Download as pdf or txt
Download as pdf or txt
You are on page 1of 19

PUBLICATIONS

Water Resources Research


RESEARCH ARTICLE An improved multilevel Monte Carlo method for estimating
10.1002/2016WR019475
probability distribution functions in stochastic oil
Key Points: reservoir simulations
 Develop an improved multilevel
Monte Carlo method Dan Lu 1, Guannan Zhang 2, Clayton Webster 2,3, and Charlotte Barbier 4
 Apply the method to an oil reservoir
model 1
Computer Science and Mathematics Division, Climate Change Science Institute, Oak Ridge National Laboratory, Oak
 Evaluate the effectiveness and
Ridge, Tennessee, USA, 2Department of Computational and Applied Mathematics, Oak Ridge National Laboratory, Oak
efficiency of the proposed method
Ridge, Tennessee, USA, 3Department of Mathematics, University of Tennessee, Knoxville, Tennessee, USA, 4Computational
Sciences and Engineering Division, Oak Ridge National Laboratory, Oak Ridge, Tennessee, USA
Correspondence to:
D. Lu,
lud1@ornl.gov
Abstract In this work, we develop an improved multilevel Monte Carlo (MLMC) method for estimating
cumulative distribution functions (CDFs) of a quantity of interest, coming from numerical approximation
Citation:
of large-scale stochastic subsurface simulations. Compared with Monte Carlo (MC) methods, that require
Lu, D., G. Zhang, C. Webster, and
C. Barbier (2016), An improved a significantly large number of high-fidelity model executions to achieve a prescribed accuracy when
multilevel Monte Carlo method for computing statistical expectations, MLMC methods were originally proposed to significantly reduce the
estimating probability distribution
computational cost with the use of multifidelity approximations. The improved performance of the MLMC
functions in stochastic oil reservoir
simulations, Water Resour. Res., 52, methods depends strongly on the decay of the variance of the integrand as the level increases. However,
9642–9660, doi:10.1002/ the main challenge in estimating CDFs is that the integrand is a discontinuous indicator function whose
2016WR019475.
variance decays slowly. To address this difficult task, we approximate the integrand using a smoothing
function that accelerates the decay of the variance. In addition, we design a novel a posteriori optimiza-
Received 7 JUL 2016
tion strategy to calibrate the smoothing function, so as to balance the computational gain and the
Accepted 17 NOV 2016
Accepted article online 21 NOV 2016 approximation error. The combined proposed techniques are integrated into a very general and practical
Published online 30 DEC 2016 algorithm that can be applied to a wide range of subsurface problems for high-dimensional uncertainty
quantification, such as a fine-grid oil reservoir model considered in this effort. The numerical results reveal
that with the use of the calibrated smoothing function, the improved MLMC technique significantly
reduces the computational complexity compared to the standard MC approach. Finally, we discuss sever-
al factors that affect the performance of the MLMC method and provide guidance for effective and effi-
cient usage in practice.

1. Introduction
Subsurface flow and transport models are routinely employed in prediction of fluid behavior, and in making
economic and management decisions. These models describe fluid flow, e.g., oil, gas, and water, through a
porous medium, so the porous media properties have a significant impact on the model outputs such as
mass flow rate and pressure. Almost all the porous media properties used in subsurface flow simulation are
subject to uncertainties due to incomplete or inaccurate measurements [Hill and Tiedeman, 2007;
Tartakovsky, 2013]. Such uncertainties may be quite large, because direct measurements of the properties
are only available at a limited number of boreholes, and indirect measurements, inferred from well tests or
subsurface data, usually involve averaged responses over a large scale. Thus, analyzing the uncertainties of
the properties and quantifying their influences on predictions of the fluid behavior have become critical to
facilitate science-informed decision making in subsurface environment management.
A widely used approach for describing uncertainty in subsurface flow is to represent the porous media
property (e.g., permeability) as a parametrized random field. This naturally leads to simulating subsurface
flows using partial differential equations (PDEs) with random input parameters. Since realistic random field
models often need a rather large number of stochastic degrees of freedom (>100 s) for their accurate repre-
sentation of the heterogeneous porous media property, the main computational challenge is to efficiently
C 2016. American Geophysical Union.
V solve these stochastic PDEs. Several stochastic techniques like perturbation/moment methods [Zhang,
All Rights Reserved. 2002], stochastic Galerkin methods [Ghanem, 1999], stochastic collocation methods [Nobile et al., 2008a,

LU ET AL. AN IMPROVED MLMC METHOD 9642


Water Resources Research 10.1002/2016WR019475

2008b; Barajas-Solano and Tartakovsky, 2016], and model order reduction methods [Pasetto et al., 2014]
have been applied to porous media flow problems [Ghanem, 1998; Zhang and Lu, 2004; Li and Zhang, 2007;
Laloy et al., 2013; Zhang et al., 2013]. Among these techniques, the moment methods have wide application
for their easy implementation. However, solving moment equations is very time consuming [Franssen et al.,
2009] and the solutions may be a poor approximation when the random permeability field is highly hetero-
geneous [Guadagnini and Neuman, 1999a, 1999b]. The stochastic Galerkin, stochastic collocation, and model
order reduction methods have drawn significant interest recently for their impressive approximation accu-
racy and computational efficiency [Gunzburger et al., 2014]. However, these surrogate-based methods are
not suitable for highly heterogeneous problems either, due to the explosion in computational cost with
increasing stochastic dimensions [Ma and Zabaras, 2009].
Hence, standard Monte Carlo (MC) simulation, due to its mild dependence on the dimensionality, is still the
method of choice, and sometimes the only choice for stochastic subsurface flow problems with extremely
high heterogeneity. In MC simulations, the effect of random model parameters on predictions is commonly
analyzed in the following procedure: first generate a set of realizations of the parameterized random field,
then solve the underlying PDE model to obtain predictions for each realization, and last calculate predictive
statistics, e.g., moments or distribution functions, of the quantities of interest (QoI). Such procedure is, how-
ever, computationally demanding because for each individual realization, the PDE model has to be numeri-
cally solved on a sufficiently fine mesh to minimize the discretization error. Furthermore, the notoriously
slow convergence rate of the standard MC method requires a large number of such simulations so as to
reduce the sampling error and thus obtain accurate approximation [Smith, 2013].
To improve computational efficiency of the standard MC methods, multilevel Monte Carlo (MLMC) methods,
which integrate multifidelity approximations with MC simulations, were proposed [Heinrich, 2001]. The pri-
mary concept behind multilevel methods is to solve the underlying PDE model on a sequence of geometric
meshes with increasing resolution [Giles, 2008], where the coarsest and the finest meshes are labeled as the
lowest and the highest levels, respectively. As opposed to the standard MC method that performs only on
the highest level, the MLMC method calculates the predictive statistics by conducting sampling on multiple
levels in a way that achieves the same accuracy but reduces the overall cost. The computational saving
comes from the fact that the MLMC method can smartly allocate more samples on the coarse levels with
inexpensive model executions, and fewer samples on the fine levels with expensive model runs. Moreover,
the implementation of the multilevel method is as easy as the standard MC method [Nobile and Tesei,
2015].
The goal of this effort is to develop an improved MLMC algorithm to enhance the computational efficiency
of approximating cumulative distribution functions (CDFs) of QoIs. The CDFs provide detailed descriptions
of the uncertainty in prediction [Tversky and Kahneman, 1992; Goovaerts, 2001; Boso and Tartakovsky, 2016];
however, the MLMC technique has been mainly used to estimate statistical expectations [Barth et al., 2011;
Giles and Reisinger, 2012; Charrier et al., 2013; Teckentrup et al., 2013; Teckentrup, 2013] and has not been
applied to CDF approximation until very recently in Giles et al. [2015]. Since the MLMC method can be
viewed as a variance reduction technique, its efficiency depends heavily on the decay of the variance of the
QoI as the level increases, where a fast variance decay leads to a small computational cost to achieve a pre-
scribed accuracy e > 0. Theoretical studies show that asymptotically, i.e., e approaches zero arbitrarily close-
ly, the MLMC method can outperform the standard MC as long as there exists a reduction in the variance
[Cliffe et al., 2011; Efendiev et al., 2014; Dodwell et al., 2015]. Nevertheless, conducting highly accurate simula-
tions in the asymptotic regime is usually not realistic, because of irreducible discretization errors caused by
the limited computer resources and the restriction of the measurement scale of the permeability field. In
the preasymptotic regime, i.e., e is not extremely small, the variance must decay sufficiently fast to show the
superiority of the MLMC method. Unfortunately, this is difficult to achieve in CDF approximation, because
the QoI is the indicator function whose discontinuity will dramatically deteriorate the variance decay [Zhang
et al., 2016]. To tackle this challenge, a smoothing function was used to approximate the indicator function,
so as to accelerate the variance decay. Since the use of the smoothing function introduces an additional
error into the MLMC estimator, a careful calibration of the smoothing function is required to maximize the
acceleration as well as control the approximation error. However, the calibration strategy in the previous
study was designed based on asymptotic error analysis, which may provide insufficient smoothing effect in
the preasymptotic regime and consequently deteriorates the MLMC performance.

