Non-Destructive, Near-Real-Time Nuclear Safeguards Monitoring at a Reprocessing Facility

Home , Nuclear reprocessing

The Multi-Isotope Process Monitor: Non-destructive, Near-Real-Time Nuclear Safeguards Monitoring at a Reprocessing Facility

Dissertation

Presented in Partial Fulfillment of the Requirements for the Degree Doctor of Philosophy in the Graduate School of The Ohio State University

Christopher Robert Orton, M.S.

Graduate Program in Nuclear Engineering

The Ohio State University

2009

Dissertation Committee:

Richard N. Christensen, Advisor

Richard S. Denning

Xiaodong Sun

Abstract

The IAEA will require advanced technologies to effectively safeguard nuclear material at envisioned large scale nuclear reprocessing plants. This dissertation describes results from simulations and experiments designed to test the Multi-Isotope Process

(MIP) Monitor, a novel safeguards approach for process monitoring in reprocessing plants. The MIP Monitor combines the detection of intrinsic gamma ray signatures emitted from process solutions with multivariate analysis to detect off-normal conditions in process streams, nondestructively and in near-real time (NRT). Three different models were used to predict spent nuclear fuel composition, estimate chemical distribution during separation, and simulate spectra from a variety of gamma detectors in product and raffinate streams for processed fuel. This was done for fuel with various irradiation histories and under a variety of plant operating conditions. Experiments were performed to validate the results from the model. Three segments of commercial spent nuclear fuel with variations in burnup and cooling time were dissolved and subjected to a batch

PUREX method to separate the uranium and plutonium from fission and activation products. Gamma spectra were recorded by high purity germanium (HPGe) and cadmium zinc telluride (CZT) detectors. Hierarchal Cluster Analysis (HCA) and

Principal Component Analysis (PCA) were applied to spectra from both model and experiment to investigate spectral variations as a function of acid concentration, burnup

ii level and cooling time. Partial Least Squares was utilized to extract quantitative information about process variables, such as acid concentration or burnup. The MIP

Monitor was found to be sensitive to the induced variations of the process and was capable of extracting quantitative process information from the analyzed spectra.

iii

Dedication

Dedicated to my wife, children and God

May we, His children, use atomic power for peace and prosperity

Acknowledgments

There are many people that have helped in the completion of this dissertation to whom I am grateful. First, I want to thank my wife and children. Their love, support and patience have made this journey not only possible, but enjoyable. My successes are their successes and I am eternally grateful for their sacrifice.

I appreciate the support of my parents and my wife’s parents. Their support has been constant and reassuring. I am also grateful for the support of our grandparents and the example of my father and grandfathers. Their lives have provided both inspiration and background for this dissertation.

I wish to express my thanks and gratitude to my advisor, Dr. Richard Christensen.

Thank you for the opportunity, the support, the advice, the patience, the guidance and ultimately the encouragement for this dissertation as well as life. I am extremely grateful to have had the opportunity to be instructed by you and to work with you. I also wish to thank Pamela Christensen for her support to me and also my family. We are grateful for the meals and the fun times. We’ve been uplifted by your service.

I wish to express my thanks to my mentor and friend at Pacific Northwest

National Laboratory, Dr. Jon Schwantes. Thank you for the project, the funding, and the education. More importantly, thank you for pushing me to perform better and for keeping your sense of humor. I appreciate the guidance and unwavering support. It has

v been a privilege to complete my dissertation with you and I’m grateful for all you’ve done.

I wish to thank the additional members of my committee: Dr. Richard Denning,

Dr. Don Miller, Dr. Xiaodong Sun, and Steve Maheras. I thoroughly enjoyed my education at The Ohio State University due in large part to the efforts of these men.

Thank you for your courses, instruction, and wise counsel. Your efforts have been invaluable and impactful. Thank you.

I wish to thank my friends and associates at Pacific Northwest National

Laboratory, including Matt Douglas, Amanda Johnsen, Sam Bryan, Tatiana Levitskaia,

Chuck Soderquist, Lori Darnell, Leah Arrigo, Mark Englemann, Chris Aardahl (and everyone else in the Advance Radioanalytical Chemistry Group), Dan Couch, Andy

Prichard, Mark Shaver, Richard Pagh, Chris Gesh, Ann Doherty, Jeremy Kephart, Larry

Greenwood, Kathie Thomas, Walt Hensley, Shane Peper, Carrie Mathews, Brian W.

Smith, Khris Olsen and Mark Killinger. All of you have made contributions, whether large or small to the successful completion of this work and also to my career path, education, and achievements at PNNL. I am grateful for your help.

Finally I would like to thank the U.S. Department of Energy’s Fuel Cycle

Research and Development program (formerly GNEP and AFCI) for their financial support of this project. Additional support for the experimental demonstrations was provided by the U.S. DOE’s NA-243 office. I would also like to thank Cathy Dixon and the AFCI fellowship program for additional financial support for my education and research. Though this work has been funded by the U.S. DOE, any opinions, findings,

vi and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the Department of Energy.

vii

Vita

1998…………………………………………………………………Richland High School

2004-2007……………...National Security Intern, Pacific Northwest National Laboratory

2005………………………...... B.S. Chemical Engineering, Brigham Young University

2007……………………………….M.S. Nuclear Engineering, The Ohio State University

2008-Present….Post Masters Research Associate, Pacific Northwest National Laboratory

Publications

1. C.R. Orton, D.Y. Parkinson, P.D. Evans, N.L. Owen, “Fourier Transform Infrared Studies of Heterogeneity, Photodegradation, and Lignin/Hemicellulose Ratios Within Hardwoods and Softwoods,” Applied Spectroscopy, Vol. 58, Num. 11, p. 1265-1271 (2004).

2. L.E. Smith, J.M. Schwantes, M. Douglas, J.J. Ressler, C. Durst, C.R. Orton, R.N. Christensen, “Next-Generation Online MC&A Technologies for Reprocessing Plants,” Global 2007: Advanced Fuel Cycles and Systems, September 13, Boise, ID (2007).

3. J.M. Schwantes, M. Douglas, C.R. Orton, C. Fraga and R.N. Christensen, “Multi- Isotope Process (MIP) Monitor: a Near-Real-Time Monitor for Reprocessing Facilities,” ANS Transactions from Annual Meeting, Anaheim, CA (2008).

4. C.R. Orton, J.M. Schwantes, S.A Bryan, T.G. Levitskaia, D.C. Duckworth, M. Douglas, O.T. Farmer, III, C.G. Fraga, S.A. Lehn, M. Liezers, S.M. Peper, and R.N. Christensen, “Advanced Safeguards Technology Demonstration at Pacific Northwest National Laboratory,” Conference Proceedings of the Institute of Nuclear Materials Management 49th Annual Meeting, Nashville, TN (2008).

5. J.M. Schwantes, M. Douglas, S. Bonde, J.D. Briggs, O.T. Farmer, L.R. Greenwood, E.A. Lepel, C.R. Orton, J.F. Wacker, A.T. Luksic, “Nuclear Archeology in a Bottle:

viii Evidence of Pre-Trinity U.S. Weapons Activities from a Waste Burial Site,” Analytical Chemistry, Vol. 81, 4, p. 1297-1306 (2009).

6. M.A. Green, L.M. Arrigo, M. Liezers, C.R. Orton, M. Douglas, S.M. Peper, J.M. Schwantes, and D.C. Duckworth, “Electrochemically-Modulated Separations for Destructive and Nondestructive Analysis for Process Monitoring and Safeguards Measurements,” Institute of Nuclear Materials Management 50th Annual Meeting, Tucson, AZ (2009).

7. L.M. Arrigo, S.A. Bryan, R.N. Christensen, M. Douglas, D.C. Duckworth, C.G. Fraga, M. Liezers, C.R. Orton, S.M. Peper, and J.M. Schwantes, “Advanced Safeguards Technology Demonstration at Pacific Northwest National Laboratory,” Institute of Nuclear Materials Management 50th Annual Meeting, Tucson, AZ (2009).

8. J.M. Schwantes, C.R. Orton, C.G. Fraga, M. Douglas, and R.N. Christensen, “The Multi-Isotope Process (MIP) Monitor: a Near-Real-Time, Non-Destructive, Indicator of Spent Nuclear Fuel Reprocessing Conditions,” Institute of Nuclear Materials Management 50th Annual Meeting, Tucson, AZ (2009).

9. C.R. Orton, J.M. Schwantes, C.G. Fraga, M. Douglas, and R. Christensen, "Experimental Validation of the Multi-Isotope Process Monitor Concept," Conference Proceedings of GLOBAL 2009 "The Nuclear Fuel Cycle: Sustainable Options & Industrial Perspectives" September 6-11, 2009, Paris, France (2009).

Fields of Study

Major Field: Nuclear Engineering

Table of Contents

Abstract ...... ii Dedication ...... iv Acknowledgments ...... v Vita ...... viii List of Tables...... xii List of Figures ...... xiii Chapter 1: Introduction ...... 1 Chapter 2: Model and Experiment Methodology ...... 16 Introduction ...... 16 Methodology of the Simulations ...... 16 Methodology of the Experiment ...... 23 Multivariate Analysis Techniques ...... 27 Conclusion...... 37 Chapter 3: Testing the Multi-Isotope Process Monitor Through Simulation ...... 39 Introduction ...... 39 Model Results and Analysis ...... 39 Initial Proof of Concept – Analysis of Organic Extract ...... 41 MIP Monitor Applied to Aqueous Raffinate ...... 47 Detector Feasibility ...... 49 Cooling Time ...... 52 Extending Multivariate Analysis to Quantitative Predictions ...... 56 Summary and Conclusion ...... 57 Chapter 4: Experimental Test of the Multi-Isotope Process Monitor ...... 59 Introduction ...... 59 Results and Analysis ...... 60 Hierarchal Cluster Analysis ...... 61 Principal Component Analysis – Aqueous ...... 64 Principal Component Analysis – Organic ...... 70 Partial Least Squares ...... 77 Summary and Conclusion ...... 78 Chapter 5: Comparison of Model to Experiment ...... 80 Introduction ...... 80 Comparison of Model to Experiment ...... 80 Summary and Conclusion ...... 86 Chapter 6: Conclusion ...... 87 Summary of Conclusions ...... 87

x Model Analysis ...... 87 Experimental Analysis ...... 88 Comparison of Model and Experiment ...... 90 Need for Further Work ...... 90 References ...... 92

List of Tables

Table 1. Burnup characteristics of segments of spent BWR fuel (ATM-105) modeled in

ORIGEN-ARP...... 17

Table 2. List of elements modeled in AMUSE and isotopes modeled in Synth...... 20

Table 3. Burn up characteristics of ATM105 segments of spent BWR fuel (ATM-105) modeled in ORIGEN-ARP ...... 24

Table 4. Predictions of burnup level of fuels G1 and G2 based on simulated spectra of the organic extract from separations performed at various acid concentrations for fuels A and P...... 56

Table 5. Predictions of the extraction acid concentration of fuels ATM 105 A and ATM

109 based on spectra (HPGe) of the organic extract from separations performed at various acid concentrations...... 78

xii

List of Figures

Figure 1. Diagram of the Material Balance Areas for a generic PUREX reprocessing facility. 8 ...... 4

Figure 2. Visual representation of Principal Component Analysis, Q-residual and T 2.

(Adapted from Wise et. al. 18 ) ...... 33

Figure 3 Simulated spectra from organic extract solutions of liquid extraction separations performed at 2.25 M nitric acid concentrations of fuel with two different burnups. Ge detector, 1 hour count...... 40

Figure 4. A dendrogram from Hierarchical Cluster Analysis of the organic fraction, Ge spectra (normalized and mean centered) for samples with variable acid concentration

(0.1, 1.0, 2.13, 2.19, 2.25, 2.30, 2.36 M) and burnup (A = 16, G1 = 21.7, G2 = 23.4, and

P = 28.7 MWd/kgU)...... 41

Figure 5. Scores plot from the PCA of simulated HPGe spectra of the organic extract of separated fuels at four different burnups (A = 16, G1 = 21.7, G2 = 23.4, and P = 28.7

MWd/kgU)...... 43

Figure 6. Loadings of the Ge, organic fraction spectral data for each Principal

Component ...... 44

xiii Figure 7. Scores plot of HPGe, organic extract spectra samples wherein the PCA model was built on normal process samples and off-normal samples were subsequently projected onto model (A = 16, G1 = 21.7, G2 = 23.4, and P = 28.7 MWd/kgU)...... 45

Figure 8. Q-residuals and T 2 for the normal model with off-normal samples included (A

= 16, G1 = 21.7, G2 = 23.4, and P = 28.7 MWd/kgU)...... 46

Figure 9. Simulated spectra from the aqueous raffinate and organic extract of 28.7

MWd/kg U fuel. Coaxial Ge detector, 1 hour count...... 48

Figure 10. Scores plot from the PCA of simulated HPGe spectra of the aqueous raffinate of separated fuels at four different burnups (A = 16, G1 = 21.7, G2 = 23.4, and P = 28.7

MWd/kgU)...... 49

Figure 11. Simulated spectra from the organic extract of a 2.25 M separation of 28.7

MWd/kg U dissolved fuel. Variety of detectors, 1 hour count, normalized by area and artificially offset...... 50

Figure 12. Scores plots from the PCA of simulated NaI spectra of the aqueous raffinate of separated fuels at four different burnups (A = 16, G1 = 21.7, G2 = 23.4, and P = 28.7

MWd/kgU)...... 51

Figure 13. Q-residual and T 2 plot of the PCA for the NaI, normal aqueous raffinate samples (A = 16, G1 = 21.7, G2 = 23.4, and P = 28.7 MWd/kgU)...... 52

Figure 14. Scores plot of the Ge, organic samples including fuel decayed for 3, 15, and

26 years (PC1 vs PC2)...... 53

Figure 15. Scores plot of the Ge, organic samples including fuel (A = 16, G1 = 21.7, G2

= 23.4, and P = 28.7 MWd/kgU) decayed for 3, 15, and 26 years (PC2 vs PC3)...... 54

xiv Figure 16. Q-residuals plot of PCA model of Ge, organic normal spectra with off-normal spectra projected onto the model (A = 16, G1 = 21.7, G2 = 23.4, and P = 28.7

MWd/kgU)...... 55

Figure 17. Gamma spectra as taken by a CZT detector of the aqueous feed, raffinate and organic extract for the ATM 109 fuel separation...... 60

Figure 18. A dendrogram from Hierarchical Cluster Analysis of the organic fraction,

HPGe spectra (normalized and mean centered) for samples with variable acid concentration and burnup...... 61

Figure 19. A dendrogram from Hierarchical Cluster Analysis of the aqueous raffinate fraction, HPGe spectra (normalized and mean centered) for samples with variable acid concentration and burnup...... 62

Figure 20. A dendrogram from Hierarchical Cluster Analysis of the raffinate and aqueous dissolver (feed) solutions HPGe spectra (normalized and mean centered) for samples with variable acid concentration and burnup...... 63

Figure 21. Unsupervised scores plot from the PCA of HPGe spectra of the aqueous feed and raffinate of experimentally separated fuels performed at a variety of acid and TBP concentrations...... 65

Figure 22. Unsupervised scores plot from the PCA of HPGe spectra of the aqueous feed and raffinate of ATM 105 Section P performed at a variety of acid and TBP concentrations (Expanded from Figure 21)...... 66

Figure 23. Plot of the distance between the PC 1 & 2 scores of the ATM 109 Section P aqueous spectra collected on the same day and on the same HPGe detector versus the time of collection...... 67

xv Figure 24. A comparison of the peak channel of the 661.7 keV peak for the feed and raffinate solutions of the dissolved ATM 105 P fuel...... 68

Figure 25. Scores plot from the PCA of HPGe spectra of the organic extract of experimentally separated fuels performed at a variety of acid concentrations with 30%

TBP in dodecane...... 71

Figure 26. Scores plot from the PCA of CZT spectra of the organic extract of experimentally separated fuels performed at a variety of acid concentrations with 30%

TBP in dodecane...... 72

Figure 27. Scores plot from the PCA of replicate HPGe spectra of the organic extract of the ATM 109 segment at a variety of acid concentrations with 30% TBP in dodecane. .. 73

Figure 28. Q residual and T 2 plot from the supervised PCA of spectra from the organic fraction of ATM 105 Section A collected on a HPGe detector, 10 replicate spectra were collected for each sample listed in the legend...... 75

Figure 29. Q residual and T 2 plot from the supervised PCA of spectra from the organic fraction of ATM 109 collected on a HPGe detector, 10 replicate spectra were collected for each sample listed in the legend...... 76

Figure 30. Q residual and T 2 plot from the supervised PCA of both the ATM 105 A and

ATM 109 sample spectra collected on a HPGe detector, 20 spectra for each sample listed in the legend (10 per fuel type)...... 77

Figure 31. Modeled and experimental CZT spectrum of ~2.5 M raffinate of the 16

MWd/kgU fuel...... 81

Figure 32. Scores plot from PCA of simulated organic extract spectra using a CZT detector...... 84

xvi Figure 33. Scores plot from PCA of experimental organic extract spectra using a CZT detector...... 85

xvii

Chapter 1: Introduction

On December 8, 1953 Dwight D. Eisenhower delivered a speech to the United

Nations General assembly entitled “Atoms for Peace” in which he advocated the transition from the destructive to the peaceful use of atomic power. His vision included the establishment of “an international Atomic Energy Agency” to serve as a resource for the development of atomic energy for the peaceful pursuits and the benefit of mankind.

He suggested that the Agency serve as a bank of uranium and other fissionable materials with the responsibility of impounding, storing and protecting the material. 1 From the vision of Dwight D. Eisenhower, the International Atomic Energy Agency (IAEA) was born.

Although it does not currently serve as a bank of fissionable material as originally envisioned, 2 the Agency is tasked with verifying that nuclear material is not diverted from peaceful uses. States who have signed the Non Proliferation Treaty (NPT) or a similar treaty are required to enter into a safeguards agreement with the IAEA. With the exception of the five Nuclear Weapons States, the signatories of the NPT have agreed not to develop, manufacture or otherwise acquire nuclear weapons or other nuclear explosive devices and have accepted safeguards on all nuclear material used in peaceful nuclear activities. 3,4 The basic measure by which the IAEA verifies the fulfillment of these obligations is through nuclear material accountancy. 5 Hence the technical objective of

IAEA safeguards is to detect the diversion of significant quantities of nuclear material from peaceful activities for unknown purposes in a timely manner. 4,6

The IAEA safeguards source and special nuclear material (SNM) at a variety of facilities. As of 2001, safeguards activities were applied routinely at over 900 facilities in

71 countries, 7 including power reactors, research reactors, conversion plants, fuel fabrication plants, enrichment plants, storage facilities and reprocessing plants in addition to other facilities. Accountancy of source and SNM at the facilities occurs for one of two reasons: (1) the material is within a country that is under safeguards agreement and/or

(2) the material being safeguarded originates from a country that has signed a safeguards agreement.

