Journal of Nuclear Science and Technology

ISSN: 0022-3131 (Print) 1881-1248 (Online) Journal homepage: https://www.tandfonline.com/loi/tnst20

Confidence interval estimation by bootstrap

method for uncertainty quantification using
random sampling method

Tomohiro Endo, Tomoaki Watanabe & Akio Yamamoto

To cite this article: Tomohiro Endo, Tomoaki Watanabe & Akio Yamamoto (2015)
Confidence interval estimation by bootstrap method for uncertainty quantification using
random sampling method, Journal of Nuclear Science and Technology, 52:7-8, 993-999, DOI:
10.1080/00223131.2015.1034216

To link to this article: https://doi.org/10.1080/00223131.2015.1034216

Published online: 27 May 2015.

Submit your article to this journal

Article views: 1904

View related articles

View Crossmark data

Citing articles: 7 View citing articles

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=tnst20
Journal of Nuclear Science and Technology, 2015
Vol. 52, Nos. 7–8, 993–999, http://dx.doi.org/10.1080/00223131.2015.1034216

ARTICLE Physor 2014

Confidence interval estimation by bootstrap method for uncertainty quantification using random
sampling method
Tomohiro Endo a∗
, Tomoaki Watanabeb, † and Akio Yamamotob, †
a
Department of Materials, Physics and Energy Engineering, Aichi 464-8603, Japan; b Graduate School of Engineering,
Nagoya University, Furo-cho, Chikusa-ku, Aichi 464-8603, Japan
(Received 21 January 2015; accepted final version for publication 18 March 2015)

Confidence interval estimation by the bootstrap method is investigated for the uncertainty quantification
of neutronics calculation using the random sampling method. The random sampling method is a simple
and practical technique to quantify an uncertainty (standard deviation) of the target parameter calculated
by a core analysis code. It is noted that a statistical error is inevitably included in the estimated uncertainty
because of the probabilistic method using random numbers. In order to estimate the statistical error of
uncertainty, we focus on the bootstrap method. The bootstrap method is one of the resampling techniques
to evaluate variance and confidence interval of a sample estimate (e.g. variance) without the assumption of
normality. Through a lattice burnup calculation for a simplified boiling water reactor (BWR) fuel assembly,
it is verified that the bootstrap method can reasonably estimate the confidence interval of uncertainty of
infinite neutron multiplication factor (kinf ) due to covariance data of JENDL-4.0. In the case of this prob-
lem, the distribution of kinf is well approximated by a normal distribution; thus, the confidence interval
of uncertainty can be also estimated by the aid of chi-squared distribution. The merit using the bootstrap
method is to simply estimate the confidence interval of uncertainty without the assumption of normality.
Keywords: bootstrap method; uncertainty quantification; random sampling method; uncertainty; covariance;
numerical analysis

1. Introduction effective methodologies, since a sophisticated sensitivity

In the present paper, the confidence interval estima- analysis code, such as a generalized adjoint analysis code
tion by the bootstrap method is investigated for the un- [7,8], is not necessary in the RS method [9–11]. Espe-
certainty quantification (UQ) of neutronics calculation cially for light water reactor (LWR) analysis, which em-
using the random sampling (RS) method. ploys a complicated multi-step calculation (lattice-core
The safety of a nuclear reactor is ensured by a core calculation) considering burnup and thermal-hydraulics
analysis. In the core analysis, the safety parameters feedback effects, the RS method is a practical approach
should satisfy their limitations. In this context, predic- to quantify uncertainties of target parameters due to
tion accuracy or uncertainty of the core analysis is cru- cross-section covariance data.
cial for the reactor safety [1,2]. There are various factors In the RS method, a large number of input pa-
for the prediction errors in the core analysis, e.g. model- rameters are randomly generated on the basis of their
ing of physical phenomena, approximations in numer- probabilistic distributions, and then target safety param-
ical calculations, fabrication tolerances for geometries eters are evaluated for each input parameter set; finally,
and material densities, and microscopic cross-section the uncertainties of target parameters are estimated by
data. Among these contributing factors, uncertainties the statistical processing, i.e. the uncertainties can be es-
(e.g. covariance data) of cross-section data [3–6] could timated by standard deviations for the samples of target
be a major factor. In order to estimate the uncertain- parameters. Thus, it is noted that the statistical errors
ties of target safety parameters due to covariance data are inevitably included in the estimated uncertainties of
of cross-section, the RS method is one of the simple and target parameters. As the number of sampling increases,

∗
Corresponding author. Email: [email protected].
†
Present address: Department of Materials, Physics and Energy Engineering, Aichi 464-8603, Japan.