LU ET AL. AN IMPROVED MLMC METHOD 9643


Water Resources Research 10.1002/2016WR019475

This work has three major innovative contributions. First, this work develops an optimization technique
to improve the calibration of the smoothing function by balancing the cost saving and the accuracy
loss, so as to achieve superior performance of the MLMC method. Specifically, we use a small number
of initial samples at each level to numerically estimate the error from smoothing. The estimate is then
used to calibrate the smoothing function to reach the near-maximal smoothing effect and meanwhile
restrict the smoothing error to the prescribed accuracy requirement. This strategy does not require any
extra computational effort, as the initial samples will be reused for MLMC estimation. The numerical
results in section 5 show that our strategy can significantly improve the efficiency of the MLMC method
in the stochastic oil reservoir simulations. Second, we design an a posteriori procedure to choose the
more efficient estimator on the fly between the standard MC and the MLMC methods. This is motivated
by the situations that the variance reduction is not fast enough, even with the smoothing step, to
ensure the superiority of the MLMC. The situations usually happen to a problem simulated on a coarse
mesh resulting in a relatively large numerical error that greatly deviates from the asymptotic condition.
For example, the problem could have a large simulation domain so that only coarse mesh can be
affordable with available computer resources. In this case, a blind use of the MLMC method may result
in a higher computational cost than using the standard MC. Our designed procedure utilizes the same
set of initial samples to predict the cost of the two sampling methods and automatically choose the
more efficient scheme according to the desired accuracy. Again, this procedure does not require any
extra model executions. Finally, we integrate all the proposed techniques into a practical algorithm and
apply it to a highly heterogeneous oil reservoir model. In analyzing the results, we discuss the factors
impacting the MLMC performance and give guidance for its efficient application. The developed algo-
rithm is mathematically general, and can be employed to a wide range of stochastic subsurface
problems.
The rest of the paper is organized as follows. In section 2, we define a two-phase flow and transport
problem and formulate the CDF approximation using the sampling methods. In section 3, the standard
MC method in approximation of the CDF is briefly reviewed. In section 4, we introduce the improved
MLMC method and present the practical algorithm for the CDF estimation. In section 5, the MLMC
approach is applied to an oil reservoir model and its effectiveness and efficiency is demonstrated in
comparison with the standard MC. In section 6, we discuss the factors affecting the efficient usage of
the MLMC approach and give guidance for its practical application. In section 7, we end this paper by
giving conclusions.

2. Problem Formulation
We consider an incompressible and immiscible two-phase flow in heterogeneous porous media, where
gravity and capillary effects are neglected and porosity is assumed a constant. The two phases are referred
to water and oil, denoted as w and o, respectively. Assume the two phase saturations having Sw 1So 51, the
simulation of water injection into an oil reservoir in a bounded domain D for a time period T can be
described by a flow equation for fluid pressure p and a transport equation for the phase saturation as
[Helmig, 1997]
8
>
> r  v ¼ g;
>
>
<
v ¼ 2KkðSw Þrp;
(1)
> @S
>   g
>
> w w
:/ þ r  f ðSw Þv ¼ ;
@t qw

subject to suitable boundary and initial conditions. In the equations, v is the Darcy velocity of the engag-
ing fluids, g is the source term, / is the porosity, q represents the density, the term f ðSw Þv describes vis-
cous force, and K is the random permeability, which we consider to be isotropic and is simulated as a
random field. Total mobility is defined by kðSw Þ :5kw ðSw Þ1ko ðSw Þ, where ki models the reduced mobility
of phase i due to presence of the other phase (i represents w or o). The mobility kw and ko depends on
the viscosity lw and lo, the irreducible oil saturation Sor, and the connate water saturation Swc in the fol-
lowing way,

LU ET AL. AN IMPROVED MLMC METHOD 9644


Water Resources Research 10.1002/2016WR019475

S2nw ð12Snw Þ2 Sw 2Swc


kw ðSw Þ5 ; ko ðSw Þ5 ; Snw 5 : (2)
lw lo 12Sor 2Swc
We are interested in estimating the mass flow rate Q at a specific time-space point ðt; xÞ, where Q5qvðt; xÞA
with A representing the cross-sectional area. Since the permeability K is treated as a random field, the output
Q is a probability distribution function.
In this effort, our goal is to approximate the CDF of the quantity of interest Q, denoted by F(q), using the
Ðq
MLMC and the standard MC methods. For any fixed point q 2 R, F(q) is defined as FðqÞ5 21 f ðQÞdQ, where
f(Q) is corresponding probability density function. In use of Monte Carlo-type approaches, the F(q) is formu-
lated as the expectation of an indicator function I ð21;q ðQÞ as
ð
  11
FðqÞ5E I ð21;q ðQÞ 5 I ð21;q ðQÞf ðQÞdQ; (3)
21

where I ð21;q ðQÞ51 for Q 2 ð21; q and I ð21;q ðQÞ50 for Q 2 ðq; 11Þ.
We aim at approximating F(q) on a bounded domain ½a; b  R. To this end, we first discretize ½a; b into a
sequence of S h with S 1 1 equidistant points, i.e.,

S h :5fa5q0 < q1 <    < qS 5bg with h5max1nS ðqn 2qn21 Þ: (4)

Then, we estimate Fðqn Þ at the discrete points fqn gSn50 . Last, we approximate F(q) through the following
piecewise Lagrange interpolant,

XS X
S
Fh ðqÞ :5 Fðqn Þ un ðqÞ5 E½I ð21;qn  ðQÞ un ðqÞ; (5)
n50 n50

where un ðqÞ represents piecewise Lagrange polynomial basis functions. When h is sufficiently small, Fh ðqÞ
can be an accurate approximation of F(q).
In practice, the PDEs in (1) are usually solved by numerical methods, e.g., finite element or finite volume
techniques, which require a discrete mesh of the spatial domain D. We denote by T M the mesh for compu-
tation with M cells in D, and use QM to represent the corresponding numerical approximation of Q.
Substituting QM into (5), we obtain an approximation of Fh ðqÞ, i.e.,

X
S
Fh;M ðqÞ :5 E½I ð21;qn  ðQM Þ un ðqÞ: (6)
n50

According to numerical analysis, it is reasonable to assume that the expected value E½I ð21;qn  ðQM Þ ! E
½I ð21;qn  ðQÞ and consequently Fh;M ðqÞ ! Fh ðqÞ as M ! 1. Assume the order of convergence is a > 0, it
has

jE½I ð21;q ðQM Þ2I ð21;q ðQÞj  c1 M2a ; (7)

for some constant c1 > 0 independent of q.


