I

DISCLAIMER

This report was prepared as an accouht of work sponsored by an agency of the United States Government. Neither the United States Government nor any agency thereof, nor any of their employees, makes any warranty, express or implied, or assumes any legal liability or responsi- bility for the accuracy, completeness, or usefulness of any information, apparatus, product, or process disclosed, or represents that its use would not infringe privately owned rights. Refer- ence herein to any specific commercial product, process, or service by trade name, trademark, manufacturer, or otherwise does not necessarily constitute or imply its endorsement, recom- mendation, or favoring by the United States Government or any agency thereof. The views and opinions of authors expressed herein do not necessarily state or reflect those of the United States Government or any agency thereof.

3 $3 tl P, op

(D 3 e

5 C n I '

3 03

DISCLAIMER

Portions of this document may be illegible in electronic image products. Images are produced from the best available original document.

Chemical and Isotopic Determination from Complex Spectra

ki and Richard B. Strittxnatter Safeguards Systems Group

Los Alamos National Laboratory Los Alamos, NM 87545 USA

ABSTRACT Challenges for proliferation detection

include remote, high-sensitivity detection of chemical effluents from suspect facilities and enhanced detection sensitivity for nuclear material. Both the identification of chemical effluents with lidar and enhanced nuclear material detection from radiation sensors involve determining constituents from com- plex spectra. In this paper, we extend tech- niques used to analyze time series to the analysis of spectral data. Pattern identification methods are applied to spectral data for domains where standard matrix inversion may not be suitable because of detection statistics. We use a feed-forward? back-propagation neural network in which the nodes of the input layer are fed with the observed spectral data. The nodes of the output layer contain the identification and concentration of the isotope or chemical effluent the sensor is to identify. We will discuss the neural network architec- ture, together with preliminary results obtained from the training process.

I. INTRODUCTION

relevant to proliferation detection are optical measurements and gamma-ray spectroscopy. When using differential absorption lidar mea- surements,' our ultimate goal is plume identi- fication in the presence of deleterious effects,

The two main sources of spectral data

This work supported by the US Department of Energy, Office of Research and Development

the most important of which are speckle and background reflectance. Similarly, accurate measurement of plutonium and uranium iso- topic concentration has widespread applica- bility for domestic nuclear materials control and IAEA verification? Determination of plu- tonium composition by gamma-ray spectro- scopy using high-resolution germanium detectors is a standard method today. The ability to extract more accurate information from lower-resolution detectors that operate at room temperature, such as NaI, would facilitate many inspection efforts.

In a matrix analysis of multi-spectral optical data, one needs to know the values of the absorption coefficients of the atmospheric constituents at various wavelengths. A feed- forward neural network acts as a black box, giving the concentrations of different species based on experimental or simulated data. The goal of this paper is to apply the back-propa- gation algorithm to the results of the low- resolution LOWTRW and the high-resolu- tion FASCODE4 transmission codes. To this end, we use a three-layer network in which the input pattern consists of spectral transmit- tance at various wavelengths. The three-node output layer refers to concentrations of H20, CO,, and 0,. By varying the concentration levels of these gases with respect to standard conditions, we generate the network training set. The simulation results coniirm the well- known fact that the neural nets are not very sensitive to the impact of noise. Further study

will extend the input-pattern feature vector by including the atmospheric ambient condi- tions.

To generate the gamma-ray spectral data we use SYNTH, a spectrum synthesizer developed by Pacific Northwest Laborat~ry.~ In this context the principal advantage of SYNTH is its ability to vary the isotope con- centration, thus enabling one to obtain a series of spectra corresponding to well-defined sam- ples.

II. SPECTRAL DATA

Molecular absorption determines radia- tive heating and cooling in the infixred, thus affecting the climate modeling. Transmission of laser beams and the determination of trace gas concentrations are exacting requirements, only partially fulfilled with high-resolution spectr~scopy.~ Gamma spectroscopy experi- ments allow one to determine the quantity of nuclides producing the radiation. The accu- racy is determined in part by our ability to extract information from the complex and overlapping spectral data.

In this section, we show three plots describing some standard optical and gamma- ray spectra. To model different concentration levels of atmospheric gases as well as isotopic compositions, we scale the intervening mate- rials by a desired factor with respect to stan- dard conditions. This approach, however time consuming, leads to an effective library of data, which consists of records giving intensi- ties at different wavelengths corresponding to desired material concentrations.

A. Optical Transmittance Figure 1 shows the amspheric trans-

mittance as a function of radiation wavelength obtained with the aid of the low-resolution LOWTRAN model. For concreteness, we

select a vertical path with the observer at an altitude of 10 km looking towards the surface of the earth. A general feature of this plot is high transmittance (low absorption) in the vis- ible part of the spectrum around 0.5 pm. On the other hand, in the infrared the transmit- tance is very low, apart from atmospheric windows around 3,5, and 10 pm.

1 .o

0.8

a, u 6 0.6

E

F

u c .-

ul

0.4 k-

0.2

0.0 5 10 15 20

Wavelength (micron)

Fig. I . Low-resolution atmospheric trans- mittance as a function of wavelength.

