Probabilistic Error Bounds for Reduced Order Modeling M&C2015

ANS MC2015 - Joint International Conference on Mathematics and Computation (M&C), Supercomputing in Nuclear Applications (SNA) and the Monte
Carlo (MC) Method • Nashville, TN • April 19-23, 2015, on CD-ROM, American Nuclear Society, LaGrange Park, IL (2015)
Probabilistic Error Bounds for Reduced Order Modeling
Mohammad G. Abdo and Hany S. Abdel-Khalik
School of Nuclear Engineering, Purdue University
400 Central Drive, Purdue Campus, NUCL Bldg., West Lafayette, IN 47906
abdo@purdue.edu; abdelkhalik@purdue.edu
ABSTRACT
Reduced order modeling has proven to be an effective tool when repeated execution of reactor
analysis codes is required. ROM operates on the assumption that the intrinsic dimensionality of the
associated reactor physics models is sufficiently small when compared to the nominal
dimensionality of the input and output data streams. By employing a truncation technique with roots
in linear algebra matrix decomposition theory, ROM effectively discards all components of the input
and output data that have negligible impact on reactor attributes of interest. This manuscript
introduces a mathematical approach to quantify the errors resulting from the discarded ROM
components. As supported by numerical experiments, the introduced analysis proves that the
contribution of the discarded components could be upper-bounded with an overwhelmingly high
probability. The reverse of this statement implies that the ROM algorithm can self-adapt to
determine the level of the reduction needed such that the maximum resulting reduction error is below
a given tolerance limit that is set by the user.
Key Words: Reduced order modeling, Error bounds
1 INTRODUCTION
Recently, there has been an increased interest in reduced order modeling algorithms for reactor
physics simulation. This is primarily driven by the best-estimates plus uncertainty (BEPU)
approach, first championed by the industry until its adoption into law by the US-NRC in 1988. To
fully realize the benefits of the BEPU approach, the uncertainties of the simulation predictions
must be properly characterized. Uncertainty characterization (UC) implies the capabilities to
identify, quantify, and prioritize the various sources of uncertainties. These three capabilities
require repeated model executions which proves to be an increasingly taxing endeavor, especially
with the continuous increase in the modeling details sought to improve fidelity.
Reduced order modeling is premised on the observation that the true dimensionality of reactor
physics simulation codes is rather small, implying that the associated uncertainty sources that
affect model behavior must also be rather small notwithstanding their nominal number is very
large. With a small number of uncertainty sources, uncertainty characterization becomes a
computationally tractable practice. This follows because the computational cost of performing UC
depends on the number of uncertainty sources, which absent reduction could number in the
millions for typical reactor physics simulation.
2 ROM ERROR BOUND CONSTRUCTION
To describe the contribution of this manuscript, basic definition of ROM is first introduced.
Consider a model of reactor physics simulation of the form:

Authors’ names, use et al. if more than 3
Page 2 of 9
 y f x (1)
where
n
x  are reactor physics parameters, e.g., cross-sections, m
y  are reactor responses of
interest e.g., eigenvalue, peak clad temperature, etc., and n and m are the numbers of parameters
and responses, respectively. The simulation, represented by the function f, is assumed to be a black
box. The goal of any ROM approach is replace the original simulation with an approximate
representation f that can be used in lieu of the original simulation for computationally intensive
analyses such as UC. To ensure reliability of the ROM approximation f , the following criterion
must be satisfied:
   f x f x   for all x S (2)
where  is a to-be-determined upper-bound, and S defines the region of applicability. If such
upper-bound exists, one can adjust the level of reduction to ensure that the bound matches the
confidence one has in the original simulation predictions. In such case, both f and f would provide
the same level of confidence for any subsequent analysis.
In our analysis, the ROM approximation f has the general form:
   f x f x Ν Κ
where both N and K are rank-deficient matrix operators such that:
m m
N ,   dim R mrN , and min( , )mr m n ,
n n
K ,   dim R nrK , and min( , )nr m n .
These matrices identify active subspaces in the space of input parameters and output responses.
The implication is that few degrees of freedom in the input space are needed to capture all possible
model variations, and the output responses have only a small number of degrees of freedom as
well, implying large degrees of correlation exist therein. This knowledge allows one to craft
uncertainty quantification and sensitivity analysis techniques in such a manner that reduces the
required number of forward and/or adjoint model executions necessary to complete the respective
analyses. See earlier work for more details on these approaches [1, 2].
In practice, the error operator is inaccessible but can be sampled and aggregated in a matrix E
whose th
ij element represents the error in the th
i response of the th
j sample, written as:
     
 
,: ,: ( )
[ ]
T T
i j y y i x x j
ij
i j
f x i i f x
f x


Q Q Q Q
E (3)
The matrix E calculates the discarded component of function f. Each row of E represents a
response, implying that if one treats each row as a matrix, it is possible to calculate a different error
bound for each response. This allows one to compute the individual responses’ errors since each
response is expected to have its own reduction error. To achieve that: consider a matrix m N
E
and a random vector N
w where N is the number of sampled responses such that wi ,