Many of the facilities under IAEA observation are easily safeguarded by material control and accounting (MC&A) due to the itemized nature of the material. For example, a nuclear power plant receives, uses, and ships uniquely labeled fuel assemblies. The assemblies are not usually breached, nor are the fissionable materials sub divided or compiled. Hence, tracking the safeguarded material is as simple as tracking each assembly and ensuring its integrity. This is not the case with a bulk handling facility such as a spent nuclear fuel reprocessing plant. The most common reprocessing method in commercial use today is the PUREX liquid extraction process. By its very nature and purpose, a reprocessing facility separates, divides, mixes and redistributes safeguarded nuclear material. As a result, material control is no longer as simple as tracking fuel bundles. Instead, MC&A becomes a process of verifying flow sheet declarations, measuring the material held up throughout the plant, tracking its path through the plant, and ensuring it reaches it proper destination.

Traditional safeguards include designating material balance areas (MBA) throughout a facility. The IAEA inspectors at a bulk handling facility, such as a reprocessing plant, perform material accountancy by monitoring the amount of material held in an MBA. Several MBA’s are common within bulk handling facilities. Typical reprocessing facilities have one MBA at the front end that includes the receipt and storage area and the head-end fuel shearing and dissolution area. A second MBA typically encompasses the process area, and a third MBA is usually defined around the product storage area (Figure 1) at the tail end of the process. Material flow into and out of the MBA’s is monitored by the analysis of discrete samples taken from various locations within the MBA for destructive assay and measuring flows at “key measuring points,” or predetermined locations at the boundaries of the MBA’s. 8,9 By tracking SNM into and out of an MBA, the amount of material in the area at any given time can be estimated. In addition to continuous monitoring, full inventories are conducted periodically to verify the amount of material contained within an MBA.

Material Material Material Balance Balance Balance Area 1 Area 2 Area 3

U Product Uranium Uranium Processing Storage

Input Spent Accoun- Fuel Dissolver Separation tability Storage Tank Spent Fuel Spent

Pu Product Plutonium Plutonium Processing Storage

Hulls Fission Products and Waste

Figure 1. Diagram of the Material Balance Areas for a generic PUREX reprocessing facility.8

Typical MC&A measurements include concentrations, isotopic compositions, and volume, density and/or weight. These measurements are made by a variety of instruments and techniques including alpha spectrometry, calorimetry, chemical titration,

K-edge absorption densitometry, manometers and vibrating-tube densimeters, neutron techniques, spectrophotometry, uranium gravimetry, mass spectrometry, x-ray fluorescence, and gamma-ray spectrometry. 8 Using all of these techniques, IAEA inspectors are tasked with verifying the material inventory of a reprocessing plant within one “significant quantity” during a fixed time period. A “significant quantity” (SQ) is defined by the IAEA as 8 kg of plutonium, 8 kg of 233 U, 25 kg 235 U in high ( ≥20% 235 U) enriched uranium, 75 kg 235 U in low (<20% 235 U) enriched uranium, or 20 tons of thorium. The fixed time period within which the SQ must be verified depends on the

material type. For plutonium in irradiated fuel, this duration is 3 months, but for plutonium in the oxide or metallic forms, the MBA must be verified below an SQ over a period of 1 month. 10 The amount of plutonium in a reprocessing plant is checked using the techniques listed above. Each of these techniques has varying amounts of error associated with them and even the most precise of these techniques is insufficient to ensure no protracted diversion has taken place at large reprocessing facilities.

For example, given a large throughput reprocessing plant, such as the Rokkasho plant in Japan (~800 tons heavy metal/year), a spent fuel input of 0.8% plutonium by weight and an optimistic nominal measurement uncertainty of 0.2%, 12.8 kg of plutonium fall inside the annual error bounds. Given a more realistic 1% nominal measurement uncertainty, 64 kg of plutonium lie within the annual error bounds. 8 Even if the nominal measurement uncertainty reached the current target of 0.1%, at some point, using traditional MC&A methods, the facility becomes too big (~1000 tons heavy metal/year) to ensure that 1 SQ could not be diverted. The timeliness goal for verifying most of the plutonium at a reprocessing facility is one month (as opposed to the one year described in the example), thus easing the task of verifying an SQ. However, no additional verification is performed to ensure cumulative uncertainties don’t reach a SQ.

In addition, many of the more accurate classical techniques have a long wait time (3 – 4 months) due to transportation of samples off-site and the time it takes to do the analysis.

For these reasons many experts have called for improvements in safeguards methods to reconcile these and other difficulties and uncertainties. 11,12,13

Many of the improvements envisioned require drastic increases in manpower, technology or both. Shortening the material accountancy period or increasing the number

of material balance areas requires many more person-hours to accomplish a timely analysis. Near-real-time accounting and traditional process monitoring each require more measurements and inspectors. Currently, a major limitation of the IAEA is the additional manpower it can supply to inspect a reprocessing plant. In 1995 it was envisioned that the Rokkasho reprocessing plant in Japan would require 45 inspectors to maintain traditional safeguards. At that time the IAEA employed 200 total inspectors 8 to safeguard over 900 facilities of which only 5 were reprocessing plants. 7 That forecast predicted that a disproportional 22.5% of IAEA inspectors would be committed to safeguard one facility. Today, the Rokkasho reprocessing plant requires between 1000 and 1200 Person Days of Inspection (PDI) out of the approximately 10,000 PDIs deployed by the IAEA’s Safeguards Department worldwide. 14 While not as disproportionate as the forecast predicted, 10% of the inspection effort is currently spent on <0.12% of safeguarded facilities and <1.3% 15 of safeguarded plutonium (or < 0.5% of safeguarded plutonium not under a Voluntary Offer Agreement). To overcome this disproportional person-hour requirement for reprocessing facilities and to prepare for the inevitable increase in their number in the future, it has been suggested that unattended, tamper-proof instruments that continuously monitor nuclear material concentrations at or better than the uncertainty level of conventional, time-lapsed, destructive analysis be developed. These instruments might be able to provide the necessary enhancements in monitoring efficiency without a significant increase in inspectors and minimize or reverse the disproportional shift of the limited inspector resources to reprocessing facilities.

Additional process information above and beyond the typical concentration of SNM in-

stream would provide inspectors the means for verifying the process and preventing diversion of nuclear material. 11

Modern MC&A techniques strive to combine containment and surveillance

(C&S) techniques with the accuracy of direct material accounting. C&S techniques maintain continuity of knowledge of the location and movement of SNM (though surveillance or traditional process monitoring), thereby reducing the frequency of direct material accountancy measurements and verification. 3

The Multi-Isotope Process (MIP) monitor has been developed as a response to the call for effective and efficient monitoring. The MIP monitor is designed to detect changes in gamma-ray signals emitted from spent fuel process streams due to changing process conditions, non-destructively and in near-real time. Pattern recognition software is used to extract meaningful information from minor changes in the complex spectra. The MIP monitor concept has the potential to combine the advantages of periodic, direct measurements with the continuity of knowledge traditionally obtained through C&S techniques. This dissertation explores the viability and effectiveness of the MIP monitor in order to improve material control and accountability at reprocessing facilities.

A Review of Current Technologies

Destructive analysis (DA) is an essential part of material control and accountability for the IAEA. It is used to determine the composition and quantity of

SNM in the process streams and accountancy tanks. In the past, DA samples were sent off-site for analysis to IAEA labs or IAEA-approved partnering labs. This method resulted in a delay of several months, from the time samples are collected until results are

reported. The Rokkasho plant minimized this delay by placing an IAEA lab on site, 17 at the cost of many more person hours devoted to analysis. Rokkasho has also increased the number of MBA’s from three to five, partially due to the addition of a MOX conversion facility, but also due to an increase in the number of key measurement points. In addition, other strategic measurement points were identified for verification of flows within MBA’s. Both of these additions increase the number of accountancy books to track as well as the number of destructive samples to analyze, and consequently the required resources for implementation. In planning for larger facilities, simply adding measurement points and/or attempting to improve on existing DA accuracy does not represent an efficient or the most effective way to track nuclear material.

Process variables are monitored at reprocessing plants to indicate the condition of the system for the purposes of process control, and more recently, to improve the ability of the inspectors to verify MC&A. The Solution Measurement and Monitoring System

(SMMS) 16,17,18 at the Rokkasho plant measures the physical process conditions, such as tank level, temperature and density. By monitoring the physical conditions of the system, the IAEA can be alerted to any general anomalies in plant operations. While SMMS maintains a continuity of knowledge over the plutonium-bearing stream, it does not have the ability to measure the Pu or U directly or assure that the SNM is present in the process stream.

Another deployed process monitoring technique at Rokkasho is a method that measures the neutrons emitted from of curium that make up a portion of dissolved spent fuel. From these neutron measurements, the amount of plutonium is back calculated using industrial knowledge of the expected plutonium-curium ratio under normal

operating conditions. 19 Though this method is online, nondestructive, gathers results in near-real-time and provides a means for indirectly estimating plutonium concentration, it has serious drawbacks. The physical size of the neutron detectors limits the deployment of the technique to the initial dissolution and the final product streams. More importantly, the technique relies on the assumption that curium behaves similarly to plutonium during processing. However, curium is not always a good chemical analog for plutonium. The best example of these differences are illustrated in their redox chemistries, where curium exists in solution as a trivalent species, while plutonium, depending on the reduction potential in solution, may simultaneously exist in four different oxidation states (III-VI). 20,21 Curium will cease to be a good plutonium analog when the plutonium speciation is dominated by other than the (III) oxidation state.

Gamma spectroscopy is currently used by IAEA inspectors at reprocessing plants 5,22 to quantify enrichment, isotopic composition or the date of discharge of spent nuclear fuel. It has also been used in conjunction with passive neutron detection to determine the burn up of spent fuel assemblies, for both safeguards and operational purposes. 23,24 Additional details of such activities can be found in other references. 9,25,26,27,28,29,30 However, gamma spectroscopy has not been employed to monitor process stream conditions in reprocessing plants.

A multivariate approach known as principal component analysis (PCA) was tested as part of a system that monitored and measured volume, density and temperature of the solution in major vessels for safeguards purposes in reprocessing plants. 31 In addition,

PCA has been combined previously with gamma-ray spectrometry to monitor international border crossings by radiation portal monitors. 32 Mathematical algorithms

have also been used to enhance poor resolution gamma spectra, gathered using the

IAEA’s portable radionuclide identifications devices. 33 However, PCA has not been used previously on process stream spectra.

Background

Modern industrial reprocessing techniques, including the PUREX and UREX+ family of separations technologies, are based on solvent extraction between organic and aqueous phases. In these bi-phase systems, product (actinide) and contaminant (fission and activation products) elements are preferentially driven (thermodynamically) to opposite phases, with small amounts of each remaining in the other phase.34 The distribution of each element, between the organic and aqueous phases, is determined by major process variables such as acid concentration, organic ligand concentration, reduction potential, and temperature. Hence, for consistent performance of the separation process, the distribution of each element between the organic and aqueous phases should be relatively constant. During “normal” operations the pattern of elements distributing into the product and waste streams at each segment of the facility should be reproducible, resulting in a statistically significant signature of the nominal process conditions. Under

“abnormal” conditions, such as those expected under some protracted diversion scenarios, patterns of elements within the various streams would be expected to change measurably.

The MIP monitoring approach utilizes changes in the concentrations of gamma- emitting elements as evidence of changes to the process chemistry.35,36,37 It exploits a suite of gamma emitting isotopes to track multiple chemical species and behaviors

simultaneously, thus encompassing a large array of elements that are affected by chemical and physical changes. In-process surveillance by the MIP monitor is accomplished by coupling the gamma spectrometry of the streams with multivariate techniques, such as Principal Component Analysis (PCA). PCA is a chemometrics tool that finds combinations of variables (principal components or PCs) that best describe the common variance between differing datasets.38 Using multivariate analysis, such as

PCA, the MIP monitoring technique is then capable of automatically evaluating the patterns of the gamma-emitting contaminants for statistically relevant signs of potential changes to the process chemistry. The MIP monitor represents the first of its technology to combine gamma-ray spectroscopy with multivariate analysis for monitoring reprocessing operations.

Purpose and Structure

Monitoring online, nondestructively and in near real time are several approaches envisioned to enhance current material accountancy techniques. By focusing on the validation of declared process conditions (flow sheet verification), one can assure that the material remains within the intended path throughout the facility. With assurance of material control, the burden placed on accountancy measurements is lightened, reducing the requirement for accurate, but expensive, destructive analysis. The MIP monitor is designed primarily as a tool for ensuring the integrity of the process and guarding against material diversions by way of process manipulation. It has the potential to achieve this goal with minimal inspector interaction, non-destructively, online and in near real time.

The objective of this dissertation is to demonstrate, through modeling and experiment,

that the MIP monitor approach can detect and quantify process changes such as acid concentration and burnup variation.

This dissertation is broken up into five chapters. Chapter 1 consists of this manuscript, containing the motivation and background of the research. Chapter 2 contains a description of the methodologies used during the model and lab experiments.

The methodology includes descriptions of the various models used to simulate spent fuel, liquid extraction, and gamma spectra of various streams within a reprocessing facility.

Spent fuel characteristics are also described in addition to descriptions of the method of batch extraction used during simulations and experiments. The characteristics of the employed detectors are also listed. Finally, this chapter contains an explanation of multivariate analysis techniques, including Hierarchical Cluster Analysis (HCA), PCA and Partial Least Squares (PLS), which were used to analyze spectral features of the simulated and collected gamma-ray spectra.

Chapter 3 discusses the results and implications of the model simulations. The simulated gamma spectra were analyzed using HCA, PCA, and PLS. The HCA and PCA are explained and illustrate the viability of the MIP approach to monitor acid concentration and burnup, two major variables of spent nuclear fuel reprocessing.

Chapter 3 also contains a discussion on the ability of PLS to extract information about the burnup level of the fuel.

Chapter 4 contains the results of the experimental validation of the MIP concept.

Spent nuclear fuel, of different irradiation histories, was subjected to batch extractions representing the first stage of the PUREX liquid extraction process. Solvent extractions were conducted on the spent fuels at various acid and TBP concentrations. Samples of

the fuel solution, pre- and post-extraction, and samples of the extract were collected and gamma counted. The recorded spectra were analyzed using HCA and PCA for evidence of the changing process conditions and these results were reported in this chapter. The efforts to extract quantitative information of major process conditions from gamma spectra using PLS were also reported.

Chapter 5 contains a discussion comparing the simulated and experimental results. Similarities and differences are examined between the spectra and their subsequent PCA analysis.

The final portion of the dissertation, Chapter 6, includes a summary of the research.

1 Dwight D. Eisenhower. “Atoms for Peace,” An address given to the 470 th Plenary Meeting of the United Nations General Assembly, December 8 th , 1953, accessed at http://world-nuclear- university.org/html/atoms_for_peace/ (January 2009). 2 D. Fisher, History of the International Atomic Energy Agency: the first forty year , Vienna, Austria: The International Atomic Energy Agency (IAEA), 1997. 3 IAEA Department of Safeguards, “IAEA Safeguards: Staying Ahead of the Game,” Vienna, Austria: The International Atomic Energy Agency, July 2007. 4 International Atomic Energy Agency, “IAEA Safeguards Glossary, 2001 Edition,” Vienna, Austria: The International Atomic Energy Agency, 2001. 5 International Atomic Energy Agency, “Safeguards Techniques and Equipment, 2003 Edition,” Vienna, Austria: The International Atomic Energy Agency, 2003. 6 INFCIRC/153 (Corrected), The Structure and Content of Agreements between the Agency and States Required in Connection with the Treaty on the Non-Proliferation of Nuclear Weapons, 1972. 7 International Atomic Energy Agency, “IAEA Safeguards: Stemming the Spread of Nuclear Weapons,” Vienna, Austria: The International Atomic Energy Agency, accessed electronically at http://www.iaea.org/Publications/Factsheets/English/S1_Safeguards.pdf , March 2009. 8 U.S. Congress, Office of Technology Assessment, “Nuclear Safeguards and the International Atomic Energy Agency, Appendix A,” OTA-ISS-615. Washington, D.C.: U.S. Government Printing Office, June 1995. 9 O. Yamamura, R. Yamamoto, S. Nomura and Y. Fujii, “Development of safeguards and maintenance technology in Tokai Reprocessing Plant,” Progress in Nuclear Energy , Vol. 50, p. 666-673, 2008. 10 J. E. Doyle, Editor. Nuclear Safeguards, Security, and Nonproliferation . Butterworth-Heinemann (Elsevier), Burlington, MA. Chapter 5, “International Safeguards Inspection: An Inside Look at the Process” by B. Boyer and M. Schanfein, p. 91, 2008. 11 U.S. Congress, Office of Technology Assessment, “Nuclear Safeguards and the International Atomic Energy Agency,” OTA-ISS-615. Washington, D.C.: U.S. Government Printing Office, June 1995.

