uncertainty quantification for inverse radiation transport student posters_0.pdf · airborne...
TRANSCRIPT
Identify the unconstrained
subspaceCorresponds to null singular values
Mentor: K. Bledsoe With: P. Hausladen, M. Blackston, J. Lefebvre Program: NESLS / Reactors and Nuclear Systems Division / Nuclear Security Modeling Group
Uncertainty Quantification for Inverse Radiation Transport
Aaron M. Bevill – University of Michigan
Signal ProcessingGain correction
Pulse-shape discrimination
Intrinsic efficiency
Object
Coded aperture
mask
Organic
scintillator
bank
Counts
y
z
Hidden parameters
True source distribution
Sampled data
~50K count rates
Forward model
Calibrated ray tracing
Parameters’ confidence region
~1M voxels’ source intensity
Quantity-of-interest confidence interval
Total source intensity
Hit pattern
Mask
intersections
Predicted
counts
3D Geometry
Cross-validation resultsPredicted minus measured
Chi-squared
confidence testReject holdup distributions
that are unlikely to produce
the data
Confidence regionAll non-negative distributions
that pass the test
Confidence intervalLower and upper bound on
holdup
Define confidence interval Evaluate confidence interval
s0 (neutrons / second)
s1
(neutr
ons / s
econd)
First, find the
CR (green
arrows)
Second, seek
min(s0 + s1) &
max(s0 + s1)
(white arrows)
CR bounded
by white oval
Convex optimization problemsSolved using Newton’s method with logarithmic barriers
Result: 3.43 < s0 + s1 < 4.52 (blue dotted lines)
Hidden parameters
True characteristics
Sampled data
4 to 8 count rates
Forward model
Transport
linearized response
Parameters’ Bayesian posterior
Mean and covariance of 2 to 6 parameters
Generalized Linear Least Squares
Linear Equality Constraints
Transform the problemSo that the residuals follow a
standard normal distribution
Minimize squared
residualsUsing linear algebra
Calculate covariance
GLLSOn the unconstrained subspace
Underdetermined Problems
Overdetermined Problems
Linearize the problem
Singular value
decomposition
This work was funded by DOE's National Nuclear Security Administration.
Open-source air sampling missions during
aboveground US nuclear weapons testing
from 1945 to 1963 provided diagnostic data,
however; these parameters do not address the
needs of National Technical Nuclear Forensics
to inform attribution. To fill this gap Oak Ridge
National Laboratory is developing the Airborne
Planning Tool, which provides predictive data
based on specified routes to optimize the
collection of quality volatile samples
(r-value > 1.0) with a sufficient quantity of
equivalent fissions for radiochemical analysis.
Nuclear Preparedness: Air Sampling for Today and the Future
Introduction
Samuel J. Cope, North Carolina State University, NESLS Intern
Mentor: Vincent J. Jodoin, Nuclear Security Modeling, RNSD
Active Air Sampler Networks
Airborne Planning Tool
Additional Developments
Right: EPA RadNet System for national
radiation monitoring
Below: IMS map for radionuclide stations
Conclusions
• Volatile-air-sample collection is necessary for
rapid nuclear forensics
• The Airborne Planning Tool optimizes
- fixed air sampler locations
- aircraft collection flight missions
• Qualitative and quantitative predictive results
as specified for radiochemical analysis
This work was funded by the Office of Defense Nuclear
Nonproliferation Research and Development (NA-22), within
the U.S. Department of Energy's National Nuclear Security
Administration and the Next Generation Safeguards Initiative
(NA-24).