The high-resolution FASCODE pro- duces the transmittance shown in Fig. 2 corre- sponding to a horizontal path located 0.5 km above the ground level, with the extent of 10 km. Strictly speaking, although transmission and transmittance are used here interchange- ably, transmission is a physical process in which light penetrates the medium without being absorbed, whereas transmittance is the percentage of energy being transmitted.

FASCODE efficiently carries out line by line computations of atmospheric absorption using line parameters, the basic line shape, and absorption equations. Because of its accu- racy, FASCODE is the basic tool today used for modeling laser light transmission.

WRVELENCTH I M l C R O N E T E R l 9.32 9.31 9 .30 9 . 2 9 9 .28 9 . 2 1 9.26 9 . 2 5

, -- 1-- 1.0

0.8

0.6

0.11

0 . 2

0.0 IUI.2 1013 1U15 1016 10)7 l U l B IOMU 1081 1082

W R V E N U M B E A . I C H - I I

Fig. 2. High-resolution atmospheric transmittance as afunction of wavenumber and wavelength.

B. GammaSpectra The computer code SYNTH allows a

user to specify physical characteristics of a gamma-ray source, the quantity of the nuclides producing the radiation, the source- to-detector distance and the type of thickness of absorbers, the size and composition of the detector (germanium or NaIj, and the elec- tronic setup used to gather data.5

As in measurements of Collins: we con- sider a 1-g plutonium sample whose composi- tion by weight is 93% Pu-239,5% Pu-240, and 0.15% Pu-241, together with trace amounts of Pu-238, Pu-242, and Am-241. The low-resolution NaI detector located a few centimeters away from the sample generates the spectrum consisting of 5 12 channels, cali- brated to 1.5 keV per channel. The synthe- sized spectrum obtained from the SYNTH code is illustrated in Fig. 3. For the prelimi- nary proof-of-concept analysis, spectra were computed that varied only in Pu-241 and Pu-239 concentrations. The synthesized spec- tra and the neural network method were used to determine only the Pu-241 content.

Fig. 3. Synthetic gamma-ray spectrum of a plutonium sample.

III. NEURAL NETWORK ARCHITECTURE A multilayer feedforward network con-

sists of a set of neurons that are logically arranged into two or more layers. There is an input layer and an output layer, each contain- ing at least one neuron. We introduce a hidden layer sandwiched between the input and out- put layers. The term feedforward means that information flows in one direction only. The strength of the connections between neurons (weights) is adjusted by the back-propagation algorithm, based on an iterative procedure.

The artificial neural networks (ANS) we consider are composed of a large number of simple processing elements called units, each interacting with others via excitatory or inhib- itory connections.' Several major aspects of the A N S model are common to all neural net- works:

Processing units. 0 State of activation.

Output function for each unit. Pattern of connections among units. Propagation rule for propagating patterns of activities.

e

e

Activation rule for combining inputs, affecting a unit with its present state to pro- duce an output. Learning rule whereby interconnections can be modified on the basis of experience. An environment within which the learning system must Operate.

Each processing unit determines a net input value based on all input connections. Because the linear composition rule applies here, we can write the net input to the ith unit as

neti = C x j w i j , i

where the index j runs over all connections to the processing unit and where the w's are referred to as weights. Once the net input is calculated, it is converted to an activation value, or simply activation, for the processing element:

ai = F i ( u i ( t - l ) , n e t j ( t ) ) ,

in which we indicated that the current activa- tion depends on the activation defined one time step prior to time t.

The output value of the processing unit is determined in terms of an output function

xi = f i (ai) . (3)

Usually ai = neti, so this function is writ- ten as

xi = f i (ne t i ) . (4)

As the new data are accumulated, the weights corresponding to different unit con- nections are adjusted; the process is called network learning. A typical neural network

consists of an input layer, an output layer, and one or more hidden layers. This is illustrated in Fig. 4.

OutputPattern

Input Pattern

Fig. 4 . Schematic representation of a neural net.

The network repsented in Fig. 4 is des- ignated to operate as a multilayer, feedfor- ward network, using the supervised mode of learning. Other potentially useful networks are the bidirectional associative memory (BAM) system, the Hopfield network, and the counterpropagation network.

IV. SIMULATION RESULTS

Table 1 lists the wave numbers selected for this analysis of optical spectra, together with the selected energies of the gamma-ray spectrum. Note that the relation between wavelength X (p) and wavenumber k (cm-') is k = loo00 / A. The low-resolution optical spectrum has the width of 5 cm-', that is, the numbers listed in the first column of Table 1 denote the lower spectral bound. The high- resolution optical spectral data are obtained by integrating over three adjacent FASCODE lines, each line having a width of 0.02 cm-'; in other words, the numbers listed in the second column of Table 1 correspond to the line cen- ter. For the preliminary proof-of-concept

TABLE 1. Wavenumbers and energies used for pattern recognition

Wavenumbers Wavenumbers Energles (cm-') for bw- (an-') for high- (Irev) for

opticalspectrum opticnlspeclrum spectrum resdution resolution gamma-ray

665.0 1072.65 58.5

730.0 1073.33 100.5 ~ ~

1050.0 1074.47 120.0

1290.0 1074.71 205.5

TABLE 2. Wakr concentration retrieval based on LOWTRAN and FASCODE simulation.