12 P.C. Durst et al., “Advanced Safeguards Approaches for New Reprocessing Facilities,” ASA-100 report, Pacific Northwest National Laboratory, PNNL-16674, June 2007. 13 International Atomic Energy Agency, “Report of the LASCAR Forum: Large Scale Reprocessing Plant Safeguards,” STI/PUB/922, IAEA, Vienna, Austria, 1992. 14 Personal correspondence with Shirley Johnson, 11 May 2009. 15 International Atomic Energy Agency, “IAEA Annual Report, Table A4. Approximate quantities of material subject to Agency safeguards at the end of 2007,” Accessed at http://www.iaea.org/OurWork/SV/Safeguards/sv.html on 13 May, 2009. 16 J. E. Doyle, Editor, Nuclear Safeguards, Security, and Nonproliferation , Butterworth-Heinemann (Elsevier), Burlington, MA. Chapter 9, “Case Study: Safeguards Implementation at the Rokkasho Reprocessing Plant” by S.E. Pickett, p. 165, 2008. 17 S.J. Johnson et al., “Development of the Safeguards Approach for the Rokkasho Reprocessing Plant,” International Atomonic Energy Agency, IAEA-SM-367/8/01, 2001. 18 M. Suzuki, M. Hori, S. Nagaoka, and T. Kimura, “Study on Loss Detection Algorithms Using Tank Monitoring Data,” Journal of Nuclear Science and Technology , Vol. 46, No. 2, p. 184-192, 2009. 19 P.M. Rinard and H.O. Menlove, “Application of Curium Measurements for Safeguarding At Reprocessing Plants,” LA-13134-MS, Los Alamos National Laboratory, 1996. 20 G.R. Choppin, B.E. Stout, “Plutonium – the element of surprise.” Chemistry in Britain , 12, 1126-1129, 1991. 21 L.R. Morss, N.M. Edelstein and J. Fuger, Editors, The Chemistry of the Actinide and Transactinide Elements , 3 rd Ed. Springer, Dordrecht, The Netherlands, 2006. 22 J. E. Doyle, Editor, Nuclear Safeguards, Security, and Nonproliferation , Butterworth-Heinemann (Elsevier), Burlington, MA., 2008. 23 C. Riffard, H. Toubon, S. Pelletier, M. Batifol, and J.M. Vidal, “MOX Fuel Characterization for Burnup Credit Application: Extension of Nondestructive Method Qualified for LEU Fuels,” Nuclear Technology , Vol. 154, p. 186-193, May 2006. 24 K. Oeda, H. Naito, M. Hirota, K. Natsume, and H. Kumanomido, “Calibration of Burnup Monitor Installed in Rokkasho Reprocessing Plant,” Journal of Nuclear Science and Technology , Vol. 37, No. 6, p. 543-547, June 2000. 25 S. Anilkumar, A.K. Deepa, K. Narayani, A.K. Rekha, P.V. Achuthan, G. Krishnamachari, and D.N. Sharma, “Estimation of 235 U concentration in some depleted uranium samples by high resolution gamma- ray spectrometry using 185 keV and 1001 keV gamma-energies of 235 U and 234m Pa,” Journal of Radioanalytical and Nuclear Chemistry , Vol. 274, No. 1, p. 161-166, 2007. 26 A. Morgenstern, C. Apostolidis, H. Ottmar and K. Mayer, “Analysis of 237 Np in spent fuel solutions,” Radiochim. Acta , Vol. 90, p. 389-393, 2002. 27 C.K. Mathews and P.R. Vasudeva Rao, “Radioactivity in Monitoring Materials Processing,” Journal of Radioanalytical and Nuclear Chemistry, Articles , Vol. 203, No. 2, p. 519-535, 1996. 28 H.T. Matsuda, B.F. de Araujo, and J.A. de Araujo, “Analytical Process Control of the Celeste R&D Installation of IPEN-CNEN/SP,” J. Radioanal. Nucl. Chem., Letters , Vol. 199, No. 6, p. 453-463, 1995. 29 F. Baumgärtner and D. Ertel, “The Modern PUREX Process and its Analytical Requirements,” Journal of Radioanalytical Chemistry , Vol. 58, p. 11-28, 1980. 30 S.G. Marathe, V.K. Rao, V.K. Bhargava, R.H. Iyer and M.V. Ramaniah, “A Gamma-Monitoring Assembly for the Analysis of Fission Products in Reprocessing Streams,” Nuclear Instruments and Methods , Vol. 127, p. 99-103, 1975. 31 M. Suzuki, M. Hori, R Asou & S. Usuda, “Numerical Consideration for Multiscale Statistical Process Control Method Applied to Nuclear Material Accountancy,” Journal of Nuclear Science and Technology , Vol. 43, No. 10, p. 1270-1279, 2006. 32 R.C. Runkle, M.F. Tardiff, K.K. Anderson, D.K. Carlson and L.E. Smith, “Analysis of Spectroscopic Radiation Portal Monitor Data Using Principal Component Analysis,” IEEE Transactions on Nuclear Science , Vol. 53, No. 3, p. 1418-1423, 2006. 33 L. Chen and Y. Wei, “Nuclide identification algorithm based on K-L transform and neural networks,” Nuclear Instruments and Methods in Physics Research A , Vol. 598, p. 450-453, 2009.

34 M. Benedict, T.M. Pigford, and H.W. Levi. Nuclear Chemical Engineering , 2nd Ed. McGraw-Hill, New York, NY,1981. 35 L.E. Smith, J.M. Schwantes, J.J. Ressler, M. Douglas, K.A. Anderson, C.G. Fraga, P.C. Durst, C.R. Orton, R.N. Christensen, “Next Generation On-line MC&A Technologies for Reprocessing Plants,” Proceedings of Global 2007 Conference on Future Nuclear Energy Systems, 2007 36 J.M. Schwantes, M. Douglas, C.R. Orton, C. Fraga and R.N. Christensen, “Multi-Isotope Process (MIP) Monitor: a Near-Real-Time Monitor for Reprocessing Facilities,” ANS Transactions from the Annual Meeting, Anaheim, CA, 2008. 37 C.R. Orton, J.M. Schwantes, S. Bryan, T. Levitskaia, D. Duckworth, M. Douglas, O.T. Farmer, C. Fraga, S. Lehn, M. Liezers, S. Peper, R.N. Christensen, “Advanced Safeguards Technology Demonstration at Pacific Northwest National Laboratory,” Proceedings of the 49 th Annual INMM Conference, Nashville, TN, 2008 38 E. Malinowski, Factor analysis in chemistry , John Wiley & Sons, New York, NY, 405p, 2008.

Chapter 2: Model and Experiment Methodology

Introduction

The Multi-Isotope Process Monitor concept was tested using both model simulations and experimental investigations. These approaches were designed to roughly reproduce operating conditions in a reprocessing plant. Both of these tests considered the processing of several spent nuclear fuels, under normal and off-normal conditions.

Gamma spectra from the process streams under the various conditions were simulated or collected, and identical multivariate techniques were applied to analyze these spectra.

Modeling, experimental and multivariate analysis approaches are summarized here.

Methodology of the Simulations

A group of three computer models were used to simulate gamma spectra from reprocessing streams. Spent boiling water reactor (BWR) fuel was modeled using

ORIGEN-ARP. 1 A common burn history at four different power levels was modeled to mimic actual fuel samples located at Pacific Northwest National Laboratory (which were used for experimental validation). The samples were from different locations on a single spent BWR fuel rod. The characterization report for the fuel segments indicates that each segment had a different average total burnup level due to their respective proximity to the top of the reactor core. 2 The details of the characterized segments can be found in Table

Section Identifier A G1 G2 P Distance from top (cm) 52 - 56 66 - 69 72 - 76 142 - 146 Approx. Burnup (MWd/kgU) 16.0 21.7 23.4 28.7

Table 1. Burnup characteristics of segments of spent BWR fuel (ATM-105) modeled in ORIGEN-ARP.

A detailed exposure history of the fuel assembly and a summary of chemical analyses of the fuel pellet compositions were available in the characterization history.

This information was used as input to ORIGEN-ARP in order to mimic the fuel composition and exposure as closely as possible. Ten elements were manually added to the default fuel composition file in ORIGEN-ARP to bring the model in closer agreement with the fuel characterization report. The fuel type selected in ORIGEN-ARP (which includes the cross section libraries) was a GE 7x7 BWR bundle with a 3% enrichment.

Due to their respective distances from the top of the reactor core, different moderator densities were used for each section. Section A, G1, G2, and P were assigned the moderator densities of 0.3, 0.5, 0.5, and 0.7 grams per cubic centimeter, respectively.

Section A had the minimum moderator density available and section P had the maximum moderator density. While moderator density is used to adjust the cross section libraries, the overall effect on the output was found to be minimal.

Calculations included 26 years of cooling after the final power down and removal of the fuel from the core, providing an estimate of the state of the fuels as of April, 2008.

The fuel composition after cooling times of 3 and 15 years were also calculated. As output, ORIGEN-ARP provides quantity, weight, and activity of the elements and

isotopes in the fuel. This output was used as a basis for predicting the elemental and isotopic composition in a simulated dissolved fuel solution.

The results from the ORIGEN-ARP simulation were used as input for Argonne’s

Model for Universal Solvent Extraction (AMUSE) 3,4,5 code to simulate solvent extraction. AMUSE, in its entirety, can calculate the steady state compositions for both the aqueous and organic phases at each contactor stage for various processes, as well as estimate the size and cost of the necessary equipment. For the purposes of this study, only a portion of the code (the Spreadsheet Algorithm for Speciation and Partitioning

Equilibria, SASPE) was used to calculate the batch extraction distribution coefficient for the primary uranium and plutonium extraction in a PUREX process. Regalbuto et. al. 3 explain that SASPE uses input compositions and the conditions of the aqueous and organic phases to calculate the distribution ratios. These calculations employ chemically correct models that use the thermodynamic activities of the major aqueous species. The approach of Bromley 6 is used to calculate these activities from aqueous-phase compositions. Solvent loading is also incorporated into the distribution coefficient. As input parameters, AMUSE requires relative concentrations of elements, acids and solvents in the contacting solutions as well as the temperature.

An aqueous feed solution, with a uranium concentration of 1.3 M typical of a

PUREX process, was modeled. 7 Concentrations of major radioactive fission products were included in, accordance to their respective ratios to the uranium in the dissolved fuel as simulated by ORIGEN-ARP. All fission and activation products were assumed to have dissolved completely into solution. The total element concentration of the selected nuclides were entered into AMUSE and the distribution coefficient was calculated based

on simulated contact with a 30% (V/V) Tri-Butyl Phosphate (TBP) in dodecane in a two- to-one volume ratio with the aqueous phase. The distribution coefficients were calculated at different nitric acid concentrations of the feed solution including 0.10, 1.00,

2.13, 2.19, 2.25, 2.31, and 2.36 M. Standard nitric acid concentration for PUREX was assumed to be 2.25 M, and simulated points at ± 2.5% (2.19 and 2.31 M) and ± 5% (2.13 and 2.36 M) of this nominal value were assumed to represent typical process variation.

The 0.10 and 1.00 M acid concentrations were chosen to represent a separation performed outside of the nominal conditions.

Some distribution coefficient calculation constants for elements described in the

ORIGEN simulation were poorly defined or missing altogether in AMUSE. For those elements included in AMUSE without the calculation constants necessary for the extraction simulation, a constant distribution coefficient of 0.001 was assumed. Due to this limitation, only the most active elements according to ORIGEN output were considered in the analysis. 8 A list of elements included in AMUSE is found in Table 2.

Element Symbol Isotopes Americium Am 241, 242, 242m, 243 Barium Ba 137m Carbon C 14 Cadmium Cd 113m Curium Cm 242, 243, 244 Cesium Cs 134, 135, 137 Europium Eu 152, 154, 155 Neptunium Np 239 Palladium Pd 107 Promethium Pm 146, 147 Plutonium Pu 238, 239, 240, 241, 242 Samarium Sm 151 Technetium Tc 99 Thorium Th 234 Uranium U 234, 237, 238 Yttrium Y 90 Zirconium Zr 93

Table 2. List of elements modeled in AMUSE and isotopes modeled in Synth.

Elements included in AMUSE but defined by the default distribution coefficient included barium (Ba), carbon (C), cadmium (Cd), cesium (Cs). In addition, the palladium (Pd) distribution coefficients as a function of acid concentration from literature 9 were used instead of the default constant. AMUSE did not have the capability to model distribution coefficients for Antimony (Sb), Tellurium (Te), Tin (Sn), Nickel

(Ni), and Niobium (Nb). These elements were left out of the model simulations.

The output from AMUSE consisted of the distribution coefficients for each chemical species. Each element was assumed to be restricted to one species, as that was the only option in AMUSE for a majority of the elements. Though AMUSE had the capability to model multiple oxidation states of plutonium, the input to AMUSE assumed that all of the plutonium existed in the IV oxidation state to simplify the analysis. This

assumption was adequate since the gamma rays from plutonium did not add appreciably to the overall spectra of the spent fuel.

The distribution coefficients generated by AMUSE were used to derive the element concentrations in both the organic extract and aqueous raffinate solutions.

Briefly, concentrations of the extracting element in the aqueous feed ( z) and the organic

(y) phase are related by the equations 7

Dz y = Eq. (1) 1+ DE F

where E/F is the volume ratio of the organic ( E) phase over the aqueous ( F) phase and D is the distribution coefficient, or

y D = Eq. (2) x

where (x) is the concentration of the element in the raffinate solution. Thus the fraction extracted from the feed solution ( ρ) can be described

Ey ρ = Eq. (3) Fz

The fraction of each element extracted combined with its relative isotopic abundances from ORIGEN-ARP output were used to distribute the nuclide activity between phases and propagate a list of nuclides and their activity. A list of these isotopes is found in

Table 2. The nuclide activity list was used as source characterization for input into a third computer program, PNNL’s Synth code, 10 to simulate gamma spectra of the feed,

raffinate and extract solutions. Synth is a 1-D radiation transport code designed to mimic the response of a selected detector type among several choices, including sodium iodide

(NaI), high purity germanium (HPGe) and cadmium zinc telluride (CZT) crystals.

Spectra are populated based on the nuclide source list and a library that attributes the gamma-rays and their branching ratios to each nuclide. The gamma spectrum from the feed, raffinate and organic phases were simulated using four different detector models, including HPGe, CZT, NaI, and lanthanum bromide (LaBr 3).

For all spectral simulations, a point source was assumed, separated from the detector by 5 cm of air. The energy calibration was set to 0.5 keV per channel and 4096 channels were used, resulting in full scale energy of 2048 keV. The live time was set to one hour. For the HPGe detector, a coaxial model of 50% relative efficiency was selected, and the default settings were used, including a resolution of 1.9 keV (0.14%) at

1332 keV. 10 A 0.1 cm 3 crystal was chosen for the CZT detector simulations using the default settings, including a resolution of 16 keV (1.2%) at 1332 keV. 10 A 2” x 2”

Thallium doped crystal was selected for the NaI detector and the default settings were

10 used, including 46 keV (7%) resolution at 662 keV. LaBr 3 detectors are not one of the detector types available as an option in Synth, however, the counting efficiency for LaBr 3

11 crystals are roughly similar to that of NaI. Given this similarity, LaBr 3 was simulated using the NaI settings with an altered energy resolution, adjusted from 7% to 3% FWHM at 662 keV. Simulating LaBr 3 in this manner did not take into account the inherent interference peaks at 788.7 keV and 1435.8 keV from 138 La contamination and other interferences from 227 Ac. 11 These effects can be significant for low activity samples, but in this case the large radiological activity of the simulated samples made the contribution

from the intrinsic radioactivity negligible. The LaBr 3 model also did not take into account the characteristic loss of resolution at low energies. 12

For each detector type, the gamma spectra for the raffinate and extract were simulated as a function of acid concentration and burnup level. Variations as a function of cooling time were also modeled exclusively in the organic extract solution. In total,

258 spectra were generated, 60 spectra per detector type (4 burnups – 1 feed, 7 raffinate,

7 extract spectra for each burnup), except for HPGe, for which the cooling time variation simulations were included, resulting in 78 spectra.

Methodology of the Experiment

Three segments of boiling water reactor (BWR) spent commercial nuclear fuel were used in the experimental analysis of the MIP monitor: ATM 105 segments A and P, and a segment of ATM 109. The dissolution and separation of the fuel was performed at the Shielded Analytical Laboratory hot cell facility at Pacific Northwest National

Laboratory. Two of the three segments originated from the same fuel rod (ATM 105) but at different vertical locations within the rod (segments A and P). As a result, these segments represented fuel of different burnup levels, but with identical time spent in the reactor and cooling. This rod was one of a group of 88 rods, known as ATM 105, that were Approved Testing Material (ATM) for the U.S. Department of Energy Office of

Civilian Radioactive Waste Management Program. The rod ADD2974 was characterized as part of the program, and segments A and P of this rod were used in the experiment. 2

ATM 105 samples were irradiated in the Cooper Nuclear Power Plant (Nebraska

Public Power District), and were fabricated by General Electric (GE) in 1972 as part of 7

x 7 assemblies. ATM 105 had an average burn up of 28 MWd/kgU, with the range of irradiation, from 15-35 MWd/kgU depending on the specific rod and the vertical location of the fuel within the rod. The burnup level over the length of ADD2974 was estimated using gamma analysis of the Cs-137 in the original fuel characterization report. 2 The estimated burnup level for sections A and P can be found in Table 1. All of the ATM 105 fuel was irradiated for various periods of time between July, 1974 and May, 1982. As of

May 2009, the segments had been cooling approximately 27 years. The fuel segments used in the demonstration had an initial enrichment of 2.94% 235 U.

Section Identifier ATM 105 A ATM 105 P Distance from top (cm) 52 - 56 142 - 146 Approx. Burnup (MWd/kgU) 15.5 - 17.5 29.0 - 31.0

Table 3. Burn up characteristics of ATM105 segments of spent BWR fuel (ATM-105) modeled in ORIGEN-ARP

The third segment of fuel used in the experiment was part of the ATM 109 fuel rod group, which was BWR fuel irradiated in the Quad Cities I reactor. The rods were fabricated by GE, post irradiation examinations were performed at GE’s Vallecitos

Nuclear Center, and then the fuel was sent to Argonne National Laboratory (ANL) followed by Pacific Northwest National Laboratory (PNNL) for use as an Approved

Testing Material. The segment of fuel dissolved for the demonstration had an initial enrichment of 3% and a burnup level of approximately 67 - 70 MWd/kgU. The fuel was irradiated from February 1979 until September 1987 at which point the rod was removed from its bundle and placed in a carrier assembly. The carrier assembly was placed back

into the reactor and the rod was irradiated from November 1989 until September

1992. 13,14 As of September 2009, the fuel had been cooling for approximately 17 years.

The dissolution and extraction of the fuel segments were performed in a hot cell by manipulator. The segments were removed from their cladding (12 – 15 grams) and dissolved in concentrated nitric acid. The undissolved fines were removed by centrifuging the solution. In the case of the ATM 105 A and 109 segments, 5 feed/dissolver samples were prepared for solvent extraction by a TBP and dodecane mixture. Each of these feed solutions had a uranium concentration of roughly 0.7 M and nitric acid concentrations of approximately 0.3, 1.3, 2.5, 3.8 and 5.1 M, respectively.