In this work, we focus on analyzing the computational cost of the MLMC method in approximation of F(q)
using the formulation in (6), given a desired root mean square error (RMSE) accuracy e > 0. For comparison,
we use both the MLMC and the standard MC methods to evaluate Fh;M ðqÞ. The approximation error
between Fh;M ðqÞ and F(q) consists of two parts: the interpolation error Fh ðqÞ2FðqÞ and the error in estimat-
ing the expectations Fh;M ðqÞ2Fh ðqÞ. When the number of interpolation points S h is sufficiently large, the
interpolation error can be negligible. In addition, increasing the number of S h does not bring extra compu-
tational costs because the calculation of Fðqn Þ at interpolation points fqn gSn50 uses the same set of QM sam-
ples. So in the following discussion, we assume the interpolation error is zero, and concentrate on the error
Fh;M ðqÞ2Fh ðqÞ.

LU ET AL. AN IMPROVED MLMC METHOD 9645


Water Resources Research 10.1002/2016WR019475

3. The Standard MC Method for Estimating CDFs


For notational simplicity, we use I n ðÞ to represent I ð21;qn  ðÞ. The standard MC estimator for E½I n ðQM Þ at
the point qn is defined by

X
NMC
ðiÞ
I MC 21
n;M :5NMC I n ðQM Þ; (8)
i51

ðiÞ
where QM is the ith sample of QM, and NMC is the total number of independent realizations of QM on the
mesh T M . Note that I MC MC
n;M is an unbiased estimator of E½I n ðQM Þ, i.e., E½I n;M 5E½I n ðQM Þ. Applying (8) to all

the interpolation points fqn gSn50 , we obtain the following MC estimator for Fh;M ðqÞ,

X
S
MC
Fh;M ðqÞ :5 I MC
n;M un ðqÞ: (9)
n50

MC MC
The error eðFh;M Þ :5Fh;M 2Fh consists of the sampling error for estimating E½I n ðQM Þ and the discretization
MC
error QM 2Q. Specifically, the mean square error (MSE) of E½jjeðFh;M Þjj2  can be bounded by

h i h i
MC
E jjeðFh;M Þjj2 5E jjFh;M
MC
2Fh jj2
h i h i
MC 2
MC
 E kFh;M 2E½Fh;M MC
k 1E kE½Fh;M 2Fh k2
 2  (10)
MC MC
 max E I n;M 2E½I n;M  1kFh;M 2Fh k2
0nS

21
 max NMC V½I n ðQM Þ1max jE½I n ðQM Þ2I n ðQÞj2 ;
0nS 0nS

where jj  jj and V½ denote the L1 norm and the variance operator, respectively. The sampling
21
error NMC V½I n ðQM Þ can be reduced by increasing the number of samples NMC , and the discretization error jE½
I n ðQM Þ2I n ðQÞj2 can be reduced by refining the mesh T M . Since both strategies require the increase of com-
putational cost, the MLMC method improves efficiency by working on the variance as explained in section 4.
MC
To make the estimator Fh;M achieve a prescribed RMSE accuracy e > 0, a sufficient condition can be
imposed such that both the sampling and the discretization errors in (10) are smaller than e2 =2. As such,
the accuracy e can be achieved by setting
 pffiffiffi 21=a
NMC 52e22 max V½I n ðQM Þ and M  e=ð 2c1 Þ ; (11)
0nS

where a and c1 are given in (7). We assume that the computational cost CM of simulating one sample of QM
can be bounded by

C M  c2 Mc ; (12)

for some constants c > 0 and c2 > 0. Then the minimum cost of the standard MC estimator to achieve e is
on the order of

c c=a
MC
CðFh;M Þ5CM NMC  2112a c1 c2 max V½I n ðQM Þ e222c=a 5Oðe222c=a Þ: (13)
0nS

4. The MLMC Method for Estimating CDFs


In this section, we first briefly recall the general procedure of the MLMC simulation in section 4.1 where we
demonstrate that the MLMC performance critically depends on the variance decay. Then in section 4.2, we
introduce the smoothing strategy to accelerate the variance decay and thus improve the computational
efficiency. At last, we summarize our MLMC algorithm for CDF estimation in section 4.3.

LU ET AL. AN IMPROVED MLMC METHOD 9646


Water Resources Research 10.1002/2016WR019475

4.1. The General Idea of the MLMC Simulation


The main idea of the MLMC method is to not just sample from one approximation QM of Q, but from a
sequence of approximations fQM‘ ; ‘50; . . . ; Lg associated with the meshes fT M‘ ; ‘50; . . . ; Lg with
M0 < M1 <    < ML 5M. Usually, M‘21 522d M‘ for ‘51; . . . ; L, where d is dimension of the spatial domain
D; ‘ is the level of mesh resolution, and M‘ is the number of cells of the mesh T M‘ . In this way, the expecta-
tion E½I n ðQM Þ can be expressed as
X
L X
L
E½I n ðQM Þ5E½I n ðQM0 Þ1 E½I n ðQM‘ Þ2I n ðQM‘21 Þ5 E½I n ðY‘ Þ; (14)
‘51 ‘50

where I n ðY0 Þ :5I n ðQM0 Þ and I n ðY‘ Þ :5I n ðQM‘ Þ2I n ðQM‘21 Þ for 1  ‘  L. As such, E½I n ðY‘ Þ for ‘50; . . . ; L
in (14) can be estimated by the following MC estimator,
N‘  
1 X N0
ðiÞ 1X ðiÞ ðiÞ
I^ n ðY0 Þ :5 I n ðQM0 Þ and I^ n ðY‘ Þ :5 I n ðQM‘ Þ2I n ðQM‘21 Þ ; (15)
N0 i51 N‘ i51
ðiÞ ðiÞ
where N‘ is the number of samples on level ‘. Note that the quantities QM‘ and QM‘21 are computed using
the same set of N‘ samples but simulated on two meshes T M‘ and T M‘21 of consecutive levels. Consequent-
ly, the MLMC estimator of E½I n ðQM Þ is
X
L
I ML
n;M :5 I^ n ðY‘ Þ: (16)
‘50

The expression in (14) indicates that the expectation on the finest level is equal to the expectation on the
coarsest level plus a sum of difference in expectation between simulations on consecutive levels. In this
way, the less accurate estimate on the coarsest level can be gradually corrected by the estimates on the fin-
er levels, thereby the MLMC estimator can achieve the same accuracy as the standard MC estimator defined
solely on the finest level.
P
ML
Similar to (9), the MLMC estimator for Fh;M ðqÞ is defined as Fh;M ðqÞ :5 Sn50 I ML
n;M un ðqÞ. In the same manner
ML ML
of (10), the error eðFh;M Þ :5Fh;M 2Fh in the MSE sense can be bounded by
h i X
L
ML
E jjeðFh;M Þjj2  max N‘21 V½I n ðY‘ Þ1max jE½I n ðQML Þ2I n ðQÞ j2 ; (17)
0nS 0nS
‘50

where the second term of discritization error is exactly the same with (10) as we define ML 5 M. To make it
pffiffiffi 21=a
consistent with the standard MC, the same sufficient condition, ML  e=ð 2c1 Þ , can be taken by
requiring the discritization error smaller than e2 =2. To achieve another condition that the sampling error
is also smaller than e2 =2; N‘ can be determined in multiple ways. Here we determine N‘ using the same
ML
strategy as in Giles [2008] such that the variance of estimator Fh;M is minimized. That is, we set
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
N‘ 5c0 V½I n ðY‘ Þ=C‘ , where C‘ 5CM‘ 1CM‘21 is the computational cost of simulating one realization of Y‘
P
and c0 > 0 is  a constant. By setting max 0nS L‘50 N‘21 V½I n ðY‘ Þ5e2 =2, we can solve
P L
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
c0 5max 0nS 2e22 ‘50 V½I n ðY‘ ÞC‘ . Thus,

!
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi X
22
L pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
N‘ 5 max 2e V½I n ðY‘ Þ=C‘ V½I n ðYk ÞCk ; (18)
0nS
k50

ML
for ‘50; . . . ; L. Then, the total cost of the MLMC estimator Fh;M can be expressed by
XL
ML
CðFh;M Þ5 C ‘ N‘ ; (19)
‘50

with N‘ defined in (18).