Short version of title as entered by author on web page
Page 3 of 9
where is binomial distribution with a probability of success of 0.9. Then w can be used to
estimate the largest and smallest eigenvalue and hence the 2-norm of E via:
 
   
2
2
1
1
1,2,
0
max 1 ,
s
i
wi s
w pdf t dt



 
 
   
 
 
E E (4)
where the multiplier 1 w  is numerically evaluated to be 1.0164 for binomial distribution and
for a success probability of 0.9. For more details about this approach and the proof of Eq. (4), the
reader may consult the following references [3, 4, 5, 6].
The main goal of this paper is to show that one can satisfy Eq. (4) for any given N and K matrices
with an overwhelmingly high probability. Computing an error bound for general reduction
operators is important because in general multi-physics models, one may obtain a reduction
operator using a lower-fidelity model when the high fidelity model is too expensive to execute in
search of the ROM active subspace. Another situation occurs when the input for a given physics
model is produced by another physics model. In such case, one could use the forward model
executions of the upstream physics model to calculate an active subspace for the downstream
physics. Therefore, it is important to capture the reduction errors for general matrix reduction
operators.
If the distribution of w and the multiplier  are selected such that the integral on the right hand
side is 0.1, the probability that the estimated bound is larger than the 2-norm of the error is given
by: 1 10 s
p 
  , where s is a small integer that corresponds to an additional number of matrix-
vector multiplications. Typically, we employ a value of s equal to 10 to ensure extremely high
probability. In support of verifying the proposed algorithm however, this manuscript will employ
s=1 to give rise to situations when the estimated error bound fails to bound the actual errors with
probability of 10%. Multiple numerical experiments will be devised to test the upper-bound and
the probability of failure as predicted by the theory.
3 NUMERICAL EXPERIMENTS AND RESULTS
This section will employ a number of experiments to demonstrate the ability to calculate an upper-
bound on the reduction error. The first experiment will focus on a direct ROM application to
identify the active subspace and calculate the associated reduction error and probability of failure.
The second experiment will employ the reduction operators determined using a given set of
conditions (low burnup, hot full power) to test its adequacy for other conditions (higher burnup
and cold conditions). This capability will prove useful in model validation activities relying on the
use of the proposed reduction techniques, where now one must determine whether the developed
reduced model will be adequate for a wide range of operating conditions.

Page 4 of 9
3.1 Case Study 1:
This case study employs a pin cell depleted to (3.0 GWd/MTU) as the reference model used to
identify the active subspaces for the parameters and the responses spaces. SCALE 6.1 is used for
the computational purposes, sequences like t-depl, t-newt, tsunami-2d and SAMS 5.0 are needed
for the depletion, neutronics calculations and sensitivity analysis respectively [7]. The original
parameter space contains 7 nuclides * 2 reactions * 238 energy groups=3332 parameters, whereas
the nominal dimension for the response space is 238 representing material flux at 238 energy
groups. For illustration, a very small rank is assumed to ensure that the actual errors are large
enough to possibly fail the theoretical error bound proposed here. In all the tests, a value of 1s 
is employed to maximize the number of failures for the sake of demonstration. In the series of
figures below, the actual probability of failure is indicated on the top of the left graphs.
Figs. 1 through 4 display the results of the first case study. The same responses are employed for
both case studies. In the odd-numbered figures, the response is the total collision rate in the energy
range 1.85 to 2.35 MeV. The even figures show the same response but in the thermal range between
0.975 and 1.0 eV. We use these small ranges to depict the power of the reduction in capturing
localized responses. In each of the figures, the left graph compares the actual error resulting from
the reduction to the error bound calculated from Eq. (4). The 45-degree solid line indicates the
limit of the failure region, i.e., when the actual error exceeds the bound predicted by the theory.
The right graphs show the actual variation of the response due to a random perturbation of 30% in
the parameters.
Figure 1. Fast Collision Rate Errors – Parameter Reduction Only

Page 5 of 9
Figure 2. Thermal Collision Rate Errors – Parameter Reduction Only
Figs. 1 and 2 show a parameter-only-based reduction, meaning that the reduction is rendered at
the parameter space. Both the level of reduction in terms of the rank of the active parameter
subspace rx, and the actual probability of failure are shown on the top of the right graph. Reader
should remember that we picked s and the multiplier in equation (4) such that the probability of
success is 0.9. In reality s is picked to be 5 which results in a probability of success of 99.999%.
Figs. 3 and 4 employ response-based reduction only, implying no reduction in the parameter space.
The rank of response active subspace ry is indicated in a similar manner.
Figure 3. Fast Collision Rate Errors – Response Reduction Only