Following dissolution, each of the fuel segments underwent the first stage of a PUREX- type extraction in a batch wise fashion. These aqueous samples were contacted with an equal volume of 30% (V/V) Tri-Butyl Phosphate (TBP) in dodecane solution and then centrifuged to separate the phases. Portions of the feed, aqueous raffinate and organic extract were removed and stored for analysis, resulting in a total of 30 samples, including

10 aqueous samples and 5 organic samples for each fuel. Ten feed samples were also prepared with ATM 105 P fuel, including 6 samples at approximately 3.5 M, and one each at approximately 5 M, 6.5 M, 7.5 M, and 9 M. The replicate feed samples were prepared for TBP variations. All of the ATM 105 P feed solutions had a uranium concentration of approximately 0.9 M. Five of the 3.5 M feeds were contacted with 20,

25, 30, 40 and 50% (V/V) TBP in dodecane. The remaining feed solutions were contacted with 30% (V/V) TBP in dodecane. All were contacted in a one-to-one volume ratio. A total of 26 additional samples were removed from the feed, raffinate and extract

solutions of the ATM 109 fuel extraction to provide additional data points to quantify method variation.

The samples were removed from the hot cell for additional preparation and analysis. The solutions were subsampled to reduce the gross amount of radiation emitted by each sample. Portions (0.1 mL) of each sample were placed in 4 mL glass vials and diluted to 1 mL with 0.5 M nitric acid (aqueous samples) or 30% (V/V) TBP in dodecane

(organic samples) in order to provide sufficient volume to facilitate collimation during counting. In the case of the organic samples, the radiation levels of the undiluted samples were already low, but the subsampling during dilution further eliminated any aqueous carry over incompletely separated within the hot cell.

Gamma counting was performed in the Radiological Process Laboratory using both high purity germanium (HPGe) and cadmium zinc telluride (CZT) detectors. The aqueous samples were counted on a track HPGe counter, using distance to optimize count rate. Aqueous samples were counted between 10 and 50 minutes. The organic samples were counted on either a 70% or a 74% relative efficiency HPGe detector at close range for a total live count time of 2 hours. Calibrations, performed on the detector daily, showed minimal drift (approximately ± 0.1 keV).

The samples were also counted using a 0.125 cm 3 CZT detector. The count times for the aqueous and organic samples were 0.5 hours and 2 hours, respectively. Dead time for all counting was kept to below 10% using distance or, in the case of the high burnup

(ATM 109) feed samples, collimation.

All samples were counted at least once on both the HPGe and CZT detectors. The

5 organic extract samples of both the ATM 105 Section A and the ATM 109 fuel were

also counted 10 times each in a random order on the HPGe detector, resulting in 100 additional spectra.

Multivariate Analysis Techniques

Several multivariate analyses were performed on the spectral dataset for each detector, including Hierarchical Cluster Analysis (HCA), Principal Component Analysis

(PCA), and the Partial Least Squares (PLS) method. The PLS Toolbox 15 for Matlab 16 was used to perform all of the multivariate analyses mention above. All of the techniques required data preprocessing before they could be applied. In our case, preprocessing consisted of normalization to unit area and mean-centering. Normalizing by area removes the effect of source intensity on the spectra while maintaining the pattern.

Though intensity is a valid indicator of fuel history or process conditions, it is also easily influenced by source/detector geometry. While this effect can be easily managed during simulations, it may be difficult to control size and geometry of the samples to maintain precision between samples during an actual deployment. Normalization is one way to reduce these artificial pattern variances. Because normalization would be used in an actual deployment, it was used exclusively for the first step of the simulated data preprocessing.

Mean centering was done prior to analysis in order to remove all information from the spectra except the intra-spectral variances. In effect, mean centering adjusts the data such that the analysis takes place in the center of the variance of interest, removing the large but uninteresting distance from the zero line. Mean centering was essential for

PCA and did not adversely affect HCA and PLS, thus it was performed for all the spectra before analysis.

The analysis of the spectra included both supervised and unsupervised pattern recognition. Supervised pattern recognition takes into account sample information, such as which samples are normal and which are off-normal, in order to establish groups of interest and compare new samples to the groups. Unsupervised pattern recognition does not consider the sample’s origins, but instead allows the samples to group based solely on the similarities and differences found in the spectra. This allows the intrinsic organization to emerge resulting in increased insight into the reaction of the samples to the analysis technique.

HCA was used to search for intrinsic groupings within the dataset. HCA is an unsupervised technique used to identify whether clustering exists within a dataset. HCA uses a dendrogram to represent the interpoint distances between samples in row space. 17

The interpoint distance between clusters was calculated using Euclidean distance between nearest neighbors. The analysis and display technique of HCA allows for easy pattern recognition.

The spectral data from the simulation were analyzed using PCA to identify trends in the spectral patterns, and consider if PCA can group the spectra according to the physical differences of burnup level, acid concentration, and, to some extent, cooling time. A successful application of PCA would be expected to cluster the samples into distinct groups in a PCA scores plot based on their physical variations. PCA accomplishes clustering by finding combinations of variables called factors or principal components that describe major trends in the data. In effect, PCA takes a large dataset,

samples with hundreds or thousand of variables (like the channels in a spectrum) and looks for common variances that group the data and combines these variations into single principle components. PCA is ideally suited for compressing correlated or redundant data while retaining the essential information found in the data. This compression step not only simplifies the original data but it has the benefits of signal averaging and capturing information that depends on how the variables (e.g., energy channel intensities) change with respect to one another.

PCA modeling was accomplished by first performing singular value decomposition (SVD) on the data matrix X containing the normalized and mean-centered spectra for the dataset. The data matrix X consists of r rows (samples) and c columns

(variables, or channels in the case of gamma spectra). SVD is a specific method for eigen analysis, which decomposes X into the product of three matrices U, S, and V. S is a diagonal matrix containing the square roots of the eigenvalues. U and V are matrices that contain the corresponding eigenvectors, one spanning the row space and the other spanning the column space. This is shown by the following equation:

 T X = U S  V  Eq. (4) r×c r×q q×q c×q 

The mathematical rank of the data matrix is denoted by q and the superscript T denotes transpose. Typically, the product of the U and S matrices is called the scores matrix (T) and the V is called the loadings matrix (P). The scores and loadings matrices can be described using pairs of scores ( t) and loadings ( p) vectors as can be seen in the following equation:

T T T T T X = TP = t1p1 + t 2p 2 + ... + t np n + ... + t qp q Eq. (5)

The scores and loading pairs in Equation 5 represent the principal components and their

18 T total number equal the mathematical rank ( q) of the data matrix. The tipi pairs are ordered according to the amount of variance captured. The model is usually truncated after the number of meaningful scores and loadings ( n) is discovered and the small amount of remaining variance, which is usually noise, is compiled into a residuals matrix

(E), resulting in the following equation:

X = T (P )T + E Eq. (6) r×c r×n c×n r×c

While PCA can be used in an unsupervised setting to find natural groups within large datasets, it can also be used for anomaly detection or supervised learning. 17, 18, 19,20 When employed in anomaly detection, a PCA model is built on a representative spectra set of benign samples and the optimal number of principal components are selected. Then new samples are projected onto the benign PCA model to observe where they fall. The distance is then calculated between the new samples and the benign group to determine whether the new samples are a normal occurrence and fall inside the benign group or an anomaly.

Prior to projection, the test spectrum vector must be normalized and mean- centered. The mean values used for mean-centering came from those originally obtained from the X matrix. The projection of a test spectrum into the PCA space is given by

Equation 7.

s = y V Eq. (7) 1×n 1×c c×n

V (same as P ) is the truncated loading matrix from the PCA of X , y is a vector of data c×n c×n r×c representing a single test spectrum, and s is a vector of the scores whose numbers represent the location of the sample in PCA score space.

PCA models were created using spectra representing normal process conditions as the benign training set. Spectra representing off-normal conditions were then projected as new samples onto the model as test spectra to see how they compared to the normal spectra in PCA space. In addition, the Q-residuals and T 2 were calculated. These two metrics were used to determine whether the off-normal samples would be identified as an anomaly. The first metric, called Q-residuals, measures the residual between the original test spectrum ( y) and the spectrum reconstructed from the PCA model. The second metric, known as T 2, is a measure of the distance between the test sample and the multivariate mean of the benign training set in PCA space. In the case of the Q-residual, the spectrum is reconstructed by multiplying its score vector ( s) and the truncated loading matrix ( V ), shown in the following equation:

w = s (V )T Eq. (8) 1×c 1×n c×n

where w is the reconstructed spectrum. The residual vector e is calculated by subtracting w from y as shown in the following equation:

e = y − w Eq. (9) 1×c 1×c 1×c

The Q-residual is then calculated as the sum of the squares of the residual values shown in the following equation:

Q = e (e )T Eq. (10) 1×1 1×c 1×c

The Q-residual metric can be described as the distance between the projected sample and the model. 17 If the test spectrum contains variation dissimilar from that found in the model spectra (as we expect an anomalous spectrum would), the projection of the test spectrum onto the model will exclude the dissimilar variation. When the spectrum is reconstituted from the PCA model, it will be markedly different than the original spectrum, resulting in a relatively high Q-residual. This approach is especially useful when many principal components are used to describe the model (such that visualization of the grouping in PCA scores space is difficult or impossible) to identify anomalies.

In the case of T2, the distance between the multivariate mean of the benign training set and the test sample is significant because the data was mean centered so that the multivariate mean was located at the origin, where the score values equal zero. T 2 was calculated by taking the sum of the squares of the s vector values as shown in the following equations:

T 2 = s (s )T Eq. (11) 1×1 1×n 1×n

While not shown in Equation 11, the score values were normalized by multiplying each score value by the inverse of its associated eigenvalue. 18 The T 2 value does not take into

consideration how well the PCA model represents the test data (as the Q-residual does), but it identifies new spectra that have unusual variation captured by the benign PCA model. A graphical representation of a simplified PCA model, Q-residual and T 2 applied to a three dimensional variable space is shown in Figure 2.

PCA Model of the Data Sample with large 3 Q-residual - First Principal unusual variation Component outside the model

2 Sample with large T2 - unusual variations inside

Variable 3 Variable the model 1 T2 Confidence Interval

Second Principal 0 Component 3 4

2 3 2

1 1 Variable 1 Variable 2 0 0

Figure 2. Visual representation of Principal Component Analysis, Q-residual and T 2. (Adapted from Wise et. al. 18 )

In the illustrated example, the system of interest has data points spread over three variables. The PCA model only requires two PCs to capture the variance. The first PC is aligned to capture a majority of the variance, while the second PC is aligned to capture the most variance possible while remaining orthogonal to the first PC. In this simplified

figure, all of the variance of the training set is captured by the two principal components shown, leaving no residual. Any sample outside this plane (thus with a Q-residual value greater than zero) would not match the population described by the model, automatically identifying it as “off-normal.” In reality, the Q-residual value will be non zero for all samples used to make the model. A confidence level can be set for the normal distribution of Q-residuals which can then act as a threshold value for “off-normal” samples when the model is applied to new samples.

The distance of a sample from the center of the PCA model is captured by T 2.

Once again, those samples with a distance above a threshold for the “normal” magnitude can be identified as “off normal.” Both the Q-residual and T 2 were considered in the analysis of the spectral sets.

After analysis by PCA, PLS method was applied to the organic extract spectra to predict the burnup level of the fuel represented by the spectra. PLS is a multivariate calibration method that utilizes an approach analogous to PCA. The general version of

PLS described and used in this manuscript is PLS1. 18 Using PLS1, it is necessary to develop separate models for each property of interest (i.e. burnup level or acid concentrations) to extract quantitative information. PLS is best understood by first examining Principal Component Regression (PCR) and then identifying how PLS differs from PCR. Summaries of the PCR and PLS explanations given here are based on descriptions from the PLS Toolbox manual 18 and the PLS tutorial by Geladi and

Kowalski. 21 PCR and PLS are widely used multivariate calibration methods and are both based on the inverse calibration model. 17, 22 They assume that a regression vector b can

be used to determine the quantity of interest, c, from a vector of measured variables r. In our case r is the gamma spectrum. The formula for such an operation is shown below:

rb = c Eq. (12)

Using a calibration matrix R and a vector c containing the respective values of the quantity of interest for each spectrum in R, the regression vector b is calculated. The calibration matrix in the model study was a sub portion of the spectra set ( X), representing the high (28.7 MWd/kgU) and low (16 MWd/kgU) burnup level sample sets at all acid concentrations. For the experimental study, it was 45 of the 50 replicate spectra from the organic solutions of both the ATM 105 Section A and ATM 109 segments. This left one spectrum representing the organic solution of the extractions performed at each of the acid concentrations considered. The vector b is estimated from the following equation:

b = R+c Eq. (13)

where R+ is the pseudo-inverse of R. Both PCR and PLS estimate R+ by replacing the original variables in the calibration matrix R with linear combinations of the variables, called factors. PCR and PLS differ in the manner by which they calculate the factors.

PCR uses Principal Components for these factors, using the method described in

Equations 4 – 6. PCR then estimates R+ as

R+ = P(TTT)-1TT Eq. (14)

and b is calculated according to Equation 13. Using Equation 12, c can then be calculated by using b in conjunction with a test sample being projected ( r) onto the PCA model built on R.

PLS differs from PCR in that it uses R and c to find factors called latent variables

(LVs) that capture variance in R but also achieve correlation with c. Essentially, PLS attempts to maximize covariance between R and c. There are numerous ways to calculate PLS parameters. The method used in this paper is known as non-iterative partial least squares or NIPALS. 21 NIPALS uses both R and c in the calculation of scores T and loadings P (similar to those from PCA or PCR) and an additional set of vectors known as weights, W. W is needed to keep the vectors in T orthogonal. NIPALS estimates R+ as

−1 −1 R + = W(P T W) (TTT) TT Eq. (15)

T and P are not the same as those calculated for PCR by the PCA of R. However, they can be viewed as PCA scores and loadings that have been mathematically rotated to be more effective for predicting c. The remaining steps for determining b and predicting the quantity of interest of an unknown r are similar to those used in PCR. This was done to predict the burnup level of fuel samples G1 and G2 using their simulated organic extract spectra at all acid concentrations. It was also used to predict the acid concentration of the extraction for the experimentally derived organic sample spectra left out of the calibration matrix.

Conclusion

Simulated spectra were prepared by way of model to represent the feed, raffinate and extract solutions of a simulated PUREX extraction of spent nuclear fuel. Seven extraction acid concentrations representing both normal and off-normal process conditions were modeled. This included fuel with four burnup levels and three cooling times. Spectra were simulated based on the response of four detectors, including HPGe,

CZT, LaBr 3, and NaI.

Spectra were collected from batch extraction feed, raffinate and extract streams from a batch PUREX extraction of 3 segments of spent nuclear fuel. Each fuel segment was extracted at 5 different acid concentrations, and the ATM 105 Section P fuel segment was also extracted at 5 TBP concentrations. The fuel segments spanned various burnup levels and cooling times. Spectra were collected on both HPGe and CZT detectors. Ten replicate spectra were collected for each of the organic fractions of the ATM 105A and

ATM 109 fuel segments.

The multivariate analysis techniques used to analyze both the simulated and experimental spectra were described in detail, including HCA, PCA, and PLS. The results of these analyses will be presented in Chapters 3 and 4.

1 A.G. Croff, “ORIGEN2: A versatile computer code for calculating the nuclide compositions and characteristics of nuclear materials,”, Nucl. Technol. , 62, 3, 1983. 2 Guenther, R. J. et al., “Characterization of Spent Fuel Approved Testing Material – ATM-105,” Pacific Northwest Laboratory, PNL-5109-105, June 1989, accessed at http://www.lsnnet.gov/docview.aspx?mode=1&lsn=DN2002138631&ic=1&im=0&sc=12&sm=0 (October 2008) 3 M. C. Regalbuto, J. M. Copple, R. Leonard, C. Pereira, G. F. Vandegrift, “Solvent Extraction Process Development for Partitioning and Transmutation of Spent Fuel,” Proceeding of the 8 th Information Exchange Meeting on Actinide and Fission Product Partitioning and Transmutation:, Las Vegas, Nevada, United States, November 9-1, pp. 373-385, 2004; © Organization for Economic Co-operation and Development, Nuclear Energy Agency (OECD-NEA): Paris, France, 2005, accessed at http://www.nea.fr/html/pt/docs/iem/lasvegas04/posterI.html (October 2009).

4 G.F. Vandegrift, D.B. Chamberlain, C. Conner, J.M. Copple, J.A. Dow, L. Everson, J.C. Hutter, R.A. Leonard, L. Nunez, M.C. Regalbuto, J. Sedlet, B. Srinivasan, S. Weber, and D.G. Wygmans, "Development and Demonstration of the TRUEX Solvent Extraction Process," Proceedings of the Symposium on Waste management, Tucson, AZ, pp. 1045-1050, Feb. 28-Mar. 4, 1993. 5 R.A. Leonard and M.C. Regalbuto, "A Spreadsheet Algorithm for Stagewise Solvent-Extraction," Solvent Extraction and Ion Exchange , Vol. 12, Issue 5, pp 909 – 930, 1994. 6 L.A. Bromley, “Thermodynamic Properties of Stroping Electrolytes in Aqueous Solutions,” AIChE J. , 19, 313, 1973. 7 M. Benedict, T.M. Pigford, and H.W. Levi, Nuclear Chemical Engineering , 2nd Ed. McGraw-Hill, New York, NY, 1981. 8 R.W. Perkins and U.P. Jenquin, “Fission and Activation Products in Nuclear Reactor Fuels and Nuclear Explosion Debris,” PNNL-11554, Pacific Northwest National Laboratory, 1997. 9 T. Ishimori and K. Watanabe, “Inorganic Extraction Studies on the System of Tri-n-butyl Phosphate- Nitric Acid,” Bulletin of the Chemical Society of Japan , Vol. 33, No. 10, 1960. 10 W.K. Hensley, A.D. McKinnon, H.S. Miley, M.E. Panisko, and R.M. Savard, “SYNTH: a spectrum synthesizer,” J. Radioanal. Nucl. Chem ., 193, 229, 2005. 11 B.D. Milbrath, B.J. Choate, J.E. Fast, W.K. Hensley, R.T. Kouzes, J.E. Schweppe, “Comparison of LaBr 3:Ce and NaI(Tl) scintillators for radio-isotope identification devices,” Nuclear Instruments and methods in Phisyics Research A , 572, p. 774-784, 2007. 12 J.M. Schwantes et al., “Medium-resolution autonomous in situ gamma detection system for marine and coastal waters,” J. Radioanal. Nucl. Chem , in press/available online. 13 S. Vaidyanathan, R.D. Reager, R.W. Warner et al., “High Burnup BWR Fuel Pellet Performance,” American Nuclear Society, Proceedings of the International Topical Meeting on Light Water Reactor Fuel Performance, Portland, OR, p. 471, March 2-6, 1997. 14 S.F. Wolf, D.L. Bowers, J.C. Cunnane, “Analysis of high burnup spent nuclear fuel by ICP-MS,” Journal of Radioanalytical and Nuclear Chemistry , Vol. 263, No. 3, p. 581-586, 2005. 15 PLS_Toolbax Version 5.0 for use with MATLAB, Eigenvector Research, Inc, Wenatchee, WA., 2008. 16 Matlab, Ver. 7.8.0.347 (R2009a), The MathWorks, Inc, 2009. 17 K.R. Beebe, R.J. Pell, M.B. Seasholtz, Chemometrics: A Practical Guide , John Wiley & Sons, Inc., New York, NY, 1998. 18 B.M. Wise et al., PLS_Toolbax Version 4.0 for use with MATLAB, Manual , Eigenvector Research, Inc, Wenatchee, WA., 2006. 19 S. Dragovic & A. Onjia, “Classification of soil samples according to their geographic origin using gamma-ray spectrometry and principle component analysis,” Journal of Environmental Radioactivity , 89, p. 150-158, 2006. 20 P.K. Hopke. “The evolution of chemometrics.” Analytica Chimica Acta , 500, p. 365-377, 2003. 21 P. Geladi, B. R. Kowalski, Anal. Chim. Acta, 185, p. 1-17, 1986. 22 H. Martens and T. Naes, Multivariate Calibration , John Wiley & Sons, Ltd., Chichester, 1989.

