even under 1064-nm irradiation or are heated because of absorption of this radiation and hence emit thermal photons at high Stokes shifts. Multichannel Raman spectrometers incorporating a diode laser emitting between 780 and 820 nm and silicon charge-coupled device (CCD) array detector present an attractive alternative approach to FT-Raman spectrometers. Although diode lasers rarely generate radiation of much greater power than about 25 roW, and hence are much less intense than Nd3+:YAG lasers, silicon CCDs generally have significantly lower noise equivalent powers than Ge or InGaAs detectors. Because the intensity of bands in Raman spectra is proportional to the fourth power of their absolute wavenumber, CCD Raman spectrometers usually yield spectra of high signal-to-noise ratio (SNR) in short measurement times, and hence are very competitive with FT-Raman spectrometers for field measurements (provided that sample fluorescence for excitation at these wavelengths is comparable). The primary drawback for CCD Raman spectrometers is the fact that the long-wavelength cut-off of
175
silicon CCDs is about 1000 nm. Thus if a 780-nm diode laser is used, the greatest Raman shift that can be observed is about 3000 cm -j, while if the wavelength of the diode laser is increased to 820 nm, the maximum Raman shift is about 2400 cm ~. The optimum technique (FT-Raman or CCD-Raman) and laser wavelength for obtaining Raman spectra of "real-world" samples must has been addressed in our laboratory. The effect of various parameters on the relative utility of FT-Raman and CCD Raman for field monitoring are summarized in Table 1, below. Note that we have not included the Ti:sapphire laser in this table in view of its expense and weight, which precludes its use for field monitoring applications. We would also note that several new NIR lasers with very attractive properties, such as the "photon cannon", are also under development 9 but, since they are not yet commercially available, they too have been omitted from Table 1.
Table 1 Factors affectin~ the relative sensitivity of FT-Raman and CCD-Raman spectrometers
Factor
1064 nm
780 nm
Laser Power
Up to 2 W at the sample
Rarely more than 10 mW at sample
Detector sensitivity
Relatively poor
Relatively high to moderate
v 4 effect (1500-cm ~ band)
x
4.2 x
Thermal emission
Relatively high
Relatively low
Fluorescence
Low for non-black samples Slightly higher than at 1064 nm
Field portability
Good (for DPY)
Medium
A detailed analysis of the relative importance of each of these parameters is beyond the scope of this paper but it is believed that the advantages of the high sensitivity of silicon CCDs, the low power of diode lasers giving rise to low thermal emission, the advantageous v 4 effect, and the improved field portability of CCD-Raman instruments more than offset the benefits afforded by the high power of the Nd3+:YAG laser and the reduced probability of sample fluorescence with 1064-nm excitation. 2.5 Goal of this W o r k The goal of the first phase of the development of both the 3D-ICAS and the "Bomb Sniffer" was to investigate the strengths and weaknesses of mid-infrared and NIR reflectance spectrometry and FT-Raman and CCD Raman spectrometry for these projects. The results of these investigations will be described below.
176 3. RESULTS OF 3D-ICAS INVESTIGATION 3.1 Mid-Infrared Reflection Spectrometry Mid-infrared external reflection spectra of neat materials almost invariably show contributions from diffuse and specular refection mechanisms, as demonstrated in Figures 2.a and 3.a. Thus, even if chalcogenide or silver halide fibers were as efficient in the mid infrared as silica fibers are in the near infrared and visible, mid-infrared external reflection spectrometry would still be precluded as the technique of choice for these projects. 3.2 Near-Infrared Reflection Spectrometry Although the SNR and information content of NIR reflection spectra of natural organic materials, such as woods, are excellent, several other types of samples tested, including brick, concrete and asphalt, yield poor NIR spectra. Surprisingly, however, several samples of asbestos and asbestos-containing materials have strong features in their NIR reflection spectra; as an example, the spectrum of the most common form of asbestos, chrysotile is shown in Figure 4.a. The features in this spectrum are the first overtones of the Si-O-H stretching vibrations that absorb about 3700 cm -1, see Figure 5. The strength of these bands, together with the significant difference between the band profile in the mid and near infrared, suggests that the anharmonicity coefficient for this vibration is fairly large, and this conclusion is reinforced by the fact that each feature in the NIR spectrum absorbs at a wavenumber that is at least 175 cm -1 below twice that of the corresponding fundamentals in the mid-infrared spectrum of chrysotile. The excellent SNR and sharpness of the NIR bands allow one to readily identify the presence of asbestos in a sample and to differentiate between all the common types of asbestos, chrysotile, crocidolite, amosite and tremolite. (a)
es
L~
Ca)
74'00
73b0
72~0
71~
70~0
Waveuumber(em-l) Figure 4: Near-infrared reflection spectra of (a) chrysotile and (b) transite.