• GIS export for collection-mission path
• Fixed sampler placement optimization
• Average particle density sensitivity
• Verification and validation of DELFIC and
HYSPLIT integration
• Fractionation ratio display for sample quality
Airborne Planning Tool GUI: Buster-Jangle Sugar shot
sample collection path and particle mass distribution
Equivalent fissions per cubic meter collected along the specified mission
Dose rate and cumulative dose at the exterior of the plane over the
specified mission
• 79 of 80 radionuclide stations mapped with
high-volume samplers and HPGe detectors
• Only sensitive to particulate gamma emitters
• 90% detection of 1 kt explosion debris within
14 days
• Commercially available, automatic system
sends data to International Data Center
• Design requirement of 300 cfm;
up to 650 cfm
International Monitoring System (IMS)
EPA RadNet
• >130 continuous high-volume samplers
across all 50 states
• 40 additional deployable monitors
available for dispatching
• Automated gamma measurements
• Alerts to laboratory staff in response to
significant increases in radiation levels
• Custom Hi-Q design, ~35 cfm
Left: Knoxville EPA RadNet air sampler unit by Hi-Q
Top Right: NaI(Tl) gamma detector above filter
Bottom right: 4-in.-diameter polyester filter with debris
• Industrial designs by Staplex, F&J, Hi-Q,
ThermoFisher Scientific, and Tisch
Environmental
• Portable designs capable of up to 70 cfm
Above: Boeing WC-135 US Air Force
radiation “sniffer” plane
Below: Mushroom cloud following
detonation of the Buster-Jangle Sugar shot
• Aerosol-collection mission planning
• DELFIC (cloud rise) + HYSPLIT (aerosol transport)
• Determination of optimal placement of fixed
samplers and deployable samplers
LEGEND:
EPA RadNet air sampler
IMS Radionuclide Aerosol Sampler
Left: Staplex high-volume portable sampler
Right: F&J low-volume portable air sampler
Off-the-Shelf Air Samplers
Acknowledgements:
Repeating History: A Review of Air Sampling Techniques and Their
Application to Modern Nuclear Forensics
• What are we looking for?
– Characteristic debris from a nuclear explosion(i.e., fallout or airborne particles)
• Refractory elements (Zr, W, Mo)
• Volatile elements (I, Xe, Br)
• Air Sampling v. Fallout Sampling
– Fallout sampling:collection of debrisafter it has landedon the ground
– Air sampling:collectionof debris mid-transport
• Ground-based airsampling (portable vs.stationary)
• Aerial sampling
– Manned vs.unmanned
– Various flight paths Acknowledgments:
• Timeline:
– 1945: First nuclear weapon tested (Trinity)
– 1946–1962: US weapon development
program operated, air sampling occurred
– 1963: Limited Test Ban Treaty is
implemented, only underground
detonations allowed
– 1992: U.S. Weapon testing ceased
– Present day: R&D of technical nuclear
forensics capabilities
• What is nuclear forensics?
– Techniques used to characterize and
support attribution of nuclear material
before/after detonation of a device
• Destructive (radiochemistry)
• Non-destructive (radiation detection)
BackgroundEmilie K. Fenske, University of Tennessee, NESLS
Mentor: Vince J. Jodoin, Nuclear Security Modeling, RNSD
Historic Air Sampling
Historic Air Sampling Data and Fallout Modeling
A Comparison
What’s Next?
• Main sample criteria: Volatile particles
• Valuable data include:
– Equivalent fissions (EF/m3)
– Particle sizes
– Ratio of volatile/refractory mass chains
• Airborne Planning Tool (APTool) underdevelopment
– Similar to Fallout Planning Tool
– Placement of air samplers in scenarios:fixed vs. portable
– Additional capabilities?
• Review and application of current airsamplers
Can an air sampler collect
adequate samples for
nuclear forensics?
Historic air sampling data &
assessment
Historic air sampling data &
assessment
Fallout modeling (DELFIC)
Fallout modeling (DELFIC)
Airborne particulate modeling
Airborne particulate modeling
Overlay of weapon test Buster-Jangle Sugar DELFIC
results and air sampler locations.
High volume ground-based air
sampler located at the Nevada
Test Site.
• SS-23: Operation Sandstone Report, 1948.
• WT-811: Distribution and Characteristics of Fallout, Upshot-Knothole, 1953.
• WT-1178: Distribution and Characterization of Fallout, OperationTeapot, 1955.
• DELFIC: Department of Defense Fallout Prediction System,1979.
• Defense Land Fallout Interpretive Code (DELFIC)
– Fallout prediction/planning tool (GUI)
– Used for DOE exercises with a focus onrefractory samples close to GZ
– Models transport of fallout, NOT airborneparticles
• Cannot directly compare air sampling data
• Applicable results include:
– Plume movement
– Volatile/refractory sample locations
Preliminary Conclusions
• Based on historic ground-based air sampling
data, adequate samples can be obtained for
radiochemical analysis.
• 1010–1012 equivalent fissions possible
• Based on DELFIC results: 50–100 miles from
GZ, samples collected should be volatile.
Historical data (equivalent fissions) from
weapon test Teapot Met at three distances from
GZ (and at ± 6 mi perpendicular to hot line)
Yes/NoYes/No
SamplerOnset
time* (hr)
Cessation
time* (hr)
Distance from GZ
(mi.)