Chapter 3: Testing the Multi-Isotope Process Monitor Through Simulation

Introduction

This chapter describes results from model simulations designed to test the Multi-

Isotope Process (MIP) Monitor, a novel safeguards approach for process monitoring in reprocessing plants introduced previously in Chapter 1. Three different models were used to predict spent nuclear fuel composition, estimate chemical distribution during separation, and simulate spectra from a variety of gamma detectors in product and raffinate streams for processed fuel. This was done for fuel with various irradiation histories and under a variety of plant operating conditions. Principal Component

Analysis (PCA) was applied to gamma spectra to investigate pattern variations as a function of acid concentration, burnup and cooling time. Hierarchical Cluster Analysis

(HCA) and Partial Least Squares (PLS) were also used in the analysis. The method of simulation and analysis are described in Chapter 2. Here, this chapter summarizes the results of a series of simulations designed to test the MIP monitor concept and explore its viability as a useful technology for MC&A.

Model Results and Analysis

An example of the spectra output by Synth is illustrated in Figure 3. Two spectra of the organic extract fraction at different burnup levels recorded on an HPGe detector are compared. Other than being slightly offset, no obvious difference in peak ratio or

pattern are observed. The spectra are offset on the y-axis largely due to the difference in burnup levels. Higher burnup spent fuels contain more fission products that result in more activity. Hence the spectra for 28.7 MWd/kg U fuel in Figure 3 has higher total counts than that of the spectra representing the 16 MWd/kg U fuel. When these spectra are normalized, however, the spectra are difficult to distinguish by visual inspection. This observation holds for spectra within both the raffinate and extract spectra sets, respectively.

No Processing

16 MWd/kgU 28.7 MWd/kg U Normalized by Area Log Scale Counts Artificially Offset Offset Artificially for Clarity

Organic Extract Fraction Separation Performed at 2.25 M Acid

0 200 400 600 800 1000 1200 1400 1600 keV

Process changes were expected to be most easily discerned within the product organic fraction stream. This was due to the absence of fission products (especially 137 Cs) that 40

produce a relatively large Compton region masking many lower energy gamma lines within the spectra from the aqueous phase samples. The organic extract samples contained much lower amounts of the most radioactive fission products, allowing analysis of more of the lower energy gamma lines.

Initial Proof of Concept – Analysis of Organic Extract

HCA was performed on the simulated organic extract data from an HPGe detector to test the MIP monitor concept. The intrinsic groupings of the samples from this analysis are illustrated in the dendrogram in Figure 4.

P 2.36 M P 2.30 M P 2.25 M 25 P 2.19 M P 2.13 M G2 2.36 M G2 2.30 M G2 2.25 M 20 G2 2.19 M G2 2.13 M G1 2.36 M G1 2.30 M G1 2.25 M 15 G1 2.19 M G1 2.13 M A 2.36 M A 2.30 M A 2.25 M 10 A 2.19 M A 2.13 M P 1 M G2 1 M G1 1 M 5 A 1 M P 0.1 M G2 0.1 M G1 0.1 M A 0.1 M 0 -0.005 0 0.005 0.01 0.015 0.02 0.025 Distance to Nearest Neighbor

Figure 4. A dendrogram from Hierarchical Cluster Analysis of the organic fraction, Ge spectra (normalized and mean centered) for samples with variable acid concentration (0.1, 1.0, 2.13, 2.19, 2.25, 2.30, 2.36 M) and burnup (A = 16, G1 = 21.7, G2 = 23.4, and P = 28.7 MWd/kgU).

The results from HCA show a tight grouping of the data within the “normal” process acid concentrations (2.13 M – 2.36 M) at each burnup level. Each of these “normal” groups are closest to the other “normal” groups of differing burnup levels, with G1 being closest to G2, followed by A and then P. The 0.1 M samples of all burnup levels are closest to each other, and this trend is also followed by the 1 M samples. The 1 M group is closest to the “normal” groupings of all burnup levels followed by the 0.1 M group. These results indicate that the samples preferentially group according to the “normal” acid concentrations, and should be easily separated from spectra representing “off-normal” acid concentration conditions regardless of burnup level.

Unsupervised PCA of simulated HPGe spectra of the organic extract grouped in a similar fashion to that of the results from cluster analysis. For PCA, all of the data (both normal and off-normal) were included in the analysis in order to observe the natural grouping of the data. Essentially all of the variance between the datasets was captured using two principal components. Unsupervised PCA resulted in easily discernable groupings that reflected both acid concentration and burnup variations (Figure 5).

A 2.19 M A 2.30 M A 2.25 M A 2.13 M A 2.36 M 0.01 A 1 M A 0.1 M A 2.25 M

0.005

G1 2.25 M

G1 1 M 0 G1 0.1 M

G2 2.25 M G2 1 M G2 0.1 M -0.005 P 2.25 M Scores PCon 2 (7.30%)

-0.01 P 1 M

P 0.1 M -0.015 -0.02 -0.01 0 0.01 0.02 0.03 0.04 0.05 0.06 Scores on PC 1 (92.70%)

Figure 5. Scores plot from the PCA of simulated HPGe spectra of the organic extract of separated fuels at four different burnups (A = 16, G1 = 21.7, G2 = 23.4, and P = 28.7 MWd/kgU).

The expanded view around the acid concentration of 2.25 M, for the 16 MWd/kgU sample group reveals that the order of the samples according to acid concentration is retained, even at small increments of variation (0.06 M). A loadings plot of the data, as shown in Figure 6, provides a graphical illustration of the relative contributions from the individual gamma energy lines of the spectra to the principal components in the PCA model.

0.5

-0.5

Loadings PC 1 on (92.70%) Loadings 500 1000 1500 2000 2500 3000 3500 4000 Variable Variables/Loadings Plot for Ge modeled organic extract from ATM 105 1

0.5

-0.5 Loadings on PC on (7.30%) Loadings 2 500 1000 1500 2000 2500 3000 3500 4000 Variable

Figure 6. Loadings of the Ge, organic fraction spectral data for each Principal Component

Here, each channel represents 0.5 keV in gamma ray energy, or “Variable” as labeled on the X-axis in Figure 6. From this Figure, principal component (PC) 1 contains a majority of the captured variance from the 60, 123, and 661 keV areas of the spectra, although several other areas contribute minor amounts of variance to this PC as well.

PC2 also captures variance in the 60, 123, 661 keV regions, with additional significant contributions coming from the 248, 723 and 1275 keV regions of the spectra. Many other regions also contribute to PC2. Though some regions of the spectra contribute to both PCs, it is important to note that the variance captured by PC2 is orthogonal to that captured by PC1.

To further test the ability of PCA to distinguish between normal and off-normal process conditions, a PCA model was created using synthetic HPGe spectra from the

organic extract for the dataset defined here as representing normal operating conditions

(2.13 – 2.36 M spectra for all burnups). For this model, two PCs were sufficient to capture 99.9% of variance between the data. Spectra from the 0.1 M and 1 M separations

(representing off-normal conditions) were then projected onto the PCA model. A scores plot was created (Figure 7) around the 99% confidence interval of the normal dataset.

This confidence limit was calculated using the Students t-distribution which assumes that the scores were distributed normally. Though the dataset is artificial and fixed (i.e. not a normal distribution), it is meant to represent a normal range, thus the confidence level serves as a general guide for anomaly detection.

0.01 G1 2.25 M A 2.25 M 0 P 2.25 M G2 2.25 M -0.01

-0.02 P 1 M G1 1 M A 1 M -0.03 G2 1 M

-0.04 Scores PC on 2 (8.41%) -0.05 Normal P 0.1 M Off Normal A 0.1 M -0.06 99% Confidence Level G2 0.1 M G1 0.1 M

-0.07 -0.02 -0.01 0 0.01 0.02 0.03 0.04 Scores on PC 1 (91.58%)

Figure 7. Scores plot of HPGe, organic extract spectra samples wherein the PCA model was built on normal process samples and off-normal samples were subsequently projected onto model (A = 16, G1 = 21.7, G2 = 23.4, and P = 28.7 MWd/kgU).

The figure illustrates that the 0.1 M and 1 M samples fell well outside the confidence interval for the normal acid concentration data, clearly and correctly identifying them as process anomalies, regardless of their burnup level.

The Q-residuals and T 2 for data shown in Figure 7 are plotted in Figure 8.

-6 x 10 6 A 0.1 M

3 P 0.1 M G1 0.1 M 4 2 G2 0.1 M

1 3 0

0 2 4 6 8 10 12

Q Q Residuals (0.01%) 2 Normal Group Normal A 1 M Off Normal 1 P 1 M G1 1 M 99% Confidence Level G2 1 M 99% Confidence Level

0 0 200 400 600 800 1000 1200 T^2 (99.99%)

Figure 8. Q-residuals and T 2 for the normal model with off-normal samples included (A = 16, G1 = 21.7, G2 = 23.4, and P = 28.7 MWd/kgU).

Here, off-normal samples are well out side the confidence intervals for both the Q- residuals and the T 2, demonstrating again the ability of the MIP monitor concept to identify anomalies in process conditions. The normal samples fall within the 99% confidence intervals, as can be seen in the expanded view in Figure 8. PCA analysis on simulated spectra demonstrates that, under ideal conditions, the MIP monitor concept is feasible when applied to the organic extract streams of a reprocessing facility.

Multivariate techniques were capable of identifying changes brought about by minor changes in major process variables within gamma spectra of these streams.

MIP Monitor Applied to Aqueous Raffinate

The MIP monitoring approach was also applied to simulated aqueous raffinate streams. Spectra from aqueous raffinate were quite different from that of organic extract, as can be seen in Figure 9. The organic phase contains roughly a factor of 100 less activity than the aqueous stream. The large Compton effect seen in the raffinate phase is due to the 137 Cs, which is preferentially stable in that phase. As a result, the Compton effect seen in the spectra of the organic phase is approximately two orders of magnitude lower than that observed for the raffinate. The large Compton effect in the aqueous phase dominates the spectrum < 400 keV.

1.E+11

1.E+10

Raffinate (Aqueous) 1.E+09 Extract (Organic)

1.E+08

1.E+07

1.E+06 Log Counts Scale 1.E+05

1.E+04 28.7 MWd/kg U

1.E+03

1.E+02 0 200 400 600 800 1000 1200 1400 1600 1800 2000 keV

Figure 9. Simulated spectra from the aqueous raffinate and organic extract of 28.7 MWd/kg U fuel. Coaxial Ge detector, 1 hour count.

Unsupervised PCA analysis was performed on simulated HPGe spectra of the aqueous raffinate in a similar manner to that applied to spectra from the organic extract. Two

PC’s were sufficient to describe 99.9% of the total variance. The scores plot for this analysis is shown in Figure 10.

-4 x 10 2

A 2.25 M 1.5 A 1 M

1 A 0.1 M

G1 2.25 M 0.5 G1 1 M G1 0.1 M 0 G2 2.25 M G2 0.1 M G2 1 M -0.5 Scores PC on 2 (7.41%) -1 P 0.1 M -1.5 P 1 M P 2.25 M -2 -12 -10 -8 -6 -4 -2 0 2 4 Scores on PC 1 (92.58%) -4 x 10

Figure 10. Scores plot from the PCA of simulated HPGe spectra of the aqueous raffinate of separated fuels at four different burnups (A = 16, G1 = 21.7, G2 = 23.4, and P = 28.7 MWd/kgU).

From Figure 10, datasets defined as normal (2.13 – 2.36 M) are distinctly different from the simulated spectra for raffinate solutions separated at 0.1 and 1 M acid concentrations.

When projected onto a PCA model built on the datasets representing the normal acid concentrations, datasets representing off-normal acid concentrations fell outside of the

99% confidence interval for the scores, Q-residuals and the T 2 tests (not shown).

Detector Feasibility

One would expect high resolution detectors to outperform those of inferior resolution for the pattern recognition requirements of the MIP monitor. However, applications of the MIP monitor in actual facilities will require robust and compact detectors at a cost of less resolution. A balance between these requirements will be 49

necessary for MIP monitor deployment. To explore the performance of available detector options, multivariate analysis of spectra simulated from four different detector types with various resolutions was conducted. Simulated spectra show obvious resolution trends

(Figure 11) for each of the detector types studied: Ge (0.14%), CZT (1.2%), LaBr 3 (3%), and NaI (7%).

NaI LaBr3 CZT Ge ArtificiallyClarity Offset for Log NormalizedScaleCounts

28.7 MWd/kg U

0 200 400 600 800 1000 1200 1400 1600 1800 2000 keV

Figure 11. Simulated spectra from the organic extract of a 2.25 M separation of 28.7 MWd/kg U dissolved fuel. Variety of detectors, 1 hour count, normalized by area and artificially offset.

PCA was performed on the spectra of raffinate using a simulated NaI detector. Use of the low resolution NaI detector on the highly active aqueous phase, make this scenario the least likely to be sensitive to process changes. When unsupervised PCA was applied

to the dataset, 3 PCs were needed to describe 99.75% of the variance within the dataset.

The scores plots for these samples are shown in Figure 12.

-4 -4 x 10 x 10 4 G2 2.25 M G1 2.25 M A 2.25 M A 2.25 M G1 2.25 M 2 P 2.25 M G2 2.25 M P 2.25 M G1 1 M 0 G1 1 M A 1 M A 1 M G2 1 M G2 1 M P 1 M -2 P 1 M

-4

-6

Scores on 1 PC (88.60%) G1 0.1 M A 0.1 M G1 0.1 M -8 G2 0.1 M A 0.1 M G2 0.1 M

P 0.1 M P 0.1 M

Scores on PC 1 (88.60%) on PC Scores -10

-12 -2.5 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2-8 -6 -4 -2 0 2 4 6 8 10 12 -4 Scores on PC 3 (1.02%) -5 Scores on PC 2 (10.13%) x 10 Scores on PC 2 (10.13%) x 10 Scores on PC 3 (1.02%)

Figure 12. Scores plots from the PCA of simulated NaI spectra of the aqueous raffinate of separated fuels at four different burnups (A = 16, G1 = 21.7, G2 = 23.4, and P = 28.7 MWd/kgU).

The two dimensional representations shown in this figure illustrate that the samples group according to both acid concentration and burnup level. There was some overlap between the normal grouping for the 16 MWd/kgU and 21.7 MWd/kgU fuels, due to the loss of spectral features with lower detector resolution. Though PC’s 1 and 2 captured 98.7% of the variance, it was found that three PC’s were required to provide adequate separation between the normal and off-normal samples in the supervised analysis.

A Q-residual and T 2 plot was prepared for a PCA model built on the simulated

NaI spectra representing normal raffinate samples (Figure 13). Datasets representing off- normal conditions, including 0.1 M and 1 M acid concentration raffinate, were projected onto the model and fell outside the 99% confidence level of the Q-residual for the normal datasets. The G1 1 M samples fell inside the T 2 99% confidence interval, which

indicated that the spectral variance captured by the model fits within the normal dataset confidence interval. However, the spectral variance not captured by the model is significant enough to place the sample outside of the normal confidence interval for the

Q-residual, clearly discerning it as off-normal.

-7 x 10 1.4

1.2

1 P 0.1 M

0.8 G2 0.1 M G1 1 M A 0.1 M 0.6 G1 0.1 M

G2 1 M Q Residuals (0.51%) P 1 M 0.4 2.25 M A 1 M A, G1, G2, P Normal Off Normal (Excluded) 0.2 0 10 20 30 40 50 99% Confidence Level 99% Confidence Level 0 0 200 400 600 800 1000 1200 1400 1600 T^2 (99.49%)

Figure 13. Q-residual and T 2 plot of the PCA for the NaI, normal aqueous raffinate samples (A = 16, G1 = 21.7, G2 = 23.4, and P = 28.7 MWd/kgU).

PCA was capable of delineating NaI spectra representing normal and off-normal conditions. However, the ability to distinguish between normal and off-normal samples decreased with the resolution of the detector used.

Cooling Time

The MIP monitor concept was further tested against more complex and realistic datasets representing three physical variables, including fuel cooling time, acid 52

concentration, and burnup level. Such complex mixtures may be expected to make up the holdings of the dissolver tank at a reprocessing facility. PCA was performed on simulated HPGe spectra of organic extract. Three PCs were necessary to capture 99.97% of the total variance. PC 1 captured greater than 88.3% of the total variance and the samples clearly grouped according to cooling time (Figure 14) along this axis, with the greatest separation between the 3 and 15 year cooled fuel. This was expected since radioactive decay is an exponential process, and would be most significant between 3 year and 15 year cooled fuel.