C 2015 Atomic Energy Society of Japan. All rights reserved.
994 T. Endo et al.

the statistical errors can be reduced; however, it requires where σ is an RS vector of microscopic cross-
more computational costs. In actual situation, compu- section data, μ
 is the mean vector, z is a column
tational resources are finite; therefore, the number of vector whose elements are standard normal ran-
sampling is limited depending on the computational dom numbers and are independent of each other,
costs of core analysis. In order to judge the adequacy of and A is a square root matrix of the covariance
estimated uncertainties, their confidence intervals are matrix :
useful information, i.e. the number of sampling can be
properly determined whether their confidence intervals  = AAT , (2)
are acceptable. If a population of target parameters
follows a normal distribution, the confidence interval where the superscript T means transpose of a
of variance could be estimated by the aid of chi-squared matrix. The square root matrix A is numerically
distribution. However, a population does not necessar- calculated by the singular value decomposition
ily follow an ideal normal distribution; thus, a practical (SVD) for  [9,11]. Namely, using the SVD, a
estimation method of the confidence interval is desirable real symmetric matrix  can be decomposed as
without the assumption of normality. follows:
For this purpose, we investigate the applicability of
bootstrap method [12–14] to evaluate the confidence in-  = UWUT , (3)
terval of uncertainty of neutronics calculation. With-
out the assumption of normality, the bootstrap method where U is an unitary matrix where each col-
enables to practically estimate the histogram, variance, umn corresponds to eigenvector of  T and
and confidence interval for a sample estimate (e.g. mean, W is a diagonal matrix where each diagonal
variance, median) by a large number of resamples, which element corresponds to the square roots of
are obtained by “random sampling with replacement” the non-negative eigenvalue of  T . Conse-
from an original data. It is noted that “random sampling quently, several hundreds of microscopic cross-
with replacement” is one of the technical terms in statis- section libraries σi are randomly generated as
tics, i.e. each element in the original data has an equal follows:
probability to be selected and it is permissible to be re-
dundantly selected.
σi = μ
 + UW1/2 zi (i = 1, 2, . . . , N) , (4)
The contents of the present paper are as follows. In
Section 2, methodologies of the RS and the bootstrap
where W1/2 is the square root matrix of W and
methods are described. Section 3 shows numerical re-
N indicates the total number of RS libraries.
sults using the bootstrap method for a burnup calcula- (3) By a forward calculation using a core analysis
tion of a typical boiling water reactor (BWR) fuel as- code system with each of sampled libraries, a
sembly. Finally, concluding remarks are summarized in target parameter xi (e.g. neutron multiplication
Section 4. factor, reactivity worth, or reaction-rate ratio) is
evaluated for each of corresponding ith libraries.
2. Methodology Consequently, N sets of target parameters are
randomly sampled.
2.1. Random sampling method (4) The uncertainty of target parameter due to co-
In this section, the RS method for core analysis is variance data can be estimated by standard de-
briefly explained as an example utilizing covariance data viation sx for the sampled target parameters xi ,
of microscopic cross-section data. Other uncertainties of i.e.
input parameters, e.g. fabrication tolerances for geome- 
tries and material densities, can be similarly treated.  ⎛ ⎞2

1 ⎝ 1  ⎠
N N

sx =  xi − xj . (5)
(1) The covariance matrix for multi-group micro- N−1 N
i =1 j =1
scopic cross-section of each nuclide is generated
using the NJOY code [15]. It is noted that the
energy group structure of the covariance matrix If N is sufficiently large, it is expected that sx can
should be consistent with that of a cross-section be estimated with statistically satisfactory preci-
library for a lattice physics code. sion.
(2) A multi energy-group microscopic cross-section
library is generated for a lattice physics code, As described above, the UQ using the RS method
based on a multivariate normal distribution with is simple and practical procedure without any (gener-
the covariance matrix generated in Step (1). alized) adjoint calculations. Furthermore, the merit of
Namely, we have the RS method is high affinity to the burnup and the
thermal-hydraulics feedback effects in the core analysis
σ = μ
 + Az, (1) [10].
 Journal of Nuclear Science and Technology, Volume 52, Nos. 7–8, July–August 2015 995