Ratio*(140:99 M.C.)
Historic
Samples
(EF)
Sn
ap
pe
r F
ox Alamo -- -- 54 1.1 (volatile) 1.7E11
Crystal
Springs4.0 9.4 55 1.01 (volatile) 7.1E11
Groom
Mine2.1 7.6 24 1.1 (volatile) 7.9E11
Su
ga
r 29 0.11 0.23 3 0.55 (refract.) 8.7E9**
36 0.21 0.41 10 0.62 (refract.) 1.6E9**
39 0.35 0.67 16 0.68 (refract.) 1.4E9**
References:
This work was funded by the Office of Defense Nuclear
Nonproliferation Research and Development (NA-22) within
the U.S. DOE’s National Nuclear Security Administration.
*from DELFIC
**Calculated with an averaged concentration over 2 hours
Schematic of a ground-based
fallout collector (open-close
collector).
Schematic of filter locations on a B-17 for aerial air sampling.
DELFIC results (ratio of mass chains 140:99) of weapon
test Snapper Fox shown on Google Earth with pins
where air samplers were located.
Sensitivity Study of INDEPTH for Verification of Facility Spent Nuclear Fuel Declarations
Scott Richards – The University of Tennessee
Mentor: Brandon R. Grogan Program: NESLS Nuclear Security Modeling Group, Reactor and Nuclear Systems Division
To be able to distinguish if a prolonged cooling period hasoccurred, the isotopes of interest for the measurements shouldbe tracked through the forward ORIGEN calculation forpotential markers.
Rather than using indirectly measured quantities, INDEPTH cansolve using isotopics. To see more direct correlations,reconstructions with different groups of isotopes can beevaluated.
Results
Future Work
INDEPTH
INDEPTH finds aminimum within thesolution space byperforming a gradientsearch, based on the sumof square errors (SSE), ofthe measurement datagiven.
The forward ORIGENcalculations are done inINDEPTH as a singleirradiation and decay.
The accuracy of thecalculations is limited bythe degeneracy of thesolution. As the examplesto the right show, there isa spread of parametersthat result in the samesolution within theuncertainty.
Correlation and Convergence
The SSE is a measure of how well the foundsolution matches the measured data; however, itwas found not to be directly correlated to howwell the final INDEPTH parameters matched thecorrect parameters from the ORIGEN simulations
One of the goals of International Safeguards analysts is toverify the initial enrichments, burnups, and cooling timevalues of spent nuclear fuel declared by facilities usingnondestructive assay (NDA) measurements.
Motivation and Background
The three measurement types analyzed did show a uniformlyworse answer for all the cases with a long cooling time betweenirradiations (cases 3, 4, 10, 11, and 12), most notably in the casesfor which there was a simulated reactor shutdown event beforethe final irradiation cycle, such as cases 10 and 11. This result wasexpected, however the magnitude of the effect was not.
Scope of Study
The Inverse Depletion Theory (INDEPTH) code attemptsto reconstruct the initial enrichment and operatinghistory by using the Oak Ridge Isotope Generation(ORIGEN) code to simulate the irradiation and cooling ofthe fuel.
The study consisted of 13 different variations on the baseparameters of interest, with case 0 being the case the others areperturbations of and most closely simulating actual fuelconditions.
Each of the listed cases had 36 baseline combinations of assemblytype (2 BWR and 2 PWR designs), enrichments (2.0 and 3.5%),burnups (20 and 45GWd/MTU), and decay times (10, 20, and 30years).
All cases were solved by INDEPTH using absolute gammas, relativegammas, and absolute gammas with gross neutron countsgenerated by simulating irradiation and cooling with ORIGEN.
This work was funded under the Next Generation Safeguards Initiative—Spent Fuel project and the work was done as part of the NESLS internship at Oak Ridge National Laboratory
Varying most parameters had no tangible effecton error percentages. The only parameter thathad an effect was decay time between cycles.Generally, the longer the decay time betweencycles, the more inaccurate the results were.
In no case was the INDEPTH solution using relative gamma themost accurate solution. In fact, over the 5 factors of comparison,the absolute gamma measurement with gross neutron countswas as accurate or up to 10 times more accurate for the caseaverages.
Only in cases with extended cooling times between irradiationswas the added gross neutron information detrimental to theaccuracy of the solutions. The cause of this effect is currently beinginvestigated; it is possibly due to the way in which the 244Cminventory is changing, which cannot be reflected in the wayINDEPTH simulates the history.