0.06

0.05

0.04

0.03

0.02

0.01

Scores PC on 2 (9.20%) -0.01

-0.02 26 Yr 15 Yr -0.03 3 Yr

-0.04 -0.15 -0.1 -0.05 0 0.05 0.1 Scores on PC 1 (83.31%)

Figure 14. Scores plot of the Ge, organic samples including fuel decayed for 3, 15, and 26 years (PC1 vs PC2).

Adding PC’s 2 and 3 allowed the groups shown in Figure 14 to be further delineated as a function of fuel burnup level and acid concentration. This organization is illustrated by

Figure 15.

0.05 P 0.1 M G1 0.1 M 0.04 A 0.1 M

P 1 M 0.03 G1 1 M P 0.1 M 0.02 A 1 M G2 0.1 M G1 0.1 M P 2.25M A 0.1 M 0.01 G1 2.25 M P 1 M P 0.1 M A 2.25 M G2 1 M

Scores PC on 3 (7.46%) P 1 M 0 G1 1 M G1 0.1 M P 2.25 M G1 1 M A 1 M P 2.25 M A 0.1 M 26 Yr -0.01 G1 2.25 M G2 2.25 M A 1 M 15 Yr G1 2.25 M A 2.25 M 3 Yr A 2.25 M -0.02 -0.04 -0.03 -0.02 -0.01 0 0.01 0.02 0.03 0.04 0.05 0.06 Scores on PC 2 (9.20%)

Figure 15. Scores plot of the Ge, organic samples including fuel (A = 16, G1 = 21.7, G2 = 23.4, and P = 28.7 MWd/kgU) decayed for 3, 15, and 26 years (PC2 vs PC3).

Although the groupings of data in Figure 15 overlap, each grouping is clearly delineated in 3-D space when PC1 is considered along with PC2 and PC3. Groups of data points around 2.25 M maintain their order according to acid concentration similar to the expanded view in Figure 5. These scores plots indicate that it may be possible to distinguish cooling time, burnup and acid concentration from the measurements taken by a single gamma detector.

A model using three PCs was created based on spectra representing the normal acid concentration (~2.25 M) for all cooling times and burnups. The remaining spectra

(0.1 and 1 M), considered to be off-normal, were then projected onto the model. The Q- residual and T 2 plot, shown in Figure 16, suggests that normal and off-normal samples can be distinguished at a confidence interval greater than 99%.

-5 x 10 5 A 0.1 M 4.5 Normal Off Normal 4 99.4% Confidence Level 99.4% Confidence Level 3.5 G1 0.1 M

15 yr A 0.1 M G2 0.1 M 3 3 yr A 0.1 M 15 yr G1 0.1 M 2.5 P 0.1 M

2 3 yr G1 0.1 M 15 yr1 P 0.1 M

Q Residuals (0.03%) 3 yr P 0.1 M 15 yr A 1 M G1 1 M 1.5 15 yr G1 1 M 3 yr P 1 M G2 1 M 3 yr A 1 M A 1 M 15 yr P 1 M 0.5 3 yr G1 1 M P 1 M 1 3 yr P 2.25 M 3 yr A 2.25 M 0.5 Normal Group 0 0 50 100 150 200 0 0 100 200 300 400 500 600 700 800 T^2 (99.97%)

Figure 16. Q-residuals plot of PCA model of Ge, organic normal spectra with off-normal spectra projected onto the model (A = 16, G1 = 21.7, G2 = 23.4, and P = 28.7 MWd/kgU).

Overall these results suggest that spectra of organic extract solution representing different extraction acid concentrations and fuels of various cooling and irradiation times may be clustered using multivariate analysis into desired groups. Here, process anomalies were detected regardless of the burnup level or cooling time of the fuel.

Extending Multivariate Analysis to Quantitative Predictions

Multivariate analysis was tested for the ability to quantify distinguishing physical features of spent fuel through gamma spectra. Here, PLS was applied to the simulated spectra of the organic extract of dissolved and separated fuel of different burnup levels.

A PLS model was created using sections A and P of the 26 year cooled fuel in an attempt to predict the burnup level of the samples from section G. Leave-one-out cross validation 117 was used to improve the model fit and determine the optimal number of

Latent Variables required to capture necessary variance within the dataset. A comparison of the measured burnup levels versus the predicted burnup levels is illustrated in Table 4.

Measuredl Predicted Bias * Label (MWd/kgU) (MWd/kgU) (%) G1 0.1 M 21.7 20.96 3.4 G1 1 M 21.7 21.10 2.8 G1 2.13 M 21.7 21.20 2.3 G1 2.19 M 21.7 21.21 2.2 G1 2.25 M 21.7 21.23 2.2 G1 2.30 M 21.7 21.22 2.2 G1 2.36 M 21.7 21.22 2.2 G2 0.1 M 23.4 23.62 0.9 G2 1 M 23.4 23.30 0.4 G2 2.13 M 23.4 23.03 1.6 G2 2.19 M 23.4 23.01 1.7 G2 2.25 M 23.4 23.02 1.6 G2 2.30 M 23.4 22.99 1.8 G2 2.36 M 23.4 22.99 1.8 * |Measured – Predicted| / Measured × 100

Table 4. Predictions of burnup level of fuels G1 and G2 based on simulated spectra of the organic extract from separations performed at various acid concentrations for fuels A and P.

Model predictions of the burnup levels of section G were within 3.5% of actual values, demonstrating the ability of the MIP concept as a potential tool for quantitatively estimating this characteristic.

Summary and Conclusion

Model simulations were used to validate the MIP monitor concept as a novel means for safeguarding nuclear reprocessing facilities. Multivariate analysis of simulated gamma ray spectra from process streams was able to successfully cluster spectra as a function of acid concentration, burnup and cooling time. As expected, the simulated spectra of the aqueous raffinate did not group as well as the simulated spectra of the organic extract. However, off-normal conditions were distinguishable from spectra representing normal conditions in both raffinate and organic extract streams.

Gamma-ray spectra from four different detectors were evaluated for potential application using the MIP approach. Of these, spectra from the HPGe detector were most easily grouped. CZT, LaBr 3 and NaI detectors, none of which require cryogenic cooling, provided slightly inferior results to that of the HPGe detector. All detectors tested, however, were capable of delineating off-normal conditions from normal conditions of select major process variables.

Both HCA and PCA were able to detect off-normal conditions from gamma spectra, providing independent confirmation of both techniques. In addition, PCA was able to correctly categorize datasets as a function of three major variables tested: cooling time, burnup level, and acid concentration. PLS analysis conducted on simulated HPGe spectra of organic extract was able to correctly predict burnup levels of unknowns to

within 3.5%. However, experimental verification will be necessary to confirm these results. Results of this modeling study suggest the MIP monitor concept, a new monitoring tool for safeguards inspectors, is valid.

1 K.R. Beebe, R.J. Pell, M.B. Seasholtz, Chemometrics: A Practical Guide , John Wiley & Sons, Inc., New York, NY, 1998.

Chapter 4: Experimental Test of the Multi-Isotope Process Monitor

Introduction

This chapter describes results from experiments designed to test the concept of the

Multi-Isotope Process (MIP) monitor, a novel safeguards approach for process monitoring in reprocessing plants introduced previously in Chapter 1. The MIP monitor concept was tested using model simulation and the results were described in Chapter 3.

The methods used in the experiment and analysis were described in Chapter 2. Three segments of commercial spent nuclear fuel with variations in burnup level and cooling time were dissolved and subjected to a batch PUREX extraction to separate the uranium and plutonium from fission and activation products. Gamma spectra were recorded by high purity germanium (HPGe) and cadmium zinc telluride (CZT) detectors. Hierarchal

Cluster Analysis (HCA) and Principal Component Analysis (PCA) were applied to the spectra to investigate spectral variations as a function of acid concentration, tri-butyl phosphate (TBP) concentration, burnup and cooling time. Partial Least Squares was utilized to extract quantitative information about process variables, such as acid concentration. This chapter summarizes the results of this series of experiments and analyses designed to validate the MIP monitor concept.

Results and Analysis

An example of the aqueous feed, raffinate and organic extract spectra as taken by a CZT detector is shown in Figure 17. The spectra are artificially offset for ease of pattern comparison.

Feed Raffinate Extract Log Scale Counts Log Comparison Clarity Comparison Artificially Offset for

0 100 200 300 400 500 600 700 800 900 1000 keV

Figure 17. Gamma spectra as taken by a CZT detector of the aqueous feed, raffinate and organic extract for the ATM 109 fuel separation.

Figure 17 illustrates the similarity observed between the aqueous feed and raffinate spectra and the relatively large difference between the organic extract and the aqueous solutions. Similar observations were made with spectra from the HPGe spectra of the same solutions (not shown).

Hierarchal Cluster Analysis

HCA was performed on the normalized and mean-centered spectra taken on the

HPGe detector of the organic extract of all three fuel samples. The resulting dendrogram is illustrated in Figure 18.

16 5.1 M 14 0.3 M 9 M 12 6.5 M 7.5 M 10 3.8 M 2.5 M 8 1.3 M 3.5 M 6 3.5 M 3.8 M 4 5.1 M 2.5 M ATM 105 A 2 1.3 M ATM 105 P ATM 109 0.3 M 0 0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1 Distance to Nearest Neighbor

Figure 18. A dendrogram from Hierarchical Cluster Analysis of the organic fraction, HPGe spectra (normalized and mean centered) for samples with variable acid concentration and burnup.

HCA showed a tight grouping of samples from Section A of the ATM 105 fuel. No obvious groupings were observed with samples from the other fuels. Replicate samples of organic extract separated at 3.5 M acid concentration from fuel P were closest to each other within the dendrogram in Figure 18. ATM 109 samples at 5.1 M and 0.3 M are the furthest from the group. A greater number of replicate samples will be necessary to further study possible organization of the spectra of the organic extract using HCA.

The aqueous raffinate samples recorded on HPGe detector were also analyzed by

HCA. The resulting dendrogram from these samples is shown in Figure 19.

10 3.8 M

9 1.3 M

8 2.5 M

7 5.1 M

6 0.3 M

5 5.1 M

4 1.3 M

3 3.8 M

2 2.5 M ATM 105 A

1 0.3 M ATM 109

0 0 1 2 3 4 5 Distance to Nearest Neighbor -3 x 10

Figure 19. A dendrogram from Hierarchical Cluster Analysis of the aqueous raffinate fraction, HPGe spectra (normalized and mean centered) for samples with variable acid concentration and burnup.

Spectra from these samples grouped according to burnup. The ATM 105 Section P raffinate samples were left out of the analysis due to some artificially introduced variance that will be explained in the PCA analysis section. One of the key differences between the organic extract and aqueous raffinate is the magnitude of Euclidean distance separating the spectra. The organic data has approximately an order of magnitude more

distance in between its spectra than the raffinate spectra. The raffinate spectra have approximately the same amount of Euclidian distance as that of the aqueous feed samples as seen in Figure 20.

20 1.3 M Feed 1.3 M Raffinate 18 2.5 M Feed 2.5 M Raffinate 16 3.8 M Raffinate 5.1 M Raffinate 14 0.3 M Raffinate 3.8 M Feed 12 5.1 M Feed 0.3 M Feed 10 5.1 M Feed 1.3 M Raffinate 8 3.8 M Raffinate 2.5 M Raffinate 6 0.3 M Raffinate 3.8 M Feed 4 5.1 M Raffinate ATM 105 A 2.5 M Feed 2 1.3 M Feed ATM 109 0.3 M Feed 0 -1 0 1 2 3 4 5 Distance to Nearest Neighbor -3 x 10

Given that all the feed samples contain the same stock of radionuclides, the spectra within the feed samples were expected to be nearly identical for each fuel, and the cluster distance within each fuel to represent stochastic variation. When Euclidean distances in

the raffinate and feed dendrograms were compared, it was apparent that the raffinate spectra were not greatly affected by the extraction.

For example, when the ATM 105 Section A spectra were compared within HCA analysis, the feed and raffinate were intermixed and are separated by Euclidean distances of 4E-04 to 6E-04. The organic samples from ATM 105 Section A, on the other hand, were separated by Euclidean distances of 6E-03 to 1E-02. From these results, the organic extract spectra appear to be affected to a greater extent by changes in acid concentration than the aqueous raffinate, suggesting that the MIP monitor would be most sensitive to process changes if applied to organic rather than aqueous streams.

Similar trends were observed when HCA was applied to spectra sample sets taken by a CZT detector. The Euclidean distances between the feed and raffinate spectra when analyzed together ranged from 1.0E-03 to 1.8E-03, within the ATM 105 Section A fuel.

The feed and raffinate groups were not distinct, but instead were intermixed. The HCA of the organic spectra from the ATM 105 Section A spectra resulted in intergroup

Euclidiean distances of 7E-03 to 9E-03. The similar distance for spectra collected on both CZT and HPGe suggests that sensitivity to major process variables like acid concentration may not be greatly diminished by using lower resolution CZT detectors.

Principal Component Analysis – Aqueous

PCA was applied to the HPGe and CZT spectra of the aqueous raffinate. The results of PCA analysis provided improved unsupervised groupings over HCA, as well as insight into temporal effects seen within spectra from ATM 105 Section P. The

unsupervised scores plot of the Principal Components (PCs) of the aqueous feed and raffinate samples is shown in Figure 21.

3.E-03 Raffinate Feed ATM 105 A 2.E-03 ATM 105 P ATM 109

1.E-03

0.E+00

-1.E-03 Scores on PC 2 (9.32%) 2 PC Scores on

-2.E-03

-3.E-03 -1.E-02 -7.E-03 -2.E-03 3.E-03 8.E-03 Scores on PC 1 (90.16%)

Figure 21. Unsupervised scores plot from the PCA of HPGe spectra of the aqueous feed and raffinate of experimentally separated fuels performed at a variety of acid and TBP concentrations.

Here, samples are distinctly grouped according to burnup and cooling time. Figure 21 also illustrates the similarity of the feed and raffinate spectra. The ATM 105 Section A and ATM 109 groups have overlap between their feed and the raffinate spectra, indicating the need for additional preprocessing and/or replicate samples at each acid concentration to be able to refine the model and make the organization of the samples by

process information possible. Interestingly, the spectra from the ATM 105 P raffinate appeared to separate distinctly from the feed and followed a semblance of order according to acid and TBP concentration (Figure 22).

3.E-03

2.E-03 50% TBP 30% TBP ATM 105 P Feed 40% TBP ATM 105 P Raffinate 25% TBP Day 2 1.E-03 20% TBP 3.5 M (2) 9 M 7.5 M 0.E+00 6.5 M

5 M -1.E-03 3.5 M Day 1 7.5 M Scores on PC 2 Scores on (9.32%) PC

-2.E-03 5 M 9 M 6.5 M 3.5 M (1)

-3.E-03 -1.E-02 -1.E-02 -8.E-03 -6.E-03 -4.E-03 -2.E-03 0.E+00 2.E-03 Scores on PC 1 (90.16%)

However, upon further inspection it became apparent during that this trend was due to a temporal counting bias that occurred over a single day in which spectra for ATM 105

Section P were collected (Day 2). This bias is observed when the distance from the PC

score of the 3.5 M raffinate spectra (Figure 22) to the other scores is calculated and plotted as a function of the time of collection (Figure 23).

0.008

50% TBP Raffinate 0.007 40% TBP Raffinate 30% TBP Raffinate 0.006 25% TBP Raffinate

0.005 20% TBP Raffinate

3.5 M Feed (2) 0.004

9 M Raffinate 0.003 7.5 M Raffinate

0.002 6.5 M Raffinate Distance from PCA of scorefrom Distance First Sample (3.5 M Raffinate) M (3.5 Sample First

0.001 5 M Raffinate 3.5 M Raffinate 0 12:16:00 13:28:00 14:40:00 15:52:00 17:04:00 18:16:00 19:28:00 20:40:00 21:52:00 23:04:00 Time of Day

Figure 23. Plot of the distance between the PC 1 & 2 scores of the ATM 109 Section P aqueous spectra collected on the same day and on the same HPGe detector versus the time of collection.

The PC 1 and 2 scores for the spectra collected on this day are shifting with time, indicating a temporal shift of the spectral pattern. The cause of the trend in PCA scores space was discovered to be due to a shift in gain in the detector electronics. A similar shift in the 661.7 keV peak was observed when the ATM 105 Section P spectral data was examined over the two days (Figure 24).

2364.600 Day 1 Day 2 661.900 2364.400

661.850 2364.200

661.800

2364.000

661.750 keV

Channel 2363.800 661.700

2363.600 661.650

2363.400 ATM 105 Section P 661.600 Day 1 Average (661.7 keV)

2363.200 661.550

d d te te e te te te e e ate a a at a a a e e in n inate inate n n F F f ffi f ffi ffin ffin a a a affin a M M Feed Feed (1) 5 M Feed 9 Raf R Raf R Feed (2) Raffi R R R M .5 M M P P P P 6 7.5 M M m Raff M M B .5 5 5 .5 9 TB TB 3.5 3 6. 7 3.5 % % 5 0 20% T 2 30% TB 40% TBP 5Raffinate

Figure 24. A comparison of the peak channel of the 661.7 keV peak for the feed and raffinate solutions of the dissolved ATM 105 P fuel.

The peak energy in the spectra for all of the ATM 105 P feed solutions taken on day one are within a standard deviation of their average. However, the peak energy increasingly deviates from this average value with time for all spectra taken on day 2. It is significant to note this total deviation is less than one channel (< 0.3 keV), an insignificant shift considering the resolution of the detector (~1.1 keV FWHM at 661.keV). A shift of this magnitude would have no effect on results of peak area calculations using traditional gamma energy analysis. In contrast, this was a significant shift with respect to the analysis of the spectra after preprocessing by PCA.

One likely explanation of this trend is the effect of temperature on the electronics used with the HPGe detector system. Temperature is known to affect electronics 1 without gain stabilization. We believe this may have resulted from variations in ambient temperatures in the count room during counting. The two subsequent days over which all of the ATM 105 Section P aqueous samples were counted reached record outside temperatures. We believe this may have strained the climate control system of the HPGe counting lab, resulting in higher than normal ambient temperatures within the counting room. The temperature logs of the room also indicated that that temperature in the room that day was above average.

The loadings of the principal components (not shown) representing the aqueous solutions for all fuel types were reviewed. They indicated the variance within and between the aqueous feed and raffinate samples were dominated by the Cs-137 peak at

661 keV. This was somewhat expected since Cs-137 does not preferentially extract from the aqueous phase into the organic extract. Figure 21 illustrates that whatever changes were made by the extraction of Pu and U were small enough to be overshadowed by minor temporal gain shifts. The aqueous spectra showed some organization according to burnup and cooling information. With gain stabilization, additional preprocessing and

PC selection, as well as the addition of replicate samples, PCA may prove sensitive to process variations within the aqueous raffinate.