177
e~
\
3~0~3
~00
36r50-36b0
35150 .... ~
--34150~4
Wavenumber (cm-l)
Figure 5: Mid-infrared reflection spectrum of chrysotile in the region of the Si-OH stretching vibration.
As noted in the Introduction, U.S. federal law mandates that all asbestos-containing materials, including transite, must be disposed of in an approved manner. Transite is an asbestos/concrete composite that has found utility where the fire-retardant properties of asbestos are required along with the mechanical durability of concrete. A further advantage of transite over asbestos is that, in light of the greater mechanical strength of transite, the probability of asbestos fibers entering the atmosphere is reduced, so that the detrimental health effects of asbestos exposure are minimized. (These beneficial properties do not, of course, prevent transite being disposed of in the same manner as neat asbestos.) The NIR reflection spectrum of transite is shown in Figure 4.b. By comparison with the spectrum in Figure 4.a, it is readily apparent that the form of asbestos contained in this transite sample is chrysotile. This result is not altogether surprising as chrysotile is the most abundant (95%) naturally-occurring form of asbestos. 3.3 Near-Infrared Raman Spectrometry Raman spectroscopic studies of asbestos with 1064-nm excitation are limited by both the low Raman scattering cross-sections of the materials and thermal heating of the sample caused by absorption of the laser radiation. Shifting the excitation wavelength to 785 nm has allowed Raman spectra to be obtained above a background for the most common asbestos minerals, chrysotile and crocidolite. Neither excitation wavelength was routinely successful for the study of transite due to thermal heating of the concrete at 1064 nm and fluorescence at 785 rim. If the NIR Raman spectra of asbestos are compared with those obtained by NIR reflection spectrometry, it was found that NIR DR spectrometry has the advantage of obtaining spectra from all the varieties of asbestos and transite in a time that was 15 to 50 times faster than the Raman measurement for equivalent SNR. For most organic samples, on the other hand, Raman spectrometry proved to be the more useful technique.
178
The design of the probe to be used for the measurement of the Raman spectra is critical to the success of these measurements. Typically, the laser radiation must first be passed through a notch filter located before the sample to eliminate laser lines at other than the desired excitation wavelength and radiation resulting from Raman scattering by the input fiber. However, since Rayleigh scattering from the sample is so much more intense than Raman scattering, it is also necessary to filter out the Rayleigh-scattered radiation before it enters the return fiber to preclude Raman scattering from the fiber caused by Rayleigh scattering at the sample. A probe design that incorporates these features was first described by Carrabba ~° and is shown schematically in Figure 6. Similar designs have recently become commercially available from three companies (Kaiser Optics, EIC Corporation and Dilor). This type of probe design has the following advantages over the initial fiber-optic probes used for Raman spectroscopyllA2'13'14'15'16: Focusing lenses which increase the intensity of the Raman signal observed by increasing the laser irradiance per the unit volume of the sample; ii.
Filters that prevent the Rayleigh radiation from entering the collection fiber and hence preventing the Raman signal from the return fiber(s) from obscuring the Raman signal from the samples being investigated;
iii.
Probe head protection by sealing the probe head behind a protective window.
We are planning to use a probe based on this design in both the 3D-ICAS and bomb sniffing projects. Fibre-Optic (single fibre)
Rayleigh Rejection Filter {not required if Holgraphic Beamsplitter is used }
tlolographic Beamsplitter {or Normal Beamsplitter}
Microscope Objective
/I / Fibre-Optic (single fibre)
Laser line Filter
Collimating
Mirror
{diagram not to scale}
Lens
Figure 6: Optimal probe design for the measurement of Raman spectra through optical fibers.
179
3.4 Summary of Results Several materials were characterized by all four of the vibrational spectroscopic techniques 17 and the results are summarized in Table 2. From this table, it can be seen that Raman spectrometry using both 780-nm and 1064-rim excitation is applicable to the identification of bricks, polymers and some woods. A few samples of asbestos yielded Raman spectra with 785-nm excitation, but the strength of the first overtone of the silanol stretching mode led to the conclusion that NIR external reflection spectrometry was by far the best technique for field monitoring of this material. While spectral contrast of mid infrared spectra was good, the low efficiency of fibers in the mid infrared region of the spectrum precludes their use.