2.2. Bootstrap method normal distribution ( p is a probability, 0 < p <

In this section, the confidence interval estimation us- 1).
ing the bootstrap method for the uncertainty of target (7) Furthermore, the acceleration factor a is esti-
parameter sx is explained. Although there are several mated by the following Jackknife method:
methods to estimate the confidence interval, the boot-  3
N (i ) N ( j)
strap bias-corrected and accelerated (BCa) method [13]
= 1 sx −
1
i N j = 1 sx
is utilized in the present study. In the BCa method, by a=−  ,
adjusting the percentile points for both bias and skew-  N  (i ) N 2  32
( j)
i = 1 sx −
1
6 N j = 1 sx
ness in the bootstrap distribution, the accuracy of esti-
mated confidence interval can be improved with reason- (8)
able computational costs. The procedures are explained
below: (i )
where sx is a standard deviation esti-
mated from a dataset of target parameters
(1) An original dataset of target parameters X = {x1 , . . . , xi −1 , xi +1 , xN } which is resampled by
{x1 , x2 , . . . , xN } is provided by the RS method eliminating xi from the original dataset X.
as described in Section 2.1. (8) Let the confidence level be (1 − α). For exam-
(2) Using Equation (5), the sample estimate of stan- ple, in the case of 95% confidence level, α = 0.05.
dard deviation sx is estimated from X. Then, by utilizing the two factors w and a esti-
(3) By RS with replacement from the original mated in Steps (6) and (7), corrected percentile
dataset X, a resample (so-called “bootstrap points α̂ L and α̂U are estimated for lower and up-
sample”) is newly generated as X∗ = {x1∗ , x2∗ , per limits of the confidence interval:
∗
. . . , xN }, where the superscript ∗ represents the
bootstrap sample. In the case of non-parametric  
w + −1 (α/2)
bootstrap method, each element xi∗ is selected α̂ L =  w +   , (9)
from all elements of X, using an uniform random 1 − a w + −1 (α/2)
number within the range of [1, N]. For exam-
ple, if X = {1, 2, 3, 4, 5, 6}, xi∗ may be redun-  
dantly selected such as X∗ = {2, 1, 2, 5, 6, 1}. w + −1 (1 − α/2)
α̂U =  w +   ,
(4) The bootstrap samples X∗b are repeatedly re- 1 − a w + −1 (1 − α/2)
sampled (b = 1, 2, . . . , B), where B is the total
(10)
number of bootstrap samples. Typically, in or-
der to adequately estimate the confidence inter-
where (x) is the cumulative distribution func-
val, the order of B is recommended to be several
tion of standard normal distribution. In the
thousands. Consequently, B bootstrap samples
case of 95% confidence level, −1 (α/2) ≈ −1.96
X∗1 , X∗2 , . . . , X∗B are provided.
and −1 (1 − α/2) ≈ +1.96 in Equations (9) and
(5) For each bootstrap sample X∗b generated in
(10), respectively.
Step (4), the “bootstrap replicate” of standard
(9) The B bootstrap replicates sx∗b are sorted in as-
deviation sx∗b is evaluated as follows:
cending order. Let p be a positive real num-
 ber within the range of 0 < p < 1 and the p Bth
 ⎛ ⎞2
 smallest values in the sorted sx∗b are denoted as
1  ⎝ ∗b 1  ∗b ⎠
N N

sx∗b = xi − xj . (6) sx
∗( p B)
. Finally, the lower and upper limits of con-
N−1 N
i =1 j =1 fidence interval for the standard deviation sx , de-
∗ ∗
noted as sx,L and sx,U , are estimated as follows:
Consequently, B bootstrap replicates sx∗1 , sx∗2 ,
. . . , sx∗B are obtained to estimate the confidence ∗
sx,L ≈ sx∗(α̂L B) , (11)
interval. By using sx∗b , the confidence interval of
sx can be practically estimated.
∗
(6) In the BCa method, the bias-correction factor w sx,U ≈ sx∗(α̂U B) . (12)
is estimated as follows:

It is noted that if the value of p B is not integer,
−1 1  ∗b  ∗( p B)
sx is numerically calculated by the linear in-
w= # sx ≤ sx , b = 1, . . . , B ,
B terpolation of two adjacent neighboring points
(7) in the sorted sx∗b .