Unsupervised PCA was also performed on the CZT spectra of the aqueous solutions. In this case, no temporal shifts were observed. In the scores plot, the fuels were clearly separated. As expected, the PC loadings favored the 0 – 150 keV energy region due to the low counting efficiency of this type of detector for higher gamma ray

energies. The small crystal size of CZT limited the collection of gamma lines over 150 keV, though the 661 keV peak was still discernable due to the large activity of Cs-137 present in the sample.

Principal Component Analysis – Organic

An unsupervised PCA model was created using HPGe Spectra from the organic extract. Three principal components (PCs) were used to capture 99.07% of the total variance. However, only two PCs were necessary to separate the three groups of fuel samples in PCA scores space (Figure 25).

0.015

1.3 M ATM 105 A 3.8 M 0.01 ATM 105 P 2.5 M ATM 109 0.3 M

0.005 9 M 5.1 M 0 3.5 M 6.5 M 7.5 M

-0.005

Scores on PC (3.77%) 3 3.5 M 1.3 M 5.1 M 2.5 M 0.3 M -0.01 3.8 M

-0.015 -0.05 -0.04 -0.03 -0.02 -0.01 0 0.01 0.02 Scores on PC 2 (20.33%)

Figure 25. Scores plot from the PCA of HPGe spectra of the organic extract of experimentally separated fuels performed at a variety of acid concentrations with 30% TBP in dodecane.

From Figure 25, the data grouped according to fuel type, but did not organize according to acid concentration. The ATM 105 P organic extract samples were observed to have the most spread within the groups. Notably, the spread of the replicate 3.5 M extraction samples of the ATM 105 P fuel likely captures variation inherent to the separations process, as a result of the stability of detector electronics, and random noise from the stochastic process of decay. In order for the PCA to show organization according to acid concentration, additional replicate spectra were necessary.

Unsupervised PCA analysis of the spectra from the organic samples collected using a CZT detector resulted in a similar scores plot (Figure 26).

0.02 5.1 M

0.01 3.8 M 0.3 M 2.5 M 1.3 M 3.8 M 2.5 M 6.5 M 0 5.1 M 7.5 M 1.3 M 3.5 M 9 M

-0.01 3.5 M (2)

-0.02 Scores on PC 2 (4.60%)

-0.03 0.3 M ATM 105 A ATM 105 P ATM 109 -0.04 -0.06 -0.04 -0.02 0 0.02 0.04 0.06 0.08 Scores on PC 1 (94.86%)

Figure 26. Scores plot from the PCA of CZT spectra of the organic extract of experimentally separated fuels performed at a variety of acid concentrations with 30% TBP in dodecane.

Two PC’s were necessary to distinguish fuel types. Datasets were not distinguishable by acid concentration. One exception to this was in the dataset from ATM 109. These data showed some tendency to group according to acid concentration. Duplicate samples will be necessary to confirm the observed trend.

Though duplicate samples were not available, due to the expense and effort required by experiments conducted within a radiological hot cell, replicate counts on the

HPGe detector were performed of the organic samples of the ATM 105 A and ATM 109 datasets. In random order, each sample was counted ten times on the HPGe detector.

While this did not attempt to capture process variance, it did succeed in introducing detector variations and random noise into the spectral patterns. The unsupervised organization in the scores plot of ATM 109 is illustrated in Figure 27.

0.005

-0.005

-0.01

-0.015 Scores on PC 2 (0.75%) 0.3 M 1.3 M -0.02 2.5 M 3.8 M 5.1 M

-0.025 -0.08 -0.06 -0.04 -0.02 0 0.02 0.04 0.06 0.08 Scores on PC 1 (99.02%)

Figure 27. Scores plot from the PCA of replicate HPGe spectra of the organic extract of the ATM 109 segment at a variety of acid concentrations with 30% TBP in dodecane.

PCA successfully grouped the datasets according to acid concentration for both ATM 109

(Figure 27) and ATM 105 Section A (not shown). Three PC’s were necessary to capture

98.87% of the variance and separate the spectra when the ATM 109 and ATM 105

Section A spectra were combined in the same analysis. When combined, PCA successfully grouped datasets according to acid concentration, burnup level and/or cooling time. Data from ATM 109 were slightly better resolved as a function of acid concentration than the data from ATM 105 Section A. No advantage was observed when the ATM 105 Section A and ATM 109 were analyzed separately.

A supervised PCA model was created based on spectra from the 2.5 M acid concentration (defined here as the normal process condition) of ATM 105 A and ATM

109 fuels to test the ability of the MIP approach for distinguishing between normal and off-normal process conditions. Spectra of dissolved fuels extracted at other acid concentrations were then projected onto the PCA space of the “normal” model. The Q- residuals and T 2 plot for the analysis of the ATM 105 Section A dataset is shown in

Figure 28.

-4 x 10 1.5

2.5 M 0.3 M 1.3 M 3.8 M 5.1 M 1 99% Confidence Level 99% Confidence Level

Q Q Residuals (4.88%) 0.5

0 0 10 20 30 40 50 60 70 80 90 100 T^2 (95.12%)

From Figure 28, sample sets representing “off-normal” acid concentrations fell outside the Q-residual 99% confidence line of data representing normal conditions. A minimum of two PCs were sufficient to distinguish the off-normal samples from the normal, but four PCs were needed and used to capture 95.12% of the total variance. Nearly all of the datasets fell within the 99% confidence interval for the T 2 indicating the captured variance did not separate the normal from the off-normal samples. However, the Q- residual confidence line correctly distinguished the samples as off-normal and normal datasets. Similar results were obtained with the ATM 109 datasets, although in this case 75

the off-normal datasets fell out side the the 99% confidence interval for both Q-residuals and T 2 (Figure 29).

-3 x 10 1.8

1 1.6

1.4 0 0 10 20 30 40 50 60 1.2

0.8

2.5 M Q Residuals (1.40%) 0.6 0.3 M (Excluded) 1.3 M (Excluded) 0.4 3.8 M (Excluded) 5.1 M (Excluded) 0.2 99% Confidence Level 99% Confidence Level 0 0 500 1000 1500 2000 2500 3000 3500 T^2 (98.60%)

A PCA model was also built around the combined normal (2.5 M) sample set for both ATM 109 and ATM 105A. Off-normal samples from both fuels were projected onto this model in a similar fashion as before to investigate whether PCA could distinguish between normal and off-normal datasets regardless of burnup level or cooling time. Using four PCs, a Q-residual and T 2 plot was generated and is shown in Figure 30.

-4 x 10

2.5 M 0.3 M 1.3 M 3.8 M 5.1 M 2 99% Confidence Level 99% Confidence Level

Q Q Residuals (0.53%) 1

0 0 10 20 30 40 50 60 Hotelling T^2 (99.47%)

Figure 30. Q residual and T 2 plot from the supervised PCA of both the ATM 105 A and ATM 109 sample spectra collected on a HPGe detector, 20 spectra for each sample listed in the legend (10 per fuel type).

From Figure 30, normal and off-normal samples are distinguishable by the Q-residual, though several normal and off-normal points are near the confidence line.

Partial Least Squares

PLS was performed on replicate spectra from fuel segments ATM 109 and ATM

105 Section A in an attempt to extract quantitative information regarding acid concentration. Forty-five of the spectra of each fuel type were used as the benign training

set, with the remaining five spectra (representing each acid concentration) used as test samples to compare model predictions with experimentally determined values. Leave- one-out cross validation 2 was used to refine the model, improve its predictive power, and determine the optimal number of Latent Variables (LV). Six LV were used for both the

ATM 109 and ATM 105 Section A sample sets. The root mean square error of prediction

(RMSEP) 17 was calculated and served as a single value quantification of the precision of fit between the model and experiment. Results of the PLS predictions for the two fuel segments are shown in Table 5.

ATM 105 A ATM 109 Measured Predicted Bias Measured Predicted Bias (M) (M) (M) (M) (M) (M) 0.3 0.8 +0.5 0.3 0.6 +0.3 1.3 1.9 +0.6 1.3 1.3 ±0.0 2.5 2.8 +0.3 2.5 2.7 -0.2 3.8 3.5 -0.3 3.8 3.8 ±0.0 5.1 4.5 -0.6 5.1 5.1 ±0.0

Table 5. Predictions of the extraction acid concentration of fuels ATM 105 A and ATM 109 based on spectra (HPGe) of the organic extract from separations performed at various acid concentrations.

The RMSEP for the ATM 109 and ATM 105 Section A models were 0.15 and

0.47, respectively. These results suggest that acid concentration of aqueous feed stream at a reprocessing facility might be determined from NDA gamma-ray measurement of the organic extract stream.

Summary and Conclusion

The Mulit-Isotope Process monitoring approach was shown through experiment to be capable of identifying off-normal process conditions for separations of fuels with

different characteristics. Multivariate analysis was able to group samples according to burnup level, acid concentration and cooling time. The aqueous spectra from both CZT and HPGe were dominated by the Compton edge created by the Cs-137 gamma line at

661 keV. However, PCA was still able to clearly distinguish the different fuels. PCA of spectra from HPGe detectors were sensitive to minor temporal spectrum shifts, possibly due to the use of detector eletronics that were not gain stabilized over periods of time in which the ambient temperature within the laboratory was changing. Organic spectra were more sensitive to acid change than spectra taken from the aqueous raffinate. Also demonstrated experimentally, the aqueous feed acid concentration was determined from

HPGe spectra of the organic streams to within 0.5 M using PLS.

The results of this study confirm the viability and validity of the MIP monitor concept. These experiments demonstrate that gamma spectroscopy, when combined with multivariate analysis, provide a potential monitoring tool for MC&A around reprocessing facilities.

1 Model DSA-2000 Digital Spectrum Analyzer technical specification sheet, © 2002 Canberra Inc. 2 K.R. Beebe, R.J. Pell, M.B. Seasholtz, Chemometrics: A Practical Guide , John Wiley & Sons, Inc., New York, NY, 1998.

Chapter 5: Comparison of Model to Experiment

Introduction

The MIP Monitoring concept was demonstrated through a series of simulations

(in Chapter 3) as well as experiments (Chapter 4). While qualitatively these studies generated similar results, there were still significant differences between the simulated and experimentally determined spectra. This chapter discusses differences between modeled and experimentally determined CZT gamma spectra and subsequent PCA analysis generated from each.

Comparison of Model to Experiment

Modeled and experimental spectra representing aqueous raffinate from similar spent nuclear fuel sources are provided in Figure 31. The raffinate stream was from a liquid-liquid extraction of dissolved spent fuel (16 MWd/kgU), separated at an approximate acid concentration of ~2.5 M acid. Apparent from Figure 31, model simulations were unable to fully reproduce the structure of the experimentally derived spectrum. Limitations likely exist in all three models used to generate the simulated spectrum including ORIGEN-ARP, AMUSE and Synth.

1.0E+07

1.0E+06

1.0E+05

1.0E+04

1.0E+03 59.5 keV 241 123 keV 1.0E+02 Am 154 Eu

Counts 105.3 keV 1.0E+01 86.6 keV 155 Eu 1.0E+00 74.7 keV 243 1.0E-01 Am

Raffinate 1.0E-02 Simulated Raffinate

1.0E-03 0 100 200 300 400 500 600 700 keV

Figure 31. Modeled and experimental CZT spectrum of ~2.5 M raffinate of the 16 MWd/kgU fuel.

Discrepancies due to the potential inaccuracy in ORIGEN output were considered as potential sources of error. Though it appears that the model misrepresents several nuclide activities, it was not clear whether this was an error in the input file or the calculation methods employed by ORIGEN. A few active elements produced by

ORIGEN were left out of the liquid extraction model and subsequently, the simulated spectra. These elements were omitted because the constants used to calculate their distribution by AMUSE were determined to be either missing or inaccurate. However, none of the omitted nuclides had gamma lines in the areas where discrepancies were identified, so these emissions would not have caused the discrepancies identified below.

Another possible, but less probable, explanation for model versus experimental peak discrepancies is the omission of peaks by the gamma library accessed by Synth to

populate the spectra. This possibility would suggest Synth is mishandling major gamma lines of various isotopes in order for this to be significant within the Compton region of the spectra. Most gamma libraries are relatively accurate with respect to major gamma lines, typically only lacking accurate information for low branching ratio peaks or very high energy lines that are difficult to detect.

The most pronounced difference between modeled and experimentally derived spectra is the Compton scatter region of the 661 keV gamma energy peak from 137 Cs.

Compton edge differences in the simulated spectra were due to limitations of the Synth code. These limitations have been previously identified and discussed for Ge detectors 1 and were attributed to neglected scatter from the shielding material. These limitations likely exist for the CZT simulation described here. Synth creates the Compton region by using the theoretical shape of the Compton spectrum, based on a peak-to-Compton ratio appropriate for the volume of the detector. The code then adds these counts to the spectrum for each peak.

For the purpose of testing the MIP approach, these issues in the simulated

Compton edge are thought to be of minor importance. Here, spectra were used for relative comparisons in the MIP monitor approach. Since the Compton edge (correctly or incorrectly simulated) would be relatively consistent between datasets, inaccurate, but consistent, simulations would not contribute significantly to the overall variance between spectral groups.

Modeled spectra were also lacking a few key peaks at 75 and 86 keV, and over- emphasized the peak at 123 keV. Since these discrepancies were also apparent on the feed spectra for the 16 MWd/kgU fuel (before separation), these issues likely originated

from either ORIGEN-ARP or Synth, rather than AMUSE. The spectral region beyond the 137 Cs (661 keV) peak was not shown, as the counts are too low on the experimental spectrum to produce any meaningful peaks.

The missing gamma line at 75 keV in the simulated spectrum was due to either inaccuracies in handling 243 Am or the Compton contribution from 137 Cs. Upon further investigation, 243 Am contributes a significant number of counts in the 75 keV region of the modeled spectrum. However, this contribution did not rise above the modeled

Compton background. It is unclear whether the modeled Am-243 activity is the result of this missing peak or an overestimated Compton effect from the 661 keV gamma line.

The simulated spectra were also missing a gamma line at 86 keV seen in the experimental spectra. At the 86 keV line, the model identified 155 Eu as the main contributor. However, 155 Eu also has a 105 keV gamma ray, with a similar branching ratio (21% for 105 keV versus 31% for 86 keV), for which a peak was not apparent in the experimental spectrum. This does not rule out 155 Eu as the cause, because the detector has a sharp drop of efficiency with energy. Synth does not model the efficiency curve perfectly, leaving the root cause of the missing 86 keV peak questionable. 243 Am contributes a small fraction of its decay gammas to the 86 keV line as well, though its low branching ratio (0.34%) makes its activity an unlikely source of error. Additional analysis is necessary to identify the cause of this discrepancy.

The relatively oversized simulated peak at 123 keV is primarily produced by Eu-

154. One possible explanation for the overemphasis is the inability of Synth to accurately model the rapid decline of efficiency of the CZT crystal with energy. It is

possible that this decrease in efficiency is sufficient to account for the smaller peak in the experimental spectrum.

The spectra from both the model and the experiment were analyzed using an unsupervised Principal Component approach. CZT spectra of the organic fraction of the liquid extraction were analyzed, following normalization by area and mean centering. A scores plot of the unsupervised PCA of the simulated spectra is found in Figure 32.

-3 x 10 4 1 M 2.25 M 0.1 M 2 2.25 M

0 1 M 2.25 M

0.1 M 1 M

-2 0.1 M 2.25 M -4 Scores PC on 2 (4.41%)

1 M 16 MWd/kgU -6 21.7 MWd/kgU 23.4 MWd/kgU 28.7 MWd/kgU 0.1 M -8 -0.04 -0.03 -0.02 -0.01 0 0.01 0.02 Scores on PC 1 (95.39%)

Figure 32. Scores plot from PCA of simulated organic extract spectra using a CZT detector

Figure 32 clearly illustrates how the samples grouped according to burn up and acid concentration in PCA scores space. Simulated aqueous samples also grouped in a like manner.

0.02 5.1 M

0.01 3.8 M 0.3 M 2.5 M 3.8 M 2.5 M 1.3 M 6.5 M 0 5.1 M 7.5 M 1.3 M 3.5 M 9.0 M

-0.01 3.5 M (2)

-0.02 Scores PC on 2 (4.60%)

15.5 - 17.5 MWd/kgU -0.03 0.3 M 29 - 31 MWd/kgU 67 - 70 MWd/kgU

-0.04 -0.06 -0.04 -0.02 0 0.02 0.04 0.06 0.08 Scores on PC 1 (94.86%)

Figure 33. Scores plot from PCA of experimental organic extract spectra using a CZT detector

Similar PCA, as conducted using corresponding experiment spectra, is shown in Figure

33. From this figure it is clear that experimental samples do not group as distinctly in

PCA scores space as the simulated spectra. Due to resource constraints, multiple samples at each acid concentration were not available, which would have enhanced PCA of the experimental spectra. However, the experimental spectra are still able to be grouped according to burnup, and there may be potential for grouping according to the extraction acid concentration with advanced preprocessing of the data. The close proximity of spectra taken from duplicate samples at 3.5 M acid concentration for the ATM 109 fuel

(67-70 MWd/kgU) suggests process variations in replicate spectra were relatively small.

Further analysis of both the experimental and simulated spectra will be necessary to delineate these pattern shifts.

While comparisons of model and experimental spectra have identified important differences and shortcomings of simulations, it is important, also, to acknowledge the similarity of the model spectra to the experiment. For instance, several prominent peaks are visible in both the simulated and experimentally derived spectra, including the 661,

123, 60, 43, 36, and 32 keV lines. Also, simulations correctly demonstrated the capability of compiling multivariate analysis and gamma spectroscopy as a promising technique for safeguards technology development. Further study is necessary to determine the root causes of the minor differences between model and experiment and to test and validate the MIP approach within more realistic, continuous flow systems.

Summary and Conclusion

The results of the experimental validation of the MIP modeling concept were compared to model simulations. PCA scores plots generated from modeled spectra were more distinguishable by both burnup level and acid concentration than with experimentally derived spectra. Several peaks were missing from the simulated gamma spectra, and potential causes of this discrepancy were hypothesized, including limitations in Synth, ORIGEN-ARP and AMUSE. Nonetheless, the PCA based on the experimentally derived spectra were able to correctly identify off-normal conditions at a

99% confidence interval. Overall, the model represented the experimental observations for the purpose of testing the MIP monitoring concept.