It can be shown that to achieve the desired e, the cost of the MLMC estimation in (19) is asymptotically
lower than that of the standard MC in (13), as long as the variance V½I n ðY‘ Þ decays reciprocally with
‘, i.e., V½I n ðY‘ Þ ! 0 as ‘ ! 11. On the other hand, it is easy to verify that V½I n ðQM Þ is approximately

LU ET AL. AN IMPROVED MLMC METHOD 9647


Water Resources Research 10.1002/2016WR019475

3 constant, independent of M,
β = 0.1 which results in V½I n ðY‘ Þ
β = 0.2  V½I n ðQM Þ and consequent-
2.5 β = 0.4 ly N‘  NMC for high levels of
β = 0.8 ‘. This indicates that the

MLMC cost / MC cost


MLMC estimator simulates a
2 large number of samples at
the coarse levels with inex-
pensive model runs, but much
1.5
fewer samples at the fine lev-
els with expensive model
1 runs. In contrast, the standard
MC simulates a very large
number of samples on the fin-
0.5 est level solely with the most
expensive model executions.
The cost saving of the MLMC
0 −2 −4 −6 −8 −10 method can be significant
10 10 10 10 10
when the variance V½I n ðY‘ Þ
Prescribed accuracy ε
has a fast decay. We assume
Figure 1. A simple example illustrates the ratio of the computational cost between the the decay rate in the order of
MLMC and the standard MC for a set of variance decay rates b. When the prescribed accu- b > 0, i.e.,
racy e is sufficiently small, the MLMC method is more efficient than the standard MC with
the ratio less than 1. However, for a relatively large e, it is likely that the MLMC is less V½I n ðY‘ Þ  c3 M2b
‘ ; (20)
efficient than the standard MC especially for a small b.
for some constant c3 > 0.
Cliffe et al. [2011] proved that if there are positive constants a; b; c > 0 such that (7), (20), and (12) hold, the
minimum cost of the MLMC estimator achieving the desired accuracy e is in the order of
8
>
> Oðe22 Þ; if b > c;
>
<  
2
ML
CðFh;M Þ5 O e22 ðlog eÞ ; if b5c; (21)
>
>
>
:
O e222ðc2bÞ=a ; if b < c:

This indicates that if the variance V½I n ðY‘ Þ decays faster than the cost CM‘ increases, i.e., b > c, then the
cost on the coarsest level ‘50 will dominate. As such, the cost saving of the MLMC method will be asymp-
totically Oðec=a Þ, which reflects the ratio of the costs of simulations on the coarsest and the finest levels. On
the contrary, if the variance V½I n ðY‘ Þ decays slower than CM‘ increases, i.e., b < c, then the cost on the fin-
est level L will dominate, and the cost saving of the MLMC will be asymptotically Oðeb=a Þ. Hence, in both
cases, the MLMC approach asymptotically outperforms the standard MC; and the faster the variance decay
is, the larger cost savings the MLMC estimator can obtain.
The analysis in (21) is conducted in the asymptotic regime, namely, it is only valid when e is extremely close
to zero. However, in a real application such as a subsurface flow simulation, it is usually not affordable or
possible to reach a very high accuracy due to the irreducible discretization errors caused by the limited
computational resources and the restriction of the measurement scale of the parameter random field.
Therefore, practically, most subsurface simulations are conduced in the preasymptotic regime, in which the
desired accuracy e is usually greater than 1024 [Mukherjee, 2013]. In this case, we find out that the MLMC
method may be inferior to the standard MC when the variance does not decay sufficiently fast. Here we
give a simple example for illustration. We assume M0 54; M‘ 54M‘21 54‘21 M0 ; c1 5c2 5c3 51:0; a51:0, and
c51:0; and we consider a set of b50:1; 0:2; 0:4; and 0:8. For a fixed e > 0, we determine the maximum level
pffiffiffi 21=a
L by solving ML 5 e=ð 2c1 Þ , and calculate the number of samples N‘ for ‘50; . . . ; L based on (18).
For comparison, the cost of the standard MC using samples of QM is also computed. The results of the ratio
ML MC
CðFh;M Þ=CðFh;M Þ for a set of e are presented in Figure 1. It is observed that the MLMC estimator outperforms
the standard MC for all the values of b when e is sufficiently small (i.e., e  1028 ). However, for a relatively slow
variance decay with b50:1, the MLMC method is inferior to the standard MC for e 2 ½1027 ; 1022  that covers

LU ET AL. AN IMPROVED MLMC METHOD 9648


Water Resources Research 10.1002/2016WR019475

the usually required accuracy range in practice. Unfortunately, such slow variance decay rate can easily occur in
the MLMC estimation of the CDFs because of the jump discontinuity of the indicator function I ðÞ used in the
calculation of I n ðY‘ Þ. To address this problem, we define a smoothing function to remove the singularity, and
use it in the approximation of Fh;M ðqÞ to accelerate the variance decay and improve the computational
efficiency.

4.2. A Smoothing Technique to Improve the MLMC Efficiency


We assume that the probability density function f(Q) is r times continuously differentiable in ½a; b. A
smoothing function is defined based on rescaled translation of a polynomial g:R ! R as follows [Giles et al.,
2015]:
1. g(Q) 5 1 for Q < 21, and g(Q) 5 0 for Q > 1.
2. For Q 2 ½21; 1, g(Q) is a polynomial of degree at most r 1 1, whose coefficients can be obtained by solv-
Ð1
ing the system 21 Qj gðQÞdQ5ð21Þj =ðj11Þ; for j50; . . . ; r21, as well as the conditions gð1Þ50 and
gð21Þ51.
3. The rescaled function gððQ2qn Þ=dÞ is the smoothing function to approximate I n ðQÞ, and the constant
d > 0 is the smoothing factor.
The factor d controls the smoothing effect of the function, namely, the larger the d is, the more gently the
function changes in the neighborhood of qn. As d ! 0, essentially there is no smoothing and
gððQ2qn Þ=dÞ ! I n ðQÞ. The d plays an important role in computational efficiency of the MLMC and careful
calculation of d is critical as discussed at the end of this section.
Similar to (16), the MLMC estimator of E½gððQM 2qn Þ=dÞ is
X
L
gML
n;M :5 ^ n ðY‘ Þ;
g (22)
‘50

where

ðiÞ
!
1 XN0
QM0 2qn
^ n ðY0 Þ :5
g g ;
N0 i51 d
" ðiÞ
! ðiÞ
!#
1X N‘
QM‘ 2qn QM‘21 2qn
^ n ðY‘ Þ :5
g g 2g ;
N‘ i51 d d

for ‘51; . . . L. Correspondingly, the MLMC estimator with smoothing for Fh;M ðqÞ can be defined by

X
S S X
X L
ML
Fh;d;M ðqÞ :5 gML
n;M un ðqÞ5 ^ n ðY‘ Þ un ðqÞ:
g (23)
n50 n50 ‘50

ML
In the same manner of (10), the MSE of Fh;d;M ðqÞ for the approximation of Fh ðqÞ can be bounded by

ML
E½jjeðFh;d;M Þjj2   E½jjFh;d;M
ML ML
2E½Fh;d;M jj2  1 jjE½Fh;d;M
ML
2Fh;M jj2 1 jjFh;M 2Fh jj2 ; (24)
|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl} |fflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflffl} |fflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflffl}
eML eML eML
3
1 2

where eML ML
1 is the sampling error, e2 is the smoothing error caused by the smoothing function approxima-
tion, and e3 is the discretization error same as the one in (10). The sampling error eML
ML
1 can be bounded by