where #{sx∗b ≤ sx , b = 1, . . . , B} means the As explained above, using the bootstrap method,
number of sx∗b , satisfying the condition of sx∗b ≤ limited number of samples can be effectively utilized to
sx , and −1 ( p) is a quantile function of standard estimate the confidence interval of uncertainty without
996 T. Endo et al.

further sampling. The above-mentioned procedure is 0.009

Standard deviation of kinf at 30GWd/t

within 95% CI
a specific example to estimate the confidence interval out of 95% CI
for the standard deviation of target parameters. Note reference

that the bootstrap method has capability to estimate

0.006
confidence intervals for other sample estimates: not only
the variance for one kind of target parameter but also
the covariance (or the correlation coefficient) between
two different kinds of target parameters. Moreover, 0.003
even in the case of the cross-section adjustment method 0 50 100 150 200
m-th trial
using the RS method [11], the bootstrap method is
applicable to estimate the confidence intervals for the Figure 1. 95% bootstrap confidence intervals for standard
adjusted cross-section and covariance data. deviations of kinf (30 GWd/t assembly-average burnup).

3. Numerical analysis (5) The procedures of Step (4) were carried out 200
3.1. Verification of the bootstrap method times, thereby 200 bootstrap confidence intervals
were obtained.
In order to verify the confidence interval estimated
(6) Using Step (5), we estimated a coverage prob-
by the bootstrap method (i.e. the BCa method), the un-
ability ps that the reference standard deviation
certainty of neutron multiplication factor is investigated
sk,ref was contained within the estimated confi-
in the burnup calculation using a lattice physics code. ∗ ∗
dence interval [sk,L , sk,U ]. If the confidence in-
The verification procedures are given as follows:
terval estimated by the BCa method is adequate,
(1) Using CASMO-4 as a lattice physics code [16], the value of ps should be close to 95%.
the burnup calculation was carried out for a sim-
plified BWR 9 × 9 type-B fuel assembly. For
As an example, Figure 1 shows the 95% bootstrap
simplicity, the 235 U enrichment was 3.7wt% for ∗ ∗
confidence intervals [sk,L , sk,U ] at 30 GWd/t assembly-
all fuel pins and there were no burnable poison
average burnup. In Figure 1, each plot point of +
fuel rods.
and × corresponds to the values of sk (50) obtained by
(2) The covariance data of JENDL-4.0 were used as
Equation (5); the gray bars (marked + ) indicate that
uncertainties in microscopic cross-section data
the confidence interval contains the reference value sk,ref ;
[3,4]. The uncertainties of nuclides were treated
on the other hand, the black bars (marked × ) indicate
for 18 heavy nuclides: 234 U, 235 U, 236 U, 238 U,
237 that sk,ref exists outside of the intervals. In the case of 30
Np, 238 Pu, 239 Pu, 240 Pu, 241 Pu, 242 Pu, 241 Am,
242 GWd/t, the coverage probability ps is (200 − 11)/200 ≈
Am, 243 Am, 242 Cm, 243 Cm, 244 Cm, 245 Cm and
246 94.5% and nearly equal to the confidence level (95%).
Cm. In addition, their capture, fission, elastic
Figure 2 shows the variation of ps with respect to
scattering microscopic cross-section (σc , σf , σs ),
the assembly-average burnup. As shown in Figure 2,
and number of neutrons emitted per fission (ν̄)
ps ranges from 91% to 96%. According to the above-
were randomly sampled on the basis of multi-
mentioned results, it is confirmed that the confidence
variate normal distribution with their covariance
data (see Section 2.1).
(3) CASMO-4 burnup calculations were carried out 100%
for 10,000 times with randomly sampled 10,000
libraries using the RS method. Then the uncer-
tainties (standard deviations) of infinite neutron 98%
multiplication factor kinf were estimated for all
burnup steps. Because of the adequately large 96%
number of sampling, the estimated uncertainties
ps [-]

were regarded as the reference values (denoted

as sk,ref ) which would be close to the true values. 94%
(4) CASMO-4 burnup calculations were carried out
for 50 times with 50 libraries which were ran- 92%
domly chosen from the original 10,000 libraries.
Then the uncertainties of kinf (denoted as sk (50))
90%
were estimated in the same way. Furthermore, 0 10 20 30 40 50
using the BCa method, their 95% confidence in-
∗ ∗ Assembly-average burnup [GWd/t]
tervals [sk,L , sk,U ] were evaluated as described in
Section 2.2. The total number of the bootstrap Figure 2. Variation of coverage probability ps with respect to
samples B was 10,000. the assembly-average burnup.
 Journal of Nuclear Science and Technology, Volume 52, Nos. 7–8, July–August 2015 997