1 W.K. Hensley, A.D. McKinnon, H.S. Miley, M.E. Panisko, and R.M Savard. “Synth: A Spectrum Synthesizer.” Journal of Radioanalytical and Nuclear Chemistry, Articles, Vol. 193, No. 2, p. 229-237, 1995.

Chapter 6: Conclusion

Summary of Conclusions

The Mulit-Isotope Process Monitor concept was validated through model and experiment. In Chapter 2, methods used during simulations and lab experiments were described in detail, as well as the multivariate analysis techniques used to analyze both the simulated and experimental spectra. The results of these analyses were presented in

Chapters 3 and 4, and a comparison of results from modeling to experimental investigations was presented in Chapter 5.

Model Analysis

Model simulations were used to validate the MIP monitor concept. Simulated spectra were prepared by way of model to represent the feed, raffinate and extract solutions of a simulated PUREX extraction of spent nuclear fuel. Seven extraction acid concentrations representing both normal and off-normal process conditions were modeled. This included fuel with four burnup levels and three cooling times. Spectra were simulated based on the response of four detectors, including HPGe, CZT, LaBr 3, and NaI.

Multivariate analysis of simulated gamma ray spectra from process streams was able to successfully cluster spectra as a function of acid concentration, burnup and cooling time. As expected, the simulated spectra of the aqueous raffinate did not group as

well as the simulated spectra of the organic extract. However, off-normal conditions were distinguishable from spectra representing normal conditions in both raffinate and organic extract streams.

Gamma-ray spectra from four different detectors were evaluated for potential application using the MIP approach. Of these, spectra from the HPGe detector were most easily grouped. CZT, LaBr 3 and NaI detectors, none of which require cryogenic cooling, provided slightly inferior results to that of the HPGe detector. However, all detectors tested were capable of delineating off-normal conditions from normal conditions of select major process variables.

Both HCA and PCA were able to detect off-normal conditions from gamma spectra, providing independent confirmation of the techniques. In addition, PCA was able to correctly categorize datasets as a function of three major variables tested: cooling time, burnup level, and acid concentration. PLS analysis conducted on simulated HPGe spectra of organic extract was able to correctly predict burnup levels of unknowns to within 3.5%. Experimental verification will be necessary to confirm these results.

Results of this modeling study suggest the MIP monitor concept, a new monitoring tool for safeguards inspectors, is valid.

Experimental Analysis

The Mulit-Isotope Process monitoring approach was shown through experiment to be capable of identifying off-normal process conditions for separations of fuels with different characteristics. For the experiment, spectra were collected from batch extraction feed, raffinate and extract streams from a batch PUREX extraction of 3

segments of spent nuclear fuel. Each fuel segment was extracted at 5 different acid concentrations, and the ATM 105 Section P fuel segment was also extracted at 5 TBP concentrations. The fuel segments spanned various burnup levels and cooling times.

Spectra were collected on both HPGe and CZT detectors. Ten replicate spectra were collected for each of the organic fractions of the ATM 105A and ATM 109 fuel segments.

Multivariate analysis was able to group samples according to burn up, acid concentration and cooling time. The aqueous spectra from both CZT and HPGe were dominated by the Compton edge created by the Cs-137 gamma line at 661 keV.

However, PCA was able to clearly distinguish the different fuels. PCA of spectra from

HPGe detectors were sensitive to minor temporal spectrum shifts, possibly due to the use of detector eletronics that were not gain stabilized over periods of time in which the ambient temperature within the laboratory was changing. Organic spectra were more sensitive to acid change than spectra taken from the aqueous raffinate. Also demonstrated experimentally, the aqueous feed acid concentration was determined from

HPGe spectra of the organic streams to within 0.5 M using PLS.

Comparison of Model and Experiment

The results of the experimental validation of the MIP modeling concept were compared to model simulations. PCA scores plots generated from modeled spectra were more distinguishable by both burnup level and acid concentration. Several peaks were missing from the simulated gamma spectra, and potential causes of this discrepancy were hypothesized, including limitations in Synth, ORIGEN-ARP and AMUSE. Nonetheless, the PCA based on the experimentally derived spectra were able to correctly identify off- normal conditions at a 99% confidence interval. Overall, the model reasonably characterized the experimental observations for the purpose of testing the MIP monitoring concept.

Need for Further Work

The Multi-Isotope Process monitor concept has been shown to be viable by model and experiment. Additional research and development will be required to demonstrate its full range of capabilities and to implement it as an MC&A tool. This would include testing the approach under additional process changes, including replicates of the experiments discussed in this dissertation and other changes with multiple replicates such as temperature. Testing on a continuous extraction flow loop will also be necessary, to investigate the approach in a setting that is closer to industrial methods.

Advanced methods of spectra collection and multivariate analysis might also be necessary to increase the effectiveness of the MIP monitor approach. Coincidence counting may provide increased pattern sensitivity and consistency and is under development as part of the MIP monitor. Additional multivariate analysis techniques are

being reviewed to search for the optimal handling of the pattern variances resulting from process changes. These research areas may provide further capability refinement.

Online, non destructive and near-real time process monitoring is one improvement envisioned for material control and accountancy. By focusing on the validation of declared process conditions, one can assure that the material remains within the intended path throughout the facility. With assurance of material control, the burden placed on accountancy measurements is lightened, reducing the requirement for increasingly accurate and expensive destructive analysis. A framework for how process monitoring would relieve the burden on destructive analysis has not yet been widely discussed. A framework for safeguards implementation will be required before process monitoring can achieve its full effectiveness as an accountancy tool.

The MIP monitor is designed to ensure the integrity of the process thereby assuring the international community that no safeguarded material has been diverted for unintended purposes by way of process manipulation. It has the potential to achieve this goal with minimal inspector interaction, non-destructively, online and in near real time.

References

Dwight D. Eisenhower, “Atoms for Peace,” An address given to the 470 th Plenary Meeting of the United Nations General Assembly, December 8 th , 1953. Accessed at http://world-nuclear-university.org/html/atoms_for_peace/ (January 2009).

D. Fisher, History of the International Atomic Energy Agency: the first forty years , Vienna, Austria: The International Atomic Energy Agency (IAEA), 1997.

IAEA Department of Safeguards, “IAEA Safeguards: Staying Ahead of the Game,” Vienna, Austria: The International Atomic Energy Agency, July 2007.

International Atomic Energy Agency, “IAEA Safeguards Glossary, 2001 Edition,” Vienna, Austria: The International Atomic Energy Agency, 2001.

International Atomic Energy Agency, “Safeguards Techniques and Equipment, 2003 Edition,” Vienna, Austria: The International Atomic Energy Agency, 2003.

INFCIRC/153 (Corrected). The Structure and Content of Agreements between the Agency and States Required in Connection with the Treaty on the Non-Proliferation of Nuclear Weapons, 1972.

International Atomic Energy Agency, “IAEA Safeguards: Stemming the Spread of Nuclear Weapons,” Vienna, Austria: The International Atomic Energy Agency, accessed electronically at http://www.iaea.org/Publications/Factsheets/English/S1_Safeguards.pdf , March 2009.

U.S. Congress, Office of Technology Assessment, “Nuclear Safeguards and the International Atomic Energy Agency,” OTA-ISS-615, Washington, D.C.: U.S. Government Printing Office, June 1995.

O. Yamamura, R. Yamamoto, S. Nomura and Y. Fujii, “Development of safeguards and maintenance technology in Tokai Reprocessing Plant,” Progress in Nuclear Energy , Vol. 50, p. 666-673, 2008.

J. E. Doyle, Editor, Nuclear Safeguards, Security, and Nonproliferation , Butterworth- Heinemann (Elsevier), Burlington, MA., 2008.

P.C. Durst et al., “Advanced Safeguards Approaches for New Reprocessing Facilities,” ASA-100 report, Pacific Northwest National Laboratory, PNNL-16674, June 2007.

International Atomic Energy Agency, “Report of the LASCAR Forum: Large Scale Reprocessing Plant Safeguards,” STI/PUB/922, IAEA, Vienna, Austria, 1992.

Personal correspondence with Shirley Johnson, 11 May 2009.

International Atomic Energy Agency, “IAEA Annual Report, Table A4. Approximate quantities of material subject to Agency safeguards at the end of 2007,” Accessed at http://www.iaea.org/OurWork/SV/Safeguards/sv.html on 13 May, 2009.

S.J. Johnson et al., “Development of the Safeguards Approach for the Rokkasho Reprocessing Plant,” International Atomonic Energy Agency, IAEA-SM-367/8/01, 2001.

M. Suzuki, M. Hori, S. Nagaoka, and T. Kimura, “Study on Loss Detection Algorithms Using Tank Monitoring Data,” Journal of Nuclear Science and Technology , Vol. 46, No. 2, p. 184-192, 2009.

P.M. Rinard and H.O. Menlove, “Application of Curium Measurements for Safeguarding At Reprocessing Plants,” LA-13134-MS, Los Alamos National Laboratory, 1996.

G.R. Choppin, and B.E. Stout, “Plutonium – the element of surprise.” Chemistry in Britain , 12, 1126-1129, 1991.

L.R. Morss, N.M. Edelstein and J. Fuger, Editors, The Chemistry of the Actinide and Transactinide Elements , 3 rd Ed. Springer, Dordrecht, The Netherlands, 2006.

C. Riffard, H. Toubon, S. Pelletier, M. Batifol, and J.M. Vidal, “MOX Fuel Characterization for Burnup Credit Application: Extension of Nondestructive Method Qualified for LEU Fuels,” Nuclear Technology , Vol. 154, p. 186-193, May 2006.

K. Oeda, H. Naito, M. Hirota, K. Natsume, and H. Kumanomido, “Calibration of Burnup Monitor Installed in Rokkasho Reprocessing Plant,” Journal of Nuclear Science and Technology , Vol. 37, No. 6, p. 543-547, June 2000.

S. Anilkumar, A.K. Deepa, K. Narayani, A.K. Rekha, P.V. Achuthan, G. Krishnamachari, and D.N. Sharma, “Estimation of 235 U concentration in some depleted uranium samples by high resolution gamma-ray spectrometry using 185 keV and 1001 keV gamma-energies of 235 U and 234m Pa,” Journal of Radioanalytical and Nuclear Chemistry , Vol. 274, No. 1, p. 161-166, 2007.

A. Morgenstern, C. Apostolidis, H. Ottmar and K. Mayer, “Analysis of 237 Np in spent fuel solutions,” Radiochim. Acta , Vol. 90, p. 389-393, 2002.

C.K. Mathews and P.R. Vasudeva Rao, “Radioactivity in Monitoring Materials Processing,” Journal of Radioanalytical and Nuclear Chemistry, Articles , Vol. 203, No. 2, p. 519-535, 1996.

H.T. Matsuda, B.F. de Araujo, and J.A. de Araujo, “Analytical Process Control of the Celeste R&D Installation of IPEN-CNEN/SP,” J. Radioanal. Nucl. Chem., Letters , Vol. 199, No. 6, p. 453-463, 1995.

F. Baumgärtner and D. Ertel, “The Modern PUREX Process and its Analytical Requirements,” Journal of Radioanalytical Chemistry , Vol. 58, p. 11-28, 1980.

S.G. Marathe, V.K. Rao, V.K. Bhargava, R.H. Iyer and M.V. Ramaniah, “A Gamma- Monitoring Assembly for the Analysis of Fission Products in Reprocessing Streams,” Nuclear Instruments and Methods , Vol. 127, p. 99-103, 1975.

M. Suzuki, M. Hori, R Asou & S. Usuda, “Numerical Consideration for Multiscale Statistical Process Control Method Applied to Nuclear Material Accountancy,” Journal of Nuclear Science and Technology , Vol. 43, No. 10, p. 1270-1279, 2006.

R.C. Runkle, M.F. Tardiff, K.K. Anderson, D.K. Carlson and L.E. Smith, “Analysis of Spectroscopic Radiation Portal Monitor Data Using Principal Component Analysis,” IEEE Transactions on Nuclear Science , Vol. 53, No. 3, p. 1418-1423, 2006.

L. Chen and Y. Wei, “Nuclide identification algorithm based on K-L transform and neural networks,” Nuclear Instruments and Methods in Physics Research A , Vol. 598, p. 450-453, 2009.

M. Benedict, T.M. Pigford, and H.W. Levi, Nuclear Chemical Engineering , 2 nd Ed. McGraw-Hill, New York, NY, 1981.

L.E. Smith, J.M. Schwantes, J.J. Ressler, M. Douglas, K.A. Anderson, C.G. Fraga, P.C. Durst, C.R. Orton, R.N. Christensen, “Next Generation On-line MC&A Technologies for Reprocessing Plants,” Proceedings of Global 2007 Conference on Future Nuclear Energy Systems, 2007.

J.M. Schwantes, M. Douglas, C.R. Orton, C. Fraga and R.N. Christensen, “Multi-Isotope Process (MIP) Monitor: a Near-Real-Time Monitor for Reprocessing Facilities,” ANS Transactions from the Annual Meeting, Anaheim, CA, 2008.

C.R. Orton, J.M. Schwantes, S. Bryan, T. Levitskaia, D. Duckworth, M. Douglas, O.T. Farmer, C. Fraga, S. Lehn, M. Liezers, S. Peper, R.N. Christensen, “Advanced Safeguards Technology Demonstration at Pacific Northwest National Laboratory,” Proceedings of the 49 th Annual INMM Conference, Nashville, TN, 2008.

E. Malinowski, Factor analysis in chemistry , John Wiley & Sons, New York, NY, 405p, 2008.

A.G. Croff, “ORIGEN2: A versatile computer code for calculating the nuclide compositions and characteristics of nuclear materials,” Nucl. Technol. , 62, 3, 1983.

Guenther, R. J. et al., “Characterization of Spent Fuel Approved Testing Material – ATM-105,” Pacific Northwest Laboratory, PNL-5109-105, June 1989, accessed at http://www.lsnnet.gov/docview.aspx?mode=1&lsn=DN2002138631&ic=1&im=0&sc =12&sm=0 (October 2008)

M. C. Regalbuto, J. M. Copple, R. Leonard, C. Pereira, G. F. Vandegrift, “Solvent Extraction Process Development for Partitioning and Transmutation of Spent Fuel,” Proceeding of the 8 th Information Exchange Meeting on Actinide and Fission Product Partitioning and Transmutation:, Las Vegas, Nevada, United States, November 9-1, pp. 373-385, 2004; © Organization for Economic Co-operation and Development, Nuclear Energy Agency (OECD-NEA): Paris, France, 2005, accessed at http://www.nea.fr/html/pt/docs/iem/lasvegas04/posterI.html (October 2009).

G.F. Vandegrift, D.B. Chamberlain, C. Conner, J.M. Copple, J.A. Dow, L. Everson, J.C. Hutter, R.A. Leonard, L. Nunez, M.C. Regalbuto, J. Sedlet, B. Srinivasan, S. Weber, and D.G. Wygmans, "Development and Demonstration of the TRUEX Solvent Extraction Process," Proceedings of the Symposium on Waste management, Tucson, AZ, pp. 1045-1050, Feb. 28-Mar. 4, 1993.

R.A. Leonard and M.C. Regalbuto, "A Spreadsheet Algorithm for Stagewise Solvent- Extraction," Solvent Extraction and Ion Exchange , Vol. 12, Issue 5, pp 909 – 930, 1994.

L.A. Bromley, “Thermodynamic Properties of Stroping Electrolytes in Aqueous Solutions,” AIChE J. , 19, 313, 1973.

R.W. Perkins and U.P. Jenquin, “Fission and Activation Products in Nuclear Reactor Fuels and Nuclear Explosion Debris,” PNNL-11554, Pacific Northwest National Laboratory, 1997.

T. Ishimori and K. Watanabe, “Inorganic Extraction Studies on the System of Tri-n-butyl Phosphate-Nitric Acid,” Bulletin of the Chemical Society of Japan , Vol. 33, No. 10, 1960.

W.K. Hensley, A.D. McKinnon, H.S. Miley, M.E. Panisko, and R.M. Savard, “SYNTH: a spectrum synthesizer,” J. Radioanal. Nucl. Chem ., 193, 229, 2005.

B.D. Milbrath, B.J. Choate, J.E. Fast, W.K. Hensley, R.T. Kouzes, J.E. Schweppe, “Comparison of LaBr 3:Ce and NaI(Tl) scintillators for radio-isotope identification

devices,” Nuclear Instruments and methods in Phisyics Research A 572, p. 774-784, 2007.

J.M. Schwantes et al.,”Medium-resolution autonomous in situ gamma detection system for marine and coastal waters,” J. Radioanal. Nucl. Chem., in press/available online.

S. Vaidyanathan, R.D. Reager, R.W. Warner et al., “High Burnup BWR Fuel Pellet Performance,” American Nuclear Society, Proceedings of the International Topical Meeting on Light Water Reactor Fuel Performance, Portland, OR, p. 471 March 2-6, 1997.

S.F. Wolf, D.L. Bowers, J.C. Cunnane, “Analysis of high burnup spent nuclear fuel by ICP-MS,” Journal of Radioanalytical and Nuclear Chemistry , Vol. 263, No. 3, p. 581-586, 2005.

PLS_Toolbax Version 5.0 for use with MATLAB, Eigenvector Research, Inc., Wenatchee, WA, 2008.

Matlab, Ver. 7.8.0.347 (R2009a), The MathWorks, Inc., 2009.

K.R. Beebe, R.J. Pell, M.B. Seasholtz, Chemometrics: A Practical Guide , John Wiley & Sons, Inc., New York, NY, 1998.

B.M. Wise et al., PLS_Toolbax Version 4.0 for use with MATLAB, Manual , Eigenvector Research, Inc, Wenatchee, WA, 2006.

S. Dragovic & A. Onjia, “Classification of soil samples according to their geographic origin using gamma-ray spectrometry and principle component analysis,” Journal of Environmental Radioactivity , 89, p. 150-158, 2006.

P.K. Hopke, “The evolution of chemometrics,” Analytica Chimica Acta , 500, p. 365-377, 2003.

P. Geladi, B. R. Kowalski, Anal. Chim. Acta , 185, p. 1-17, 1986.

H. Martens and T. Naes, Multivariate Calibration , John Wiley & Sons, Ltd., Chichester, 1989.