Table 2 Spectral quality of mid-infrared, near-infrared and Raman Spectra of several classes of samples Sample Class
Mid Infrared
Near Infrared
Raman (1064 nm)
Raman (785 nm)
Bricks
G
G-P
M
G*
Polymers
M
VG
G
G
Asbestos
G
VG
M-P
G-P
Woods
P
VG
M-P
VG
Asphalt
VP
VP
VP
VP
VP = very poor, P = poor, M = moderate, G = good, VG = very good * Fluorescence profile indicative of bricks observed under all conditions
4. R E S U L T S ON E X P L O S I V E M A T E R I A L S Since most explosive materials used for terrorism are usually transported and used in the form of a gel or a plastic putty (Semtex, C4, etc.), the application of NIR reflectance spectrometry for the remote characterization of these materials is precluded. In this case, fiber-optic Raman spectrometry is the only logical choice. NIR-Raman spectra can be easily measured without removing the materials from their glass (and even plastic TM) containers. Most of the pure explosives do not fluoresce significantly even under 632.8-nm excitation and acceptable spectra can also be obtained with either diode or Nd3+:YAG lasers. The largest problem associated with the automated characterization of materials with either type of laser is usually the presence of a continuous spectral background caused by very weak thermal emission, as even a relatively low background can lead to the failure of many spectral searching or discriminant analysis algorithms. Most current baseline-correction algorithms, however, require a significant amount of interaction by a trained spectroscopist. Since in their
180
final manifestation, the instruments being developed in this project will be operated by nonscientists, we came to the conclusion that there is a strong need for developing a classification algorithm that can be applied without operator interaction. The most common way of correcting for the baseline is for the spectroscopist to select regions where no vibrational modes are believed to contribute to the spectrum. One or more data points in these regions are typically fitted to piecewise linear functions or a polynomial function that represents the operator's best estimate of the baseline. This estimated baseline is then subtracted from the measured spectrum to yield the baseline-corrected spectrum. The key to automating a baseline-correction algorithm is to find a way of selecting the baseline points without operator interaction. During the manual selection of these points, the operator typically selects spectral regions where the slope is either constant or only varying very slowly, i.e., regions where the second derivative of the spectrum is approximately zero. We have automated this modus operandi by calculating the second derivative of the spectrum and selecting short regions at approximately 100-cm -1 intervals where the second derivative is approximately constant. The central data point in each region is designated as representing the baseline point; these points are then linked by straight lines and subtracted from the measured spectrum. The success of this approach can be seen from Figure 7.
t~
\
!OriginalS p e c t r u m ~
SecondDeffvativeSpectrum
Baseline-Corrected Spectrum 3500
3000
2500
2000 Raman Shift
1500
1000
500
(cm-l)
Figure 7: Steps involved in automated baseline correction algorithm.
181
Although we have not yet incorporated a fiber-optic probe for the remote measurement of the Raman spectra of explosive materials, we are at the point where we can confidently predict that, using a probe of the type shown in Figure 6 above, the measurement of machineinterpretable spectra of explosive materials in a variety of environments is possible.