X
L
eML
1  max N‘21 V½gn ðY‘ Þ
0nS
‘50


 (25)
X L
QM‘ 2qn QM‘21 2qn
21
5max N‘ V g 2g
0nS
‘50
d d

and the smoothing error eML


2 can be bounded by

LU ET AL. AN IMPROVED MLMC METHOD 9649


Water Resources Research 10.1002/2016WR019475

 
2
 QM 2qn 
eML ML 2  2I n ðQM Þ   c4 d2ðr11Þ ;
2 5jjFh;d;M 2Fh;M jj  max E g (26)
0nS d

for some constant c4 > 0. The proof of (26) is based on the Taylor expansion in a straightforward way.
To achieve the desired accuracy e in estimating Fh ðqÞ, a sufficient condition is that eML 2 ML 2
1  e =4; e2  e =4,
and e3  e =2. Here for the sake of comparison between the MLMC and the standard MC methods, eML
ML 2
3
is set to have the same error bound as in (10), so the same condition can be used for both methods
to determine the finest mesh T M and the highest level L. That is, M 5 ML is determined by
pffiffiffi
max 0nS jE½I n ðQM Þ2I n ðQÞj  e= 2. Based on the triangle inequality, we have max 0nS jE½I n ðY‘ Þj
pffiffiffi
max 0nS jE½I n ðQML Þ2I n ðQÞj, so that we use, in practice, max 0nS jE½I n ðY‘ Þj  e= 2 to determine the
required maximum level L and the finest mesh T M .
The key step of the smoothing technique is to properly determine the value of d, so as to balance the errors
as well as maximize the variance decay. It is easy to see that the optimal choice of d is the maximum value
that can be reached under the condition of eML 2
2  e =4. In other words, we aim to enlarge d such that e2 is
ML
2
as close to e =4 as possible. In Giles et al. [2015], d was determined by exploiting the asymptotic upper
bound in (26). They first set c4 d2ðr11Þ 5e2 =4 such that d5ð4c4 Þ1=ð2r12Þ e1=ðr11Þ , then by ignoring the unknown
constant ð4c4 Þ1=ð2r12Þ they defined d as

dG :5e1=ðr11Þ : (27)

Here we denote the d calculated in (27) as dG to distinguish our d defined below. The calculation of dG is rea-
sonable in the asymptotic regime where the omitted constant is negligible compared to the dominant term
e1=ðr11Þ . However, in the preasymptotic regime, such choice of d might be much smaller than the optimal
value thus dramatically deteriorates the MLMC performance, as demonstrated in section 6.
In this effort, we design an optimization scheme to estimate the optimal value of d for any given e. Our strat-
egy is to define d by solving the equation eML 2
2 5e =4. Specifically, given the accuracy e, the set of interpola-
tion points S h 5fqn ; n50; . . . ; Sg, the polynomial order r, and the maximum level L, we first estimate dn for
a single point qn by solving

 Ninit " ðiÞ


! #
1 X QM 2qn ðiÞ  e
g 2I n ðQM Þ 5 ; (28)
Ninit  i51 dn 2

ðiÞ
based on a set of initial samples fQM gNi51
init
; then we define d as the maximum value in fdn ; n50; . . . ; Sg, i.e.,

do :5 max dn : (29)
0nS

Here we denote our d calculated in (29) as do to distinguish the dG defined above. As Ninit ! 1, the choice
of d based on (28) and (29) is optimal in the setting. Practically, a larger Ninit value would give better optimi-
zation and note that all the Ninit samples can be reused in the MLMC estimation, so that there will be no
waste of model executions as long as Ninit  NL .

4.3. An Improved MLMC Algorithm for CDF Estimation


It has been theoretically proved that the MLMC method with the smoothing technique is never inferior to
the standard MC asymptotically [Giles et al., 2015]. However, practically we can almost never expect e ! 0
and L ! 1 due to the irreducible discretization errors; and the L is usually limited to a value of Lmax whose
mesh resolution T MLmax is determined by the affordable computing resources and the measurement scale of
the parameter random field. So in the real-world application, even with our improved smoothing technique,
it is possible that the maximum variance decay is still not fast enough to make the MLMC outperform the
standard MC, especially for large e as illustrated in section 5. Here we develop a practical algorithm for CDF
estimation by comparing the performance of the two sampling methods first based on a set of initial sam-
ples. This preprocessing adaptively chooses the more efficient method according to the prescribed accuracy
e. Our final algorithm is summarized as follows.

LU ET AL. AN IMPROVED MLMC METHOD 9650


Water Resources Research 10.1002/2016WR019475

Algorithm
1. Input: The RMSE accuracy e, interpolation points S h 5fqn ; n50; . . . ; Sg, a sequence of discrete meshes
fT M‘ ; ‘50; . . . ; Lmax g, the computational costs fCM‘ ; ‘50; . . . ; Lmax g, and the smoothing polynomial
order r;
2. Output: an estimate of the CDF, F(q);
3. Procedure:
4. L 5 0;
5. while L  Lmax do
6. Set L5L11;
7. Draw Ninit initial samples of QML and QML21 ;
8. Determine d using (28) and (29) with the Ninit samples;
9. for n50; . . . ; S do
MC P init ðiÞ
10. Compute the sample mean I~ n;ML : 5 Ni51 I n ðQML Þ=Ninit ;
PNinit ðiÞ MC
11. Compute the sample variance V~ ½I n ðQML Þ :¼ i¼1 ðI n ðQML Þ2I~ n;ML Þ2 =Ninit ;
PNinit ðiÞ ðiÞ
12. Compute the sample mean I~ n ðYL Þ : 5 i51 ðI n ðQML Þ2I n ðQML21 ÞÞ=Ninit ;
P Ninit ðiÞ
13. Compute the sample mean g ~ ML
n ðYL Þ : 5 i51 gn ðYL Þ=Ninit , where

ðiÞ ðiÞ ðiÞ


gn ðYL Þ : 5gððQML 2qn Þ=dÞ2gððQML21 2qn Þ=dÞ; i51; . . . ; Ninit ;
PNinit ðiÞ 2
14. Compute the sample variance V~ ½gn ðYL Þ :¼ i¼1 ~ ML
ðgn ðYL Þ2g n ðYL ÞÞ =Ninit ;
15. end for
pffiffiffi
16. if max jI~ n ðYL Þj  e= 2 or L5Lmax then
0nS
17. for ‘50; . . . ; L do
18. Estimate N‘ according to,
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi X !
L qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
N‘ : 5 max 4e ~ 22
V ½gn ðY‘ Þ=C‘ ~
V ½gn ðY‘ ÞCk ;
0nS
k50

19. end for


20. Estimate NMC based on (11), i.e.,

NMC : 5 max 2e22 V~ ½I n ðQML Þ;


0nS

P
21. MC
Compute the costs CðFh;ML
ML
Þ5CML NMC and CðFh;d;M L
Þ5 L‘50 C‘ N‘ ;
MC ML
22. if CðFh;ML
Þ  CðFh;d;M L
Þ then
MC
23. Calculate Fh;M L
ðqÞ in (9) by drawing extra NMC 2Ninit samples of QML ;
24. else
ML
25. Calculate Fh;d;M L
ðqÞ in (23) by drawing extra N‘ 2Ninit samples of QM‘ ;
26. end if
27. Break;
28. end if
29. end while

The Algorithm may not lead to the optimal choice between the standard MC and the MLMC methods,
because the true values of the L, fN‘ gL‘50 and NMC are unknown and they were estimated using sample
means and sample variances a posteriori based on the Ninit samples. Theoretically, it can be proved that
these estimates will converge to the true values with probability one as the initial sample size goes to infini-
ty. However, practically, with finite initial samples, these estimates depend on the sample size and specific
set of samples. This is true for all the estimates using the sampling methods. So, in application of the Algo-
rithm, it is suggested properly enlarging the initial sample size to improve the estimation accuracy and
reduce the possibility of making inappropriate decision, and those samples will be reused for the CDF
estimation.