134Cs + 137Cs + 154Eu degeneracy space: 5yrs cooled
134Cs + 137Cs + 244Cm degeneracy space: 5yrs cooled
134Cs + 137Cs + 244Cm degeneracy space: 15yrs cooled
Library Enrichments Burnup (GWd/MTU) Cooling Times (years)
BWR 8x8 2.00% 20 10, 20, 30
BWR 8x8 3.50% 20 10, 20, 30
BWR 8x8 3.50% 45 10, 20, 30
PWR 15x15 2.00% 20 10, 20, 30
PWR 15x15 3.50% 20 10, 20, 30
PWR 15x15 3.50% 45 10, 20, 30
BWR 10x10 2.00% 20 10, 20, 30
BWR 10x10 3.50% 20 10, 20, 30
BWR 10x10 3.50% 45 10, 20, 30
PWR 17x17 2.00% 20 10, 20, 30
PWR 17x17 3.50% 20 10, 20, 30
PWR 17x17 3.50% 45 10, 20, 30
Variants in Cases
Case
Number
Number of
Cycles
Length of
Cycle (d)
Length Down
between
cycles (d)
Power (MW/MTU)Burn
time (d)Notes
Case 0/Base 5 330 30 12.12/27.27 1650Closest to comparison data for
actual assemblies
Case 1 1 1770 0 11.30/20.45 1770
Closest to how INDEPTH
simulates conditions, with 3
back to back burn cycles and
one decay
Case 2 5 440 30 9.091/20.46 2200 Longer power cycles
Case 3 5 270 105 14.81/33.34 1350
Longer down times between
power cycles but similar
overall cycle length
Case 4 5 330 105 12.12/27.27 1650
Longer down times between
power cycles but same length
power cycles
Case 5 4 420 30 11.90/26.79 1680 Fewer cycles
Case 6 6 270 30 12.35/27.78 1620 More cycles
Case 7 5 330 30(20.20 or 10.10)/
(45.45 or 22.73)1650
Different powers on cycles
(first cycle higher)
Case 8 5 330 30(20.20 or 10.10)/
(45.45 or 22.73)1650
Different powers on cycles
(middle cycle higher)
Case 9 5 330 30(20.20 or 10.10)/
(45.45 or 22.73)1650
Different powers on cycles
(last cycle higher)
Case 10 5 330 30 12.12/27.27 1650300 day decay between the
4th and 5th cycle
Case 11 5 330 30 12.12/27.27 1650900 day decay between the
4th and 5th cycle
Case 12 5 330 30 12.12/27.27 1650900 day decay between the 1st
and 2nd cyle
Case Burnup Down Enrich Burn Power
0 31.37% 129.05% 123.71% 30.41% 29.47%
1 62.24% 215.32% 223.05% 36.22% 63.89%
2 28.64% 136.40% 123.54% 46.48% 28.52%
3 6.37% 22.01% 4.02% 251.58% 81.22%
4 6.43% 41.11% 5.23% 260.12% 83.16%
5 49.04% 118.78% 157.67% 60.80% 47.80%
6 50.39% 131.97% 158.32% 51.90% 46.54%
7 40.04% 111.86% 115.33% 20.91% 25.83%
8 59.57% 120.12% 145.06% 35.74% 47.96%
9 24.59% 94.87% 53.50% 30.39% 30.26%
10 8.72% 25.24% 7.99% 174.34% 81.56%
11 8.00% 23.72% 8.96% 312.03% 97.79%
12 15.46% 77.93% 10.46% 357.55% 121.14%
Absolute Gamma/Relative Gamma Error
Case Burnup Down Enrich Burn Power
0 13.69% 74.95% 12.34% 35.51% 22.12%
1 10.91% 20.60% 9.89% 50.99% 14.74%
2 8.89% 66.17% 8.59% 38.86% 16.63%
3 7.12% 85.59% 41.48% 44.84% 16.13%
4 7.58% 115.06% 59.92% 48.59% 15.54%
5 17.14% 54.73% 18.99% 66.52% 28.09%
6 12.85% 69.25% 13.47% 44.15% 23.29%
7 19.29% 91.40% 12.66% 22.37% 22.90%
8 11.04% 78.45% 11.82% 32.72% 23.26%
9 7.78% 90.28% 3.85% 45.73% 24.44%
10 8.01% 94.24% 70.72% 44.07% 24.40%
11 10.11% 108.95% 59.65% 38.96% 20.35%
12 16.62% 85.48% 61.01% 42.86% 8.30%
Absolute Gamma+Gross Neutron Count/Relative Gamma Error
Case Burnup Down Enrich Burn Power
0 43.64% 58.08% 9.98% 116.77% 75.05%
1 17.52% 9.57% 4.43% 140.78% 23.07%
2 31.05% 48.51% 6.95% 83.62% 58.31%
3 111.86% 388.93% 1031.33% 17.82% 19.86%
4 117.92% 279.87% 1145.82% 18.68% 18.69%
5 34.95% 46.08% 12.05% 109.40% 58.75%
6 25.50% 52.48% 8.51% 85.07% 50.04%
7 48.19% 81.71% 10.98% 106.95% 88.68%
8 18.53% 65.31% 8.15% 91.53% 48.50%
9 31.65% 95.16% 7.20% 150.46% 80.77%
10 91.87% 373.40% 885.24% 25.28% 29.91%