5. NEURAL NETWORKS FOR DISCRIMINANT ANALYSIS The goal of both the projects described above is to classify materials as falling into a certain class, e.g., "asbestos" or "not asbestos", "explosive" or "non-explosive". As noted above, it is not economically viable to employ an experienced spectroscopist operating every one of these instruments. Their software must, therefore, completely preclude interaction by a trained scientist and must provide a technician with an output where there is a very high probability that the material does indeed fall into the category indicated by the instrument. We have investigated the application of artificial neural networks for this purpose. 5.1 Introduction Pattern classification, clustering and feature extraction of molecular vibrational spectra are important applications of chemometrics. Recently, artificial neural networks have been shown to be robust and efficient computing technologies which complement chemometric techniques for the exploratory analysis of infrared and Raman spectra 19'2°'21. Neural networks can also be constructed as pre-processors for conventional spectral analysis methods such as principal components analysis (PCA), cluster analysis, partial least squares (PLS), and spectral library searching 21'22. The modularity of artificial neural networks allows for the embedding of trained networks into programmed computing systems such as expert systems resulting in the creation of powerful and flexible pattern recognition systems known as hybrid intelligent systems or expert neural network tools 23"24'25'26. Feed-forward or mapping neural networks are common neurocomputing paradigms for the stand-alone pattern recognition of infrared and Raman spectra. Mapping networks trained by error back-propagation are the most prevalent pattern classification neural networks. In practice these neural network based pattern recognition systems have demonstrated near optimal classification performance 27. Spectral classification accuracies of 95 % are routinely achieved by properly trained feed-forward back-propagation networks 19'2°. In light of the demonstrated success of such networks for spectral classification, our current research on the pattern classification and identification of materials from their Raman spectra has concentrated on the use of this type of neural network. 5.2 Results When the spectra of all members of a particular class of compounds are very similar, and significantly different from the spectra of members of all other possible classes, neural computing allows these materials to be rapidly and accurately classified. For example, it is a simple matter to distinguish the Raman spectra of woods from the spectra of most other materials such as polymers and pharmaceuticals. Once a material has been classified as a wood, it is then possible to train a neural network to recognize the difference between the spectra of hard and soft woods, i.e., to perform sub-classification. On the other hand, when the molecular structure of materials that one would like to fall in the same class are quite
182
different, the potential of neural networks for successful discriminant analysis is less apparent. Explosives, which usually contain nitro groups but nonetheless have very different chemical structures, fall into this category. Nevertheless, surprisingly successful results were achieved in the tests that are described below. The Raman spectra of 21 explosives and a number of non-explosives were available for this study. 21 spectra of non-explosives were selected as being representative of materials that may be encountered in airport luggage. Ten training sets were used, with the spectra of 18 explosives and 18 non-explosives being randomly chosen for each set. The three remaining spectra in each class were used for testing. This system of using approximately 90% of the database for training and approximately 10% of the database for testing was repeated ten times. Thus, ten independent network training sessions (each taking around 90 seconds) were performed. The network was trained to give output values of + 1 for explosives and -1 for non-explosives. For the first studies of this system (i.e., before any attempt at baseline correction or optimization of the parameters associated with the neural network had been made), the results were remarkably successful. A graphical representation of the neural network output is given in Figure 8. Since most of the spectra of explosives that had been selected for this study did not exhibit non-zero spectral baselines while the spectra of several of the non-explosives showed the effects of fluorescence and/or thermal emission, application of the automated baseline-correction algorithm baseline correction described above did not significantly improve the results in this case. However, when the spectra of compounds falling inside the category of interest exhibit baseline changes, the effect of baseline correction is expected to be far more important.
1 0.8
0000
0 0 0 0 @ 0 0 0 0 0 0 0 0 0 0
0000
0000
0
0
0
0.6 0.4
¢3.
0.2
0 0 0
I
I
I
I
I
I
5
10
15
20
25
30
II o explosive
I 0 nonexplosive
-0.2 Z -0.4 -0.6
0
-0.8
0 -1
0000000000
0
0000000000
000000
Figure 8: Classification of explosive and non-explosive materials from their Raman spectra using a neural network.
183 6. CONCLUSIONS The remote identification or classification of materials by vibrational spectrometry using fiber-optic probes is always better achieved by NIR Raman spectrometry than by midinfrared DR spectrometry because of both the improved efficiency of optical fibers in the near infrared and the absence of mixed reflection mechanisms. While quantitative measurements of appropriate samples by near-infrared spectrometry have proved to be highly successful, the inapplicability of NIR spectrometry for many inorganic materials sometimes precludes its use for the characterization of hazardous wastes. Raman spectrometry provides useful data for a wider variety of materials than NIR, but neither technique can be used for the classification of all materials of interest in hazardous waste sites. For example, NIR reflection spectrometry is more useful for the remote identification of asbestos-containing materials whereas NIR Raman spectrometry is more useful for plastics and organic solvents and chemicals (including explosives). These classes of materials can be readily distinguished using artificial neural networks, if necessary after baseline correction of Raman spectra. An instrument that will allow both NIR reflection and NIR Raman spectra to be obtained through fiber-optic probes is being developed in our laboratory and preliminary results should be available in about a year.
7. A C K N O W L E D G E M E N T S We wish to gratefully acknowledge the Department of Energy (through a subcontract from Coleman Research Corporation) and the Federal Avaiation Administration for the financial support for this work and to the Federal Bureau of Investigation for sending us the samples of explosives. We would also like to thank the Perkin-Elmer Corporation for the donation of the System 2000 Fourier tansform spectrometer and Renishaw and Bruker Instruments Companies for allowing one of us (IRL) to make measurements on the instruments in their Applications Laboratories in England. We would also like to note that Professor Ray von Wandruszka's research group at the University of Idaho is developing the fiber optic probe for the Bomb Sniffer.