LU ET AL. AN IMPROVED MLMC METHOD 9651


Water Resources Research 10.1002/2016WR019475

Figure 2. Application to an oil reservoir simulation. (a) The log permeability field log ðKÞ from which we collect measurements; location I represents the injection well and P represents
the production well in which the mass flow rate is the QoI; (b) locations of 36 measurements drawn from Figure 2a; conditioning on these measurement data, the realizations of log ðKÞ
field are generated; (c) a demonstration of one realization of log ðKÞ field.

5. Application to an Oil Reservoir Simulation


5.1. Model Description
To demonstrate efficiency and effectiveness of the improved MLMC method, we apply the Algorithm to a
synthetic study of an oil reservoir simulation. The numerical experiment is designed based on the fine-
grid model from the tenth SPE comparative solution project [Christie and Blunt, 2001]. The original model
is three dimensional (3-D) having size of 365:763670:56351:816 m3 and total 603220385 grid cells. The
initial pressure is 6000 psi. There are four production wells at the four corners of the domain, each pro-
ducing at 4000 psi bottom hole pressure, and a central injection well with an injection rate of 5000 bar-
rels/d. All wells are vertical and completed throughout the formation. The simulation of this sufficiently
fine model is very time costly. With the efficient reactive flow and transport simulator, PFLOTRAN [Mills
et al., 2007], running in 10 processors, it takes about 8 h to complete one simulation of 100 days. To test
and demonstrate the MLMC algorithm in reasonable time, the oil reservoir model is simplified in the fol-
lowing ways: (i) it has two-dimensional domain with the same spatial resolution as one layer of the origi-
nal model, i.e., a 365:763670:56 m2 domain with a total of 60 3 220 grid cells; (ii) the synthetic
permeability (K) field is taken from the first layer of the original geological model as shown in Figure 2a
and the porosity is set to a constant of 0.1; (iii) the injection rate is changed to 62.90 barrels/d to make
the model physically reasonable; and (iv) the QoI is the oil production rate at the well P (Figure 2a) after 2
days from the initial time.
Although it is a simplification of the 3-D case, the 2-D model considered here is still a high-dimensional and
highly heterogeneous. The number of stochastic degrees of freedom of the model is 60 3 220; this large
number of parameters is extremely challenging for the stochastic Galerkin and stochastic collocation meth-
ods [Ma and Zabaras, 2009], and MC sampling methods are typically the method of choice. In addition, the
model has large variability in permeability. As shown in Figure 3, the fitted spherical model of the sampled
variogram indicates that the variance of the permeability is about 1.13 m2, and the coefficient of variation is about
rK =hKi5297%. This large variability suggests that the computational cost would be significantly high to
obtain an accurate uncertainty quantification and an advanced algorithm is required to improve the
computational efficiency. Moreover, the theoretical study in Cliffe et al. [2011] showed that the computa-
tional improvement of MLMC over the standard MC were similar for the 2-D and 3-D cases. Our algo-
rithm is tested on a workstation with four AMD Opteron 6376 Abu Dhabi 2.3 GHz CPUs, each of which
has 16 cores, i.e., a total of 64 cores.

5.2. Application of the MLMC Algorithm


We are interested in investigating the influence of the permeability uncertainty on the mass flow rate Q of production
well P. Specifically, we want to estimate the CDF of log ðQÞ using the better estimator between the standard MC and

LU ET AL. AN IMPROVED MLMC METHOD 9652


Water Resources Research 10.1002/2016WR019475

1.5 the MLMC. We assumed that the logarithm of


Sampled the random permeability field, log ðKÞ, is a
Fitted Spherical model Gaussian process whose mean and covariance
structure can be inferred from the 36 synthetic
data regularly situated in Figure 2b. Condi-
1 tioned on such data and based on the sam-
pled variogram shown in Figure 3, random
Variogram realizations of the permeability field were gen-
erated using the sequential Gaussian simula-
tion program (SGSIM) in GSLIB [Deutsch and
0.5 Journel, 1998; Lu et al., 2011]. One realization is
plotted in Figure 2c for illustration; all the real-
izations have the same log ðKÞ values at the
36 locations but differ from each other
elsewhere.
0 To implement the Algorithm, we first gen-
0 50 100 150
erated a set of candidate meshes for the
Lag distance (Meter)
definition of fT M‘ gL‘50 . Unlike the theoreti-
Figure 3. Sampled and fitted variogram using the spherical model for the cal setting with M ! 1, the finest mesh
log(K) permeability field shown in Figure 2a. in this study was set to the measurement
scale of the random K field with 60 3 220
grid cells. The coarser meshes were constructed by halving the resolution of their succeeding finer meshes
in each direction. In this way, we obtained seven candidate meshes having resolutions
134; 237; 4314; 8328; 15355; 303110, and 60 3 220. The coarsest mesh is chosen such that the length
of one grid cell is smaller than the correlation length of the random field, in order to reduce the variance
V½gn ðY‘ Þ [Cliffe et al., 2011]. According to Figure 3, the correlation length is about 122.1 m, so the coarsest
resolution considered here is 4 3 14, where in the x direction 365:76=4 91:4 < 122:1 and in the y
direction 670:56=14 47:9 < 122:1. In this setting, the last five meshes were used in the calculation,
denoted as T M‘ for ‘50; 1; 2; 3; 4, respectively, and the Lmax 5 4. The five meshes are depicted in Figure 4,
where the log ðKÞ values of the coarse meshes were upscaled from the finest mesh using arithmetic mean.
More advanced upscaling methods can be applied [Wen and Go mez-Hernandez, 1996; Li et al., 2011a, 2011b], and
they are expected to be able to further improve the computational efficiency of the MLMC method. A
good upscaling method would reduce the difference of neighboring levels and correspondingly accelerates
the variance decay, so that the sample size required by the high levels will be reduced, which reduces the
computational costs of the MLMC. Application of advanced upscaling methods to the MLMC approach will
be pursued in the future research. In addition, we chose 20 interpolation points uniformly distributed in the
bounded domain ½23; 5 for the CDF estimation at the discrete points fqn g19 n50 ; and we then used piecewise line-
ar Lagrange interpolation to approximate the final CDF curve. We considered a set of e values,
e 2 f0:05; 0:04; 0:03; 0:02; 0:01; 0:009; 0:008; 0:007; 0:006; 0:005g. They represent different levels of accuracy
requirement and all are reasonable and realistic given the fact that the CDF varies in the range of ½0; 1. The com-
putational cost CM‘ was measured using the actual computational time simulated at mesh grid M‘ using the

Figure 4. Five meshes fT M‘ g4‘50 used for the MLMC simulation, where the finest mesh at level 4 has the same resolution 60 3 220 with the measurement scale in Figure 2a.

LU ET AL. AN IMPROVED MLMC METHOD 9653


Water Resources Research 10.1002/2016WR019475

simulator PFLOTRAN running in


Reference one processor, and the time is an
1 average of 100 independent runs.
Estimated (ε=0.05)
Estimated (ε=0.01) The polynomial order r used in
the smoothing function is usually
0.8
unknown; here we took a set of r
5f3; 5; 7; 9; 11g to discuss its
CDF 0.6 impact on the computational effi-
ciency of the MLMC method.