0.010 0.010

0.008 0.008

Standard deviation of kinf

Standard deviation of kinf 0.006 0.006

0.004 0.004

0.002 0.002

0.000 0.000
10 100 1000 10000 10 100 1000 10000

Number of sampled libraries Number of sampled libraries

(a) Bootstrap method (b) Assumption of normality

Figure 3. Variation of 95% confidence intervals with respect to the number of randomly sampled libraries (30 GWd/t assembly-
average burnup).

interval for the uncertainty (standard deviation) of kinf n ( p is a probability, 0 < p < 1); and the con-
can be reasonably estimated using the BCa method. fidence level corresponds to (1 − α). For exam-
ple, if (1 − α) = 95% and N = 50, X−1 49 (0.975) ≈
70.222 and X−1 49 (0.025) ≈ 31.555 in Equations
3.2. Comparison with the confidence interval (13) and (14), respectively.
under the assumption of normality
(4) Steps (1), (2) and (3) were carried out from N =
In this section, the bootstrap confidence interval is 10 to N = 10,000.
compared with another confidence interval, which is es-
timated under the assumption of normality. The com- As an example, Figure 3 shows the variation of 95%
∗ ∗
parison procedures are as follows: confidence intervals [sk,L , sk,U N
] and [sk,L , sk,U
N
] with re-
spect to the number of randomly sampled libraries N,
(1) N microscopic cross-section libraries were ran- at the assembly-average burnup of 30 GWd/t. For com-
domly sampled without replacement from the parison, Figure 4 shows both widths of confidence in-
∗ ∗
original 10,000 libraries obtained in the verifi- tervals, i.e. (sk,U − sk,L N
) and (sk,U − sk,L
N
). As shown in
cation of Section 3.1. By the CASMO-4 burnup Figures 3 and 4, the width of confidence interval be-
calculations with these sampled libraries, the un- comes narrower as the number of sampled libraries N
∗ ∗
certainties (standard deviations) of kinf (denoted increases. Approximately, both widths of (sk,U − sk,L )
as sk (N)) were estimated in the same way as de-
scribed in Section 3.1.
(2) Using the BCa method for the result of Step (1), bootstrap method assumption of normality
∗ ∗
the 95% confidence intervals [sk,L , sk,U ] were 0.0100
evaluated. The total number of the bootstrap
Width of 95% confidence interval

samples B was 10,000.

(3) Now, let us assume that a population of kinf fol-
lows a normal distribution. Under this assump-
tion, the confidence intervals of sk (N) (denoted
N
as [sk,L , sk,U
N
]) can be approximately estimated
0.0010
by using chi-squared distribution:

N−1
N
sk,L ≈ sk (N) , (13)
X−1
N−1 (1 − α/2)

 0.0001
N−1
N
sk,U ≈ sk (N) , (14) 10 100 1000 10000
X−1
N−1 (α/2)
Number of sampled libraries

where X−1
n ( p) is the quantile function of chi- Figure 4. Comparison of widths of 95% confidence intervals
squared distribution with a degree of freedom (30 GWd/t assembly-average burnup).
998 T. Endo et al.
1.05
of hypothesis tests are needed to assess the normality.
On the other hand, the bootstrap method has an
advantage to simply estimate the confidence intervals
Infinite neutron multiplication factor