11 126.46% 459.34% 665.95% 12.49% 20.81%
12 107.54% 109.69% 583.03% 11.99% 6.85%
Absolute Gamma+Gross Neutron Count/Absolute Gamma Error
Initial comparison of fallout modeling codes within Fallout Planning Tool (FPTool) and
Specialized Hazard Assessment Response Capability (SHARC)
Atomic fallout modeling provides emergency
response teams with critical maps, data, and
projections used to plan for or react to atmospheric
nuclear explosions. Current modeling programs
have room for improvements in their outputs in
terms of accuracy and precision when compared to
exposure data of known historical events and in
providing direct predictions on radiation health and
safety. Two modeling codes named AIRborne
RADiation (AIRRAD) and DEfense Land Fallout
Interpretive Code (DELFIC) are being compared to
understand the advantages of each model.
Introduction
Katie M. Cook – Texas A&M UniversityMentor: Vincent Jodoin Program: NESLS Nuclear Security Modeling (NSM) Group Nuclear Security and Isotope Technology Division (NSITD)
Method
Based on comparing code manuals and file structures:
– DELFIC offers dynamic cloud rise with wind andmeteorological updates in time while this is notpossible with AIRRAD.
– More specifically, DELFIC accounts forfractionation in exposure calculations.
– AIRRAD utilizes a computed grid rotation anglewhich is not used in DELFIC.
– DELFIC output files include a parameter recapand more detailed annotations over AIRRAD
Results
To identify differences between DEFLIC and
AIRRAD based on:
– Code manuals
– File structures
– Map types
– Ease and usability
Objectives
• Codes AIRRAD and DELFIC are utilized within the
programs Fallout Planning Tool (FPTool) and
Specialized Hazard Assessment Response
Capability (SHARC).
• 7 historical atomic tests and 3 different map types
were used for comparisons.
• FPTool and SHARC produce different types of
fallout projection maps based on date, designated
exposure duration, location, and shot data.
Method
• Above are two examples of map types generated from SHARC for twodifferent atomic tests, representing a sampling of the total 84 maps generatedbetween SHARC and FPTool.
• DELFIC overlays are often more conservative with larger hot spot areas andlarger overall projections which are also shown in the above maps.
• Initial generated DELFIC overlays tend to be cut off more frequently (alsoshown above), but map domains can be updated in the input file.
• Runtimes for AIRRAD and DELFIC range from 12 – 30 seconds for eachmap type per each historical atomic test.
• SHARC offers an auto-calculate grid for DELFIC that can increase runtimefor larger map overlays compared to AIRRAD or standard DELFIC runs.
Results
• Both AIRRAD and DELFIC can be run through the
SHARC graphical user interface (GUI) for easy
map generation.
• FPTool better utilizes the capabilities of DELFIC
map types due to special overlay tab in GUI.
• DELFIC is currently providing higher resolution
and more conservative overlays in the selected
map types analyzed.
• Due to the ability to update wind and
meteorological data in time, DELFIC is better
suited for modeling offsite clouds that have
travelled for longer times than onsite clouds.
• Both DELFIC and AIRRAD provide similar high
and low exposure data per each map type for
each historical atomic test.
• Neither AIRRAD or DELFIC had faster runtimes
for the selected tests analyzed.