0.4 5.3. Numerical Results


The results are presented in Fig-
ures 5–7 where the MLMC
0.2 results are based on the simula-
tion with r 5 3. To evaluate the
accuracy of the Algorithm, we
0 constructed a reference CDF of
log(Q) based on 100,000 samples
−3 −2 −1 0 1 2 3 4 5 simulated on the finest mesh
log(Q) T M4 . The reference curve is illus-
trated in Figure 5, along with the
Figure 5. Reference CDF of log(Q) based on 100,000 samples simulated on the finest
mesh, and two estimated CDFs based on the Algorithm for the desired accuracy estimated CDFs using the Algo-
e50:05 and 0:01. rithm for e50:05 and 0:01. The
figure indicates that the Algo-
rithm is able to estimate the dis-
tribution with desired accuracy. As shown, with e changing from 0.05 to 0.01, the estimated CDF gets closer to
the reference, giving a more accurate approximation; and when e50:01 there is negligible difference between
the estimated and reference distributions. In fact, the calculated RMSEs of all the estimation meet the accuracy
requirement. As depicted in Figures 6c and 6d where we compare the actual RMSEs with the desired accuracy e
using their ratios, the calculated RMSE is in the range of e; actually, it is less than e in almost all cases for both
the standard MC and the MLMC estimation. This suggests the effectiveness of the Algorithm. Note that the cal-
culated RMSE cannot be exactly equal to e as it is a random variable whose value depends on the sample size
and the specific set of samples; and we only aim at RMSE being reasonable close to e. However, to achieve the
same accuracy, the computational costs of the two sampling methods are dramatically different. For example,
Figure 6a shows that with e > 0:02, the standard MC outperforms the MLMC with slightly small computational
time. In contrast, when a high accuracy is required with e < 0:02, the computational time of the MLMC method
is significantly smaller, as shown in Figure 6b. Moreover, the cost savings of the MLMC increase as the desired
accuracy increases and become very outstanding at e50:005, where the standard MC requires about 800 h and
the MLMC method only needs 480 h. The reasons causing the relative efficiency of the two sampling methods
are multiple, related to prescribed accuracy e, decay of the mean max 0nS jE½I n ðY‘ Þj, decay of the variance
max 0nS V½gn ðY‘ Þ and increase of the cost CM‘ along with the level ‘.
The e value and the max 0nS jE½I n ðY‘ Þj decay determine the maximum level L and the finest mesh T M .
pffiffiffi
For a large e and a fast decay of max 0nS jE½I n ðY‘ Þj, the condition max 0nS jE½I n ðYL Þj  e= 2 is relatively
easy to satisfy at lower levels with small L. When L is small, the variance V½gn ðY‘ Þ just starts decay and is
usually in the same order of magnitude with V½I n ðQM Þ; also the used meshes are relatively coarse at lower
levels and the cost CM‘ is usually cheap. So in this case, it is likely that the cost of the MLMC is larger than
that of the standard MC. In contrast, for a small e and a slow decay of max 0nS jE½I n ðY‘ Þj, the condition
pffiffiffi
max 0nS jE½I n ðYL Þj  e= 2 usually needs to be achieved at a large L or even with L5Lmax , and in this situ-
ation the MLMC tends to be superior to the standard MC with marked cost savings. Generally speaking, for
a high heterogeneous problem with small correlation length, the decay of max 0nS jE½I n ðY‘ Þj would be
slow and it is likely that the computational improvement of the MLMC over the standard MC would be
more pronounced [Cliffe et al., 2011]. The actual efficiency between the two sampling methods is problem
dependent, lying on the rates of the variance decay and the cost increase.

LU ET AL. AN IMPROVED MLMC METHOD 9654


Water Resources Research 10.1002/2016WR019475

1000
(a) MC (b)
MLMC

Computational time (hour)


Computational time (min)
800
6.6
3.8
3.5 5.5 36.7 37.4 347.1 600
292.6 MC
MLMC
400

200

0
0.05 0.04 0.03 0.02 0.01 0.009 0.008 0.007 0.006 0.005
ε ε

MC
2 (c) 2 (d) MLMC (r=3)
MLMC (r=5)
MLMC (r=7)
1.5 1.5 MLMC (r=9)
MLMC (r=11)
RMSE/ε

RMSE/ε

1 1

0.5 0.5

0 0
0.05 0.04 0.03 0.02 0.01 0.009 0.008 0.007 0.006 0.005
ε ε

Figure 6. Comparison of the computational efficiency between the standard MC and the MLMC methods for a set of desired accuracy e. Figures 6a and 6b compare their computational
time; (c and d) ratios of the calculated RMSE and the desired accuracy e.

Figure 7 explains the relative computational efficiency between the standard MC and the MLMC using accu-
racies e50:05; 0:03; 0:02; and 0:01 for an example. Figure 7a plots max 0nS jI~ n ðYL Þj decay with levels and
pffiffiffi
the e= 2 threshold; Figure 7b indicates that variance max 0nS V~ ½I n ðQM‘ Þ is almost constant for all the lev-
els, and max 0nS V~ ½gn ðY‘ Þ is smaller than max 0nS V~ ½I n ðQM‘ Þ and decreases with levels; Figure 7c depicts
the number of model simulations required by the standard MC for different accuracies (where each symbol
in the solid line represents the NMC for a given e) and by the MLMC (where each dashed line shows the N‘ at
levels ‘  L for a given e); and Figure 7d shows the computational time CM‘ for a single model run on
pffiffiffi
the mesh T M‘ . For example, when e50:05, L 5 1 satisfies the condition that max 0nS jI~ n ðYL Þj  e= 2 in
the Algorithm (Figure 7a). With L 5 1, the variance max 0nS V~ ½gn ðYL Þ is not effectively smaller than
max 0nS V~ ½I n ðQML Þ (Figure 7b), which results in a relatively small reduction of NL compared to NMC . As
shown in Figure 7c, for e50:05, when L 5 1, NL 5 71 and NMC 5193; the difference is 122 simulations on level
1. Moreover, Figure 7d indicates that the time of a single simulation at level 1 is very fast, suggesting that
the cost of the MLMC at level 1 be close to the total cost of the standard MC. Since the total cost of the
MLMC also includes the simulation time at the level 0, this results in the MLMC less efficient than the stan-
dard MC in the case of e50:05. On the other hand, when e50:01; L5Lmax 54 is required (Figure 7a). For
L 5 4, the variance max 0nS V~ ½gn ðYL Þ is greatly smaller than max 0nS V~ ½I n ðQML Þ (Figure 7b) and this
results in NL 5 1122 and NMC 54949 (Figure 7c); the difference is 3827 simulations on the highest level 4.

LU ET AL. AN IMPROVED MLMC METHOD 9655


Water Resources Research 10.1002/2016WR019475

Figure 7. Explanation
pffiffiffi of the reasons causing the relative computational efficiency between the standard MC and the MLMC methods. (a) max 0nS jI~ n ðY‘ Þj decay with levels (solid red
line) and the e= 2 threshold (dashed black lines) for e50:05; 0:03; 0:02; and 0:01; (b) the change of variance max 0nS V~ ½I n ðQM‘ Þ and max 0nS V~ ½gn ðY‘ Þ with levels; (c) the number of
simulations required by the standard MC (where each NMC value corresponds to a given e) and by the MLMC (where each dashed line represents N‘ for a given e); and (d) the
computational time of a single simulation on the mesh T M‘ .

The relatively large reduction of NL compared to NMC makes outstanding difference in the cost between the
MLMC and the standard MC because a single simulation on level 4 is very computationally expensive (Fig-
ure 7d). Thus, in the case of e50:01, the MLMC is significantly more efficient than the standard MC.
Overall, in this study, for e  0:03, the computational cost of the standard MC is slightly lower than
that of the MLMC. However, for this relatively low accuracy requirement, the estimated CDF is not very
accurate; as shown in Figure 5 where the estimated distribution for e50:05 deviates the reference a lot
with the absolute error up to 0.11 and relative error up to 42%. For e  0:02, the MLMC starts performing
better and gives outstanding cost savings as e  0:01. For this relatively high accuracy requirement,
the estimated CDF is quite accurate with the largest absolute error of 0.008 and relative error of 2% at
e50:01 (Figure 5). The relative efficiency between the two sampling methods is problem dependent as
analyzed above. Hence, practically it is important to use the Algorithm to compare their costs before
application.
Besides the four factors discussed above, the performance of the MLMC in estimation of the CDF is also
affected by the use of the smoothing function, the smoothing factor d and the function polynomial order r.
In the following section 6, we discuss these influences in detail for the cases with e  0:01 and give guid-
ance on the effective use of the MLMC method for the CDF approximation.