1.04

without the assumption of normality. As discussed in

Section 3.1, the confidence intervals can be reasonably
estimated using the bootstrap method, without any
1.03

normality tests. In more general cases, input parameters

(e.g. fabrication tolerances) may follow non-normal
distributions (e.g. uniform distribution). In such cases,
1.02

although target parameters (e.g. nuclide number den-

sities) do not necessarily follow normal distributions,
it is expected that the bootstrap method is applicable
1.01

to estimate the confidence intervals of uncertainties

without complicated mathematical procedures.
1.00

-4 -2 0 2 4 4. Conclusion
The RS method is one of the simple and practical
Theoretical quantile
UQ techniques. It is expected that the RS method has an
Figure 5. Normal quantile–quantile plot of kinf (30 GWd/t advantage for complicated numerical analysis, such as
assembly-average burnup). an LWR analysis which is a multi-step calculation with
burnup and thermal-hydraulics feedback effects. Note
that statistical errors are inevitably included in the es-
timated uncertainties in the RS method. Therefore, the
and (sk,UN
− sk,L
N
) decrease inversely proportional to confidence interval for estimated uncertainties is useful
√ information to determine the number of RS of input pa-
N, as suggested in the double-logarithmic graph of
rameters. In the present paper, the applicability of the
Figure 4. In addition, the bootstrap confidence intervals
bootstrap method, specifically the BCa method, was in-
are nearly equal to those under the assumption of nor-
∗ ∗ vestigated to evaluate the confidence interval of uncer-
mality, although the widths of (sk,U − sk,L ) are slightly
tainty estimated by the RS method. The merit of the
narrower than those of (sk,U − sk,L ).
N N
bootstrap method is to easily estimate the confidence in-
Figure 5 shows a normal quantile–quantile plot of terval of uncertainty only using limited number of sam-
kinf , which consists of 10,000 results obtained from ples, without further sampling.
all microscopic cross-section libraries. As shown in Through the burnup calculation for a BWR fuel as-
Figure 5, the linearity of plot points suggests that the sembly, it is confirmed that the bootstrap confidence
frequency distribution of kinf is well approximated by a interval is a reasonable and practical estimator for the
normal distribution. According to the Jarque–Bera test uncertainty (standard deviation) of neuron multiplica-
[17] for results of Figure 5, the null hypothesis of nor- tion factor. In the present problem, the distribution of
mality is not rejected, since the p-value is 0.28 and larger neutron multiplication factor is well approximated by
than the typical significance level 5%. Therefore, the as- a normal distribution. In such a case, based on the as-
sumption of normality is reasonable for kinf in this bur- sumption of normality, the confidence interval can be
nup calculation problem. Consequently, it is considered also estimated by the aid of quantile function of chi-
N
that the confidence intervals [sk,L , sk,U
N
] can be reason- squared distribution. On the other hand, the merit of
ably estimated in this case. the bootstrap method is to simply estimate the con-
In the case of an actual UQ procedure using the RS fidence interval without the assumption of normality;
method, the number of core analyses with randomly thus, it is expected that the bootstrap method is ap-
sampled library gradually increases until the statistical plicable to estimate the confidence intervals of uncer-
dispersions of estimated uncertainties are acceptably tainties without complicated mathematical procedures
small. As shown in Figures 3 and 4, the number of RS in the case of non-normal distribution. Furthermore, the
can be reasonably determined only from the calculated bootstrap method has capability to estimate confidence
results themselves without additional core analysis, intervals for other statistical estimates, e.g. covariance
by judging whether the estimated confidence interval or correlation coefficient between two different kinds of
is acceptable. A merit of the estimation method by target parameters. One of future tasks is further verifi-
assumption of normality is less calculation cost than cation of bootstrap method for the non-normal distri-
the bootstrap method and the confidence interval can bution and the other statistical estimates, which was not
be deterministically evaluated by Equations (13) and fully investigated in the present paper. The results of the
(14). It is noted that, however, Equations (13) and (14) present study indicate that the application of the boot-
are appropriate if the normal distribution model holds strap method for UQ of reactor physics calculations is
true for the target parameters; therefore, some kinds promising.
 Journal of Nuclear Science and Technology, Volume 52, Nos. 7–8, July–August 2015 999