Conclusions
AIRRAD or DELFIC DELFIC
Fallout projection
map
Location
Date
Fission yield
Fission type
Wind speed
Wind direction
Air pressure
Relative humidity
Altitudes
Map Type
I would like to thank Sandia National Laboratory for allowing me to assist with this project as well as Murray Purves and Jordan Lefebvre within the NSM Group for consistent software help. This work was funded by the National Nuclear Security Administration.
Acknowledgements
Sunbeam Johnnie Boy early population effectsHardtack II Humboldt groundshine overlay
AIRRAD DELFIC Comparison AIRRAD DELFIC Comparison
FPTool example map overlay from
atomic test Buster-Jangle Sugar
Quantifying Variations in Spent Nuclear Fuel Isotopic Concentrations for International Safeguards
Prediction and identification of isotopics of spent nuclear fuel(SNF) is of critical importance to international safeguards. Tothis goal, the Spent Fuel Nondestructive Assay (NDA)project seeks to predict isotopics of SNF. Of particularemphasis are non-standard reactor operating conditions,such as extended downtimes between cycles.
As more countries look to either reprocess or utilize SNFrepositories, it is important to be able to validate operatordeclarations with respect to SNF properties. Theseproperties include burnup, initial enrichment, and coolingtime.
Introduction
Michael Cooper, University of Tennessee, Knoxville
Mentor: Brandon Grogan Program: NESLS | Division: Nuclear Security and Isotope Technology | Group: Nuclear Security Modeling
To accomplish the goals highlighted, the Oak Ridge Isotope
Generation (ORIGEN) simulation code was used to model a
number of irradiation scenarios and calculate the
concentrations of isotopes of interest. ORIGEN allows for
the simulation of nuclear fuel being irradiated and decayed
in a reactor setting, based off of user inputs. This code is a
part of the SCALE software package.
For this project, a total of 156 ORIGEN runs were
performed, based on permutations of 13 irradiation history
cases wherein there were 2% and 3.5% enrichments, as
well as burnups of 20 and 45 gigawatt days per metric ton of
uranium (GWd/MTU).
Method
The results of this project show that variations in the
irradiation history of SNF can significantly alter the resulting
isotopic composition. This is important as NDA isotopes are
used to verify operator declarations. If the radiation
signatures do not match declared reactor operations, there
is a possibility of abnormal reactor behavior. The following
figures illustrate isotopic changes for several conditions.
Case 0 is the base case, with 5 uniform 300 day irradiation
cycles and 30 days downtime between cycles.
Results
This project looked at the effects of varying irradiation
histories and initial conditions for 4 fuel assembly types.
Varying the assembly type had a minimal effect between
PWR and BWR assemblies. Varying burnup and enrichment
altered the isotopics as has already been published. The
new contribution of this project is the information that altering
the irradiation history such as introducing extended
downtimes or longer irradiation periods with lower per cycle
power, can significantly alter the concentrations of NDA
isotopics. While burnup and cooling time are the major
determinants of isotope ratios, changes in the cycle history
can have a significant impact.
Summary
Figure 1. This figure shows the differences of 244Cm Isotopic
concentration for the 13 cases that were run.
Figure 1 shows that for case 11, which has a 900 day
downtime between the last two cycles, there is a 2.2%
decrease in the concentration of 244Cm. A decrease of 244Cm
could be mistaken for SNF with a slightly lower burnup if the
extended reactor downtime is not considered.
Figure 2. This figure shows the ratios of 134Cs/137Cs for
case 0 and case 11.
Figure 3 shows that altering the irradiation history can
significantly affect the NDA ratio of 134Cs/137Cs. This effect is
especially pronounced for cases 10 and 11. Case 11 has a
900 day decay between the 4th and 5th cycles, and case 10
has a 300 day decay between the same cycles. This decay
period allows 134Cs to decay away faster than 137Cs due to
its 2.07 year half-life. Case 9 shows a higher ratio of134Cs/137Cs due to that case having a higher last cycle
power, boosting the amount of fresh 134Cs.
Figure 2 shows that when there is an extended downtime
between reactor cycles, the important NDA isotopic ratio134Cs/137Cs is decreased by approximately 21.6% at
discharge and throughout cooling. This finding is very
important for NDA analysts who are examining gamma ray
signatures of SNF, and attempting to calculate
characteristics such as burnup from the SNF.
Figure 3: This figure shows how 134Cs/137Cs mass ratios vary
significantly due to changes in irradiation history. All cases
here are 10 years post-discharge.