REFERENCES 1.
R. Mackison, S . J . Brinkworth, R. M. Belchamber, R. E. Aries, D, J. Cutler, C. Deeley and H. M. Mould, Appl. Spectrosc., 46, 1020 (1992).
2.
K . P . J . Williams, Royal Society of Chemistry Meeting on FT-Raman Spectroscopy, Coventry, U.K. (1992).
3.
D . J . J . Fraser and P. R. Griffiths, Appl. Spectrosc., 44, 193 (1990).
4.
T. Hirschfeld and D. B. Chase, Appl. Spectrosc., 40, 133 (1986).
5.
A . P . Bohlke, J. D. Tate, J. S. White, J. V. Paukstelis, R. M. Hammaker and W. G. Fateley, J. Mol. Struct., 200, 471 (1989).
184 6.
A. Campion and W. Woodruff, Anal. Chem., 59, 1299A (1987).
7.
J. E. Pemberton and R. L. Sobocinski, J. Am. <'hem. Soc., 111, 432 (1989).
8.
Y. Wang and R. L. McCreary, Anal. Chem., 61, 2647 (1989).
9.
M. M. Carrabba, E1C Corporation, personal communication to I. R. Lewis (1994).
10.
M. M. Carrabba, Fiber-Optic Raman Spectrograph for in-situ Environmental Monitoring, DOE Final Report, Contract No. 02112402 {DOE/CH-9205} (1992).
11.
C. G. Zimba and J. F. Rabolt., Appl. Spectrosc., 45, 162 (1991).
12.
C. D. Allred and R. L. McCreery, Appl. Spectrosc., 44, 1229 (1990).
13.
B. Schrader, M. Tischer, R. Podochadlowski, and H. Schlemmer, Section, 19.11, Xr h International Conference on Raman Spectroscopy (ICORS XI), (Eds. R. J. H. Clark, and D. A. Long), John Wiley and Sons., Ltd., London, p. 957 (1988).
14.
C. Wang, T. J. Vickers, J. B. Schlenoff, and C. K. Mann, Appl. Spectrosc., 46, 1729 (1992).
15.
S. D. Schwab, R. L. McCreery, and F. T. Gamble, Anal. Chem., 58, 2486 (1986).
16.
D. Gerard, Analytical Raman Spectroscopy, (Eds., J. G. Grasselli and B. J. Bulkin, Wiley-Interscience, New York, Chapter 9 (1991).
17.
I. R. Lewis, N. C. Chaffin and P. R. Griffiths, Proc. 9th Int. Conf. on Fourier Transform Spectrosc., (J. E. Bertie and H. Weiser, eds.), SPIE, Bellingham, Proc. Soc. Photo-Opt Instrum. Eng., 2089, 454 (1993).
18.
D. A. C. Compton and S. V. Compton, Appl. Spectrosc., 45, 1587 (1991).
19.
N . W . Daniel, Jr. and P. R. Griffiths, Proc. 9th International Conference on Fourier Transform Spectroscopy, (J. Bertie and H. Wieser, eds.), SPIE, Bellingham, Proc. SPIE, 2089, 230 (1993).
20.
I. R. Lewis, N. W. Daniel, Jr., N. C. Chaffin, and P. R. Griffiths, Spectrochim.
Acta, in press (1994). 21.
J. Zupan and J. Gasteiger, Neural Networks for Chemists, VCH Publishers, New York, (1993).
22.
P. J. Gemperline, J. R. Long and V. G. Gregoriou, Anal. Chem. 6, 63 (1991).
23.
L. R. Medsker, Hybrid Neural Network and Expert Systems, Kluwer Academic Publishers, Boston (1994).
185
24.
R. C. Eberhart and R. W. Dobbins, Neural Network PC Tools.'A Practical Guide, Academic Press, San Diego (1990).
25.
J. Schmuller, in Expert Systems for Environmental Applications, (J. M. Hushon, ed.), ACS Symposium Series 431, American Chemical Society, Washington DC (1990).
26.
B. Soucek, Neural and Intelligent Systems Integration, John Wiley and Sons, Inc., New York (1991).
27.
R. Shadmehr and D. D'Argenio, Neural Computation, 2, 216 (1990).