Comen-

0.2

0.4

~

Fractional LOWTRAN FASCODE C% Percentage Percentage

Concen- Retrieval Retrieval trption Emr E m

0.2 10.9 25.1

0.4 0.8 21.2

0.6 16.1 1820.0 1078.75

analysis, the gamma-ray data consist of read- ings in a single channek the actual channel number is obtained by dividing the numbers in the third column of Table 1 by 1.5. Further analysis will be needed to optimize the selec- tion of channels and the clustering of channels for input to the net.

In the optical spectrum, the back-propa- gation neural network was trained by varying the concentrations of H20, C a , and 0, with respect to the standard conditions. We used the fractional concentration levels of 0.1,0.3, 0.5,0.7, and 0.9, resulting in 125 records, which served as our training set. Table 2 illus- trates the results of the recall procedure, in which we test the forecasting power of the network by using the concentration values independent of the training set.

It is evident that the wavelength selec- tion for the low-resolution spectrum was more judicious, resulting in better forecasting results based on LOWTRAN code. The high- resolution FASCODE leads to better results

0.6 1 .o 0.9 44.2

0.8 0.4 0.2 15.3

0.8 0.8 0.4 25.0

1.0 0.2 3.8 0.0

1 .o 0.6 4.4 0.0

when water concentration is not too small. This simply implies the wavelength region selected for simulation is appropriate for CO,, not water detection. Similar results were obtained when we simulated the C02 retrieval.

In this preliminary study, the isotope content retrieval is limited to Pu-241. Table 3 lists the actual and recalled values obtained after training the neural network with five input nodes and one output node. We experi- mented with different number of nodes in the hidden layer, varying between 3 and 12.

V. CONCLUSIONS We have formulated the basic algorithm

for species identification based on the neural- network supervised training. The algorithm does not need much information about the

absorption cross sections, as is the case when applying matrix inversion method^.^ In addi- tion, the algorithm is robust to noise. The examples taken from optical domain and gamma-ray spectroscopy demonstrate the usefulness of our method in proliferation detection and nuclear safeguards problems. When combined with the anomaly detection tool,l0 this method should also allow one to detect (exotic, unknown) species not found in the training data.

Actwl Amount (g)

0.001

0.002

0.005

0.010

0.020

0.040

0.050

TABLE 3. Retrieval of the Amount of Pu-241

RecPlled Amount (g)

0.002

0.003

0.005

0.009

0.020

0.040

0.049

0.075 0.075

0.100 0.097

REFERENCES 1.

2.

R. M. Measures, b e r Remote Sensing (Krieger Publishing, Malabar, Florida, 1992).

T. E. Sampson, “Plutonium Isotopic Composition by Gamma-ray Spectro- scopy,” in Passive Nondestructive Assay of Nuclear Materials, D. Reilly, N. Ensslin, H. Smith, Jr., and S miner, Eds. (U.S. Nuclear Regulatory Commis- sion, Washington, DC, 1991), pp. 221- 271.

3. F. X. Kneizys, E. P. Shettle, L. W. Abreu, J. H. Chetwynd, G. P. Anderson, W. 0. Gallery, J. E. A. Selby, and S. A. Clough, “User’s Guide to LOWTRAN 7,” Air Force Geophysics Laboratory report

4. H. J. P. Smith, D. J. Dube, M. E. Gardner, S. A. Clough, E X. Kneizys, and L. S. Rothman, “FASCODE-Fast Atmo- spheric Signature Code,” Air Force Geo- physics Laboratory report AFGL-TR-78- 008 (January 1978).

5. W. K. Hensley,A. D. McKinnon, H. S. Miley, M. E. Panisko, and R. M. Savard, “SYNTH: A Spectrum Synthesizer,” N u l . Muter. Manage. XXIII, 629-634 (1994).

6. T. G. Kyle, Annospheric Transmission, Emission, and Scatrering (Pergamon Press, Oxford, 1991).

7. D. E. Rumelhart and J. L. McClelland, Parallel Distributed Processing: Explo- rations in the Microstructure of Cognition (MIT Press, Cambridge, Massachusetts, 1986).

AFGL-TR-88-01777 (August 1988).

8. M. T. Collins, “Acquisition of Plutonium Spectral Data,” Los Alamos National Laboratory, June 1992 (unpublished memorandum).

9. J. R. Quagliano, P. 0. Stoutland, R. J. Romero, and R. K. Sander, “Chemo- metric Analysis of the Infrared Absorp- tion Spectra of Halogenated Hydrocarbon Gas Molecules,” CALIOPE Program, Second Annual Interim Technical Review, Los Alamos, March 28-30, 1995.

10. A. zardecki, “Fuzzy Control for Fore- casting and Pattern Recognition in a Time Series,” in Proceedings of IEEE Inter- national Conference on Fuzzy Systems, June 2629,1994 (IEEE Service Center, Piscataway, NJ, 1994), pp. 1815-1819.