Page 6 of 9
Figure 4. Thermal Collision Rate Errors – Response Reduction Only
Notice that the reduction errors calculated will depend on whether the parameter-based reduction
captures the important parameter directions that control the model response variations. Moreover,
the response reduction, if not captured correctly, will miss directions along which the response is
expected to vary. This situation will be clearer when we consider different physics conditions as
done in the next case study.
3.2 Case Study 2:
This case study employs the active subspaces extracted from the previous reference model to
predict the response variations at different physics conditions. We employ a 24 GWd/MTU
depleted fuel simulated at cold conditions. This emulates the effect of starting up a reactor with a
once-burned fuel.
Figs. 5 through 8 correspond respectively to Figs. 1 through 4, where now the model is being
evaluated at different physics conditions, using the reduction results from the previous case study,
i.e., same ranks for parameter and response spaces, same responses, and same size of parameter
perturbations. The idea here is to check whether the model reduced at hot conditions could be
employed at sufficiently different physics conditions.
Figure 5. Fast Collision Rate Errors - Parameter Reduction Only

Page 7 of 9
Figure 6. Thermal Collision Rate Errors - Parameter Reduction Only
Figs. 5 and 6 show that the actual errors and the predicted bounds due to the parameter reduction
are slightly higher than the errors in figs. 1 and 2. This indicates that the active subspace extracted
using the reference model have approximately the same level of accuracy at new physics
conditions.
Figs. 7 and 8 behave in a different fashion, where now results indicate that the actual errors and
their bounds have noticeably increased beyond those in Figs. 3 and 4. This indicates that the
responses at the new physics conditions are changing along new directions in the response space
that are not captured by the reference physics models. Also, notice that in all cases, the actual
probability of failure is always less than the theoretical value of 1-10-s
. Smaller number of
failures were observed in Figs. 7 and 8; the reason for this remains to be investigated. We recall
here that the failure probability is chosen to be 10% which is extremely high. In reality, the
failure probability is set to be extremely small to ensure that its dependence on core conditions is
negligible.
Figure 7. Fast Collision Rate Errors - Response Reduction Only

Page 8 of 9
Figure 8. Thermal Collision Rate Errors - Response Reduction Only
4 CONCLUSIONS:
This manuscript has investigated the ability of ROM techniques to upper-bound the error
resulting from the reduction. This is an important characteristic for any ROM to ensure reliability
of the reduced model for subsequent engineering analyses, such as uncertainty and sensitivity
analysis. More importantly, this summary has shown a practical way by which the ROM errors
can be evaluated for general reduction operators. This is invaluable when dealing with high
fidelity codes that can only be executed few times, and it is difficult to extract their active
subspaces. Another important application of this work is the reduction of multi-physics models,
where the active subspace generated by one physics model is used as the basis for reducing the
input space for another physics model.
5 ACKNOWLEDGEMENTS:
The first author would like to acknowledge the support received from the Department of Nuclear
Engineering at North Carolina State University to complete this work in support of his PhD.
6 REFERENCES:
1. Youngsuk Bang, Jason Hite, And Hany S. Abdel-Khalik, “Hybrid reduced order modeling
applied to nonlinear models,” IJNME; 91: pp.929–949 (2012).
2. Hany S. Abdel-Khalik, Et Al., “Overview of Hybrid Subspace Methods for Uncertainty
Quantification and Sensitivity Analysis,” Annals of Nuclear Energy, Vol. 52, February, (2013):
Pages 28–46. A Tutorial on Applications of dimensionality reduction and function
approximation.
3. S. S. Wilks, Mathematical statistics, John Wiley, New York, 1st ed. 1962.
4. John D. Dixon, “Estimating extremal eigenvalues and condition numbers of matrices,” SIAM
1983; 20(2): 812–814.

Page 9 of 9
5. Mohammad G. Abdo, and Hany S. Abdel-Khalik, “Propagation of error bounds due to Active
Subspace reduction,” Transactions of the American Nuclear Society, Reno, NV, Vol. 110,
pp.196-199 (2014).
6. Mohammad G. Abdo, and Hany S. Abdel-Khalik, “Further investigation of error bounds for
reduced order modeling,” submitted to ANS MC2015: Joint International Conference on
Mathematics and Computation (M&C), Super Computing in Nuclear Applications (SNA), and
the Monte Carlo (MC) method, Nashville, TN, April, 19-23, 2015.
7. SCALE: A Comprehensive Modeling and Simulation Suit for Safety Analysis and Design,
ORNL/TM- 2005/39, Version 6.1, Oak Ridge National Laboratory, Oak Ridge, Tennessee,
2011.

Probabilistic Error Bounds for Reduced Order Modeling M&C2015

More Related Content

What's hot (18)

Viewers also liked (10)

Similar to Probabilistic Error Bounds for Reduced Order Modeling M&C2015 (20)

Recently uploaded (20)

Probabilistic Error Bounds for Reduced Order Modeling M&C2015