Acknowledgments: This work was funded by the National Nuclear Security Administration.
Sigma=2%
Sigma=6%
Sigma=10%
N=2 N=4 N=8
Optimizing Training Noise Level
Characterization of Machine Learning Performance for Plutonium Production Predictions
• Models built with limited sets of nuclides for
deployability.
• How to measure core-averaged burnup from a
few specimens when local burnup is highly
variable?
• Factors affecting local burnup include:
– Position throughout core (radial & axial)
– Position within fuel pin (radial & axial)
– Operating temperature
– Neutron flux spectrum
• Many of these factors will be unknown.
Burnup as a Probe for Pu Production
Adam Drescher – The University of Texas at Austin
Mentor: Ken Dayman, Program: NESLS, Group: Nuclear Security Modeling, Division: Nuclear Security and Isotope Technology
• Can you make accurate measurements if the
operating temperature is unknown?
• Tradeoffs associated with training on a multi-
temperature dataset.
• Can the model discriminate between meaningful
and nuisance variations in the nuclide data?
• “Noise” could result from random errors in nuclide
measurements, or unspecified reactor operating
parameters such as temperature.
• Optimization problem with an ideal “level” of added
Gaussian noise.
• By adding Gaussian noise to training data,
predictions made on both clean and noisy testing
data were improved.
Reactor coreFuel pin
• A single core specimen is sufficient for determining
whole core plutonium production.
• A model can incorporate uncertainty in
measurements and reactor parameters while
achieving good burnup predictions
• A robust model for core-averaged burnup
predictions should incorporate many sources of
possible variation in fuel data.
Sets of Nuclides
Unknown Temperature?
Accounting for Noisy Data
Results & Conclusions
Developing a Model
• Reactor core-averaged burnup
• Multi-variate signature
• Deployability
• Invariant to biased core sample
• Accurate across a range of unknown parameters
• Measures a single fuel specimen
• Discriminates between meaningful and nuisancevariations in nuclide signatures
Pin Level Variations
What makes a good model?
Nuclide Concentration → Bavg
Estimated with
Bayesian inferenceNon-linear
basis
functions
Training data
Relevance Vector Machine Framework
Nested optimization problem
Sparse!
Integrated
feature
selection and
basis shaping!
• Probabilistic machine learning algorithm
• Multivariate statistical analysis
• Unique to ORNL: Integrated feature selection!
3D Inverse Depletion Challenges
Dataset Used For TrainingCombined Fuel Temp 600 K Fuel Temp 700 K
Da
tase
t U
se
d F
or Te
stin
g
Fu
el T
em
p 7
00
K
Fu
el T
em
p 6
00
K
Core Level Variations
404
39
460
39
15
6
1
10
100
1000
Actinides Actinides & Cesium Actinides Cs & Noble Gases
Ave
rage
Rel
ativ
e Er
ror
[%]
Nuclide Set
Average Error Predicting Noisy Data Burnups
Trained on Clean Data Trained on Noisy Data
Best Results
Acknowledgements: Support for my time on this project was provided by the
Consortium for Nonproliferation Enabling Capabilities. This work was funded by the
Office of Defense Nuclear Nonproliferation Research and Development (NA-22), within
the US Department of Energy’s National Nuclear Security Administration.
References
1. M.E. Tipping, “Sparse Bayesian learning and the relevance vector machine,”, The
Journal of Machine Learning Research, 1, 211-244 (2001)
2. K. Dayman, “Sparse Bayesian Regression with Integrated Feature Selection for
Nuclear Reactor Analysis”, International Conference on Mathematics &
Computational Methods Applied to Nuclear Science & Engineering
[2]
Reactor Fuel Data• Generated by TRITON
Training Data• Isotopics and
associated core-averaged burnups
RVM• Model Development
Burnup Predictions
Summary Statistics of Model Performance
Testing Data• Isotopics and
associated core-averaged burnups
Repeated for every considered set of parameters:• Sets of nuclides• Training and testing operating temperatures• Levels of Gaussian noise added
[1]
From Training on Values to Training on Distributions
[2]
• Goal: Position-independent signature of core-
averaged burnup despite highly position-
dependent local burnup.
The Rim Effect
Model fails if trained at the wrong temperature.
Models trained at multiple temperatures make good predictions across a range of temperatures.
• Can we extend the idea of training on a range of
possible values to other parameters besides
temperature?