Acknowledgements Mughabghab SF, Nobre GPA, Palmiotti G, Palumbo A,

We sincerely appreciate Nuclear Fuel Industries, Ltd. for Pigni MT, Pronyaev VG, Sayer RO, Sonzogni AA, Sum-
the technical supports for the use of randomly sampled micro- mers NC, Talou P, Thompson IJ, Trkov A, Vogt RL, van
scopic cross-section libraries. This work was supported in part der Marck SC, Wallner A, White MC. ENDF/B-VII.1
by JSPS KAKENHI, Grant-in-Aid for Scientific Research (C) nuclear data for science and technology: cross sections,
(24561040). covariances, fission product yields and decay data. Nucl
Data Sheets. 2011;112:2887–2996.
[6] Koning AJ, Rochman D. Modern nuclear data evalu-
ation with the TALYS code system. Nucl Data Sheets.
Disclosure statement 2012;113:2841–2934.
No potential conflict of interest was reported by the [7] Hara A, Takeda T, Kikuchi Y. SAGEP: Two-
authors. dimensional sensitivity analysis code based on gen-
eralized perturbation theory. Ibaraki (Japan): Japan
ORCID Atomic Energy Agency; 1984 (JAERI-M 84-027).
[8] Chiba G, Tsuji M, Narabayashi T. Uncertainty quantifi-
Tomohiro Endo http://orcid.org/0000-0002-6644-4398
cation of neutronic parameters of light water reactor fuel
cells with JENDL-4.0 covariance data. J Nucl Sci Tech-
nol. 2013;50:751–760.
References [9] Kato S, Endo T, Yamamoto A, Yamauchi H, Kimura Y.
[1] Yamamoto A, Yasue Y, Endo T, Kodama Y, Ohoka Y, Random sampling-based cross-section adjustment tech-
Tatsumi M. Uncertainty estimation of core safety pa- nique for LWR core analysis. Proceedings of ICAPP
rameters using cross-correlations of covariance matrix. 2013; 2013 Apr 14–18; Jeju Island (Korea). [CD-ROM].
J Nucl Sci Technol. 2013;50:966–978. [10] Watanabe T, Endo T, Yamamoto A, Ohoka Y, Kodama
[2] Park HJ, Lee DH, Shim HJ, Kim CH. Uncertainty Y, Ushio T. Uncertainty and correlation estimation of
propagation analysis for Yonggwang nuclear unit 4 by reload safety parameters of PWR using random sam-
McCARD/MASTER core analysis system. Nucl Eng pling method. Trans Am Nucl Soc. 2013;109:1365–1368.
Technol. 2014;46:291–298. [11] Watanabe T, Endo T, Yamamoto A, Kodama Y, Ohoka
[3] Shibata K, Iwamoto O, Nakagawa T, Iwamoto N, Ichi- Y, Ushio T. Cross section adjustment method based
hara A, Kunieda S, Chiba S, Furutaka K, Otuka N, on random sampling technique. J Nucl Sci Technol.
Ohsawa T, Murata T, Matsunobu H, Zukeran A, Ka- 2014;51:590–599.
mada S, Katakura J. JENDL-4.0: a new library for [12] Efron B. Bootstrap methods: another look at the Jack-
nuclear science and engineering. J Nucl Sci Technol. knife. Ann Stat. 1979;7:1–26.
2011;48:1–30. [13] Efron B. Better bootstrap confidence intervals. J Am Stat
[4] JENDL-4.0u [Internet]. Japan: Japan Atomic Energy Assoc. 1987;82:171–185.
Agency; 2012 Sep 14; [updated 2015 Jan 27; cited [14] Efron B, Tibshirani RJ. An introduction to the boot-
2015 Mar 18]. Available from: http://wwwndc.jaea. strap. New York: Chapman & Hall; 1993.
go.jp/jendl/j40/update/. [15] Macfarlane RE, Muir DW. The NJOY nuclear data
[5] Chadwick MB, Herman M, Obložinský P, Dunn ME, processing system version 91. Los Alamos (NM): Los
Danon Y, Kahler AC, Smith DL, Pritychenko B, Ar- Alamos National Laboratory; 1994 (LA-12740-M).
banas G, Arcilla R, Brewer R, Brown DA, Capote [16] CASMO-4 A Fuel Assembly Burn-up Program. User’s
R, Carlson AD, Cho YS, Derrien H, Guber K, Hale manual. Waltham (MA): Studsvik Scandpower Inc.;
GM, Hoblit S, Holloway S, Johnson TD, Kawano T, 2009 (SSP-09/443-U Rev 0).
Kiedrowski BC, Kim H, Kunieda S, Larson NM, Leal [17] Jarque CM, Bera AK. A test for normality of observa-
L, Lestone JP, Little RC, McCutchan EA, MacFar- tions and regression residuals. Int Stat Rev. 1987;55:163–
lane RE, MacInnes M, Mattoon CM, McKnight RD, 172.