LU ET AL. AN IMPROVED MLMC METHOD 9656


Water Resources Research 10.1002/2016WR019475

Figure 8. Discuss influence of the smoothing technique on the MLMC performance. (a) The variance of the MLMC with smoothing, max 0nS V½gn ðY‘ Þ, is smaller than that without
smoothing, max 0nS V½I n ðY‘ Þ; (b) the computational time of the MLMC with smoothing is less than that without smoothing.

6. Discussion
The use of smoothing technique plays a critical role in the computational efficiency of the MLMC method. It
accelerates the variance decay by reducing the value of V½I n ðY‘ Þ to V½gn ðY‘ Þ. As shown in Figure 8a, the
value of max 0nS V~ ½gn ðY‘ Þ with smoothing is much smaller than max 0nS V~ ½I n ðY‘ Þ without smoothing,
making the variance of the MLMC remarkably lower than that of the standard MC. The reduced variance sig-
nificantly decreases the number of samples N‘ particularly at a high level ‘ with expensive model simula-
tions, resulting in a great cost saving of the MLMC. Figure 8b illustrates that the computational time of the
MLMC with smoothing is much less than that without smoothing. Even worse, without the smoothing, the
efficiency of the MLMC is inferior to that of the standard MC. Therefore, when the MLMC method is used for
CDF estimation, it is suggested using the smoothing step to improve the computational efficiency.
Two factors affect the performance of the smoothing function, the factor d and the polynomial degree r,
where the d plays a deterministic role. A larger d causes a smoother function approximation, corresponding-
ly a faster variance decay and thus greater improvement of computational efficiency in the MLMC method.
Meanwhile, a larger d also brings a greater smoothing error. In this study, we optimize the d by setting the
smoothing error equal to the defined error bound so as to maximize the computational efficiency. This is
demonstrated in Figure 9. Figure 9a shows that given the set of r and e, our do values are in the range
between 1 and 4, and the resulting smoothing errors are right below the error bound of e2 =4 as depicted in
Figure 9b. In contrast, the dG values are between 0.3 and 0.7 (Figure 9c), much smaller than the do especially
for a large r. Clearly, the small dG values are far away from the optimal because their resulting smoothing
errors are substantially below the bound as illustrated in Figure 9d.
The difference in our do and the dG results in a marked difference in the variance of V½gn ðY‘ Þ and thus a
dramatical difference in the computational cost between the two MLMC simulations. As illustrated in Fig-
ure 10a where r 5 3 and 11 are taken for an example, the values of max 0nS V~ ½gn ðY‘ Þ based on do are
smaller than those based on dG and their difference becomes larger for higher levels, indicating that our
do requires smaller N‘ than the dG especially for large ‘. This leads to a great computational cost saving of
the MLMC using the do compared to the use of the dG. As shown in Figure 10b, the computational time
by using do is much lower than that by using dG, and the saved computational cost in the use of do is very
outstanding for the high accuracy. Surprisingly, the computational cost of the MLMC with dG is even larg-
er than that of the standard MC. This suggests the vital importance of the calculation of the d, and using
our optimization technique can maximize the d and greatly improves the computational efficiency of the
MLMC method.
The influence of r on the MLMC performance is mainly through its effect on d, i.e., dependent on how d is
calculated based on r. For example, in our calculation of do in (28) and (29), a larger r leads to a larger do

LU ET AL. AN IMPROVED MLMC METHOD 9657


Water Resources Research 10.1002/2016WR019475

Figure 9. Describe the difference of the smoothing factor calculated using our optimization method (do) and based on the method in Giles et al. [2015] (dG) for a set of r values. Our do
values shown in Figure 9a are much larger than the dG shown in Figure 9c, because the do is optimized by setting the smoothing error equal to the defined error bound e2 =4, as numeri-
cally verified in Figure 9b; while the dG values are far away from the optimal because their resulting smoothing errors are way below the error bound, as demonstrated in Figure 9d.

(Figure 9a) and thus a smaller computational cost of the MLMC as shown in Figure 10b. In contrast, the dif-
ference in r causes minor difference in dG (Figure 9c), and in this case the r has the main impact. A higher-
degree polynomial approximation is closer to the actual indicator function, resulting in a slower variance
decay (Figure 10a), thus a larger r in the calculation of dG causes a relatively higher computational cost (Fig-
ure 10b). In practice, r is unknown and it is suggested choosing the one giving a large value of d.

7. Conclusions
In this work, we developed a practical multilevel Monte Carlo (MLMC) method to improve the computation-
al efficiency in estimation of cumulative distribution functions (CDFs) of a quantity of interest, coming from
numerical approximation of large-scale stochastic subsurface simulations. The application to a highly het-
erogeneous oil reservoir model indicated that the developed MLMC method greatly reduced the computa-
tional costs in the CDF approximation, with up to 320 h saving compared to the standard MC for a high-
accuracy requirement.
The performance of the developed MLMC method is affected by the use of the smoothing function. The
smoothing technique can generally improve the computational efficiency; but it depends critically on

LU ET AL. AN IMPROVED MLMC METHOD 9658


Water Resources Research 10.1002/2016WR019475

Figure 10. Discuss influence of the smoothing factor d and the smoothing polynomial order r on the MLMC performance, where do is calculated in this study and dG is determined
according to Giles et al. [2015]. (a) Our do causes larger variance decay than the dG, which makes the computational time of the MLMC based on the do much lower than that based on
dG, as shown in Figure 10b. In the calculation of do, a larger r leads to a better MLMC performance, while in the case of dG, a larger r results in a worse MLMC performance.

the smoothing factor. In this study, we designed an a posteriori optimization strategy to precisely calculate
the smoothing factor. The results revealed that it substantially reduced the computational complexity com-
pared to the method used in Giles et al. [2015] which was designed for an asymptotic case. Overall, in appli-
cation of the MLMC method for CDF approximation, it is suggested using the smoothing function with the
optimized smoothing factor, and a relatively large polynomial order for a superior performance.
The MLMC method is not a upscaling technique, and it does not require any modification of the governing
equations of the problem; nevertheless it can be combined with the advanced upscaling methods to further
improve its computational efficiency. In addition, the multilevel idea is independent of MC sampling, so it
can be integrated with other MC approaches, such as randomized quasi-Monte Carlo [Giles and Waterhouse,
2009; Graham et al., 2011] and Markov chain Monte Carlo [Efendiev et al., 2014] (C. Ketelsen et al., A hierar-
chical multilevel Markov chain Monte Carlo algorithm with applications to uncertainty quantification in sub-
surface flow, arXiv:1303.7343, 2013). Moreover, because both the multilevel and MC methods are model
independent, the developed MLMC algorithm can be applied to a wide range of problems for high-
dimensional uncertainty quantification.
Acknowledgments
The author would like to thank the
anonymous referees for their insightful
comments and suggestions that have
References
helped improve the paper. This work is Barajas-Solano, D. A., and D. M. Tartakovsky (2016), Stochastic collocation methods for nonlinear parabolic equations with random coeffi-
partially supported by the U.S. Defense cients, SIAM/ASA J. Uncertainty Quantification, 4(1), 475–494.
Advanced Research Projects Agency, Barth, A., C. Schwab, and N. Zollinger (2011), Multilevel Monte Carlo finite element method for elliptic PDEs with stochastic coefficients,
Defense Sciences Office under Numer. Math., 119, 123–161.
contract HR0011619523; the U.S. Boso, F., and D. M. Tartakovsky (2016), The method of distributions for dispersive transport in porous media with uncertain hydraulic prop-
Department of Energy, Office of erties, Water Resour. Res., 52, 4700–4712, doi:10.1002/2016WR018745.
Science, Office of Advanced Scientific Charrier, J., R. Scheichl, and A. L. Teckentrup (2013), Finite element error analysis of elliptic PDEs with random coefficients and its applica-
Computing Research, Applied tion to multilevel Monte Carlo methods, SIAM J. Numer. Anal., 51(1), 322–352.
Mathematics program under contracts Christie, M. A., and M. J. Blunt (2001), Tenth SPE comparative solution project: A comparison of upscaling techniques, SPE J, 4(4), 308–317.
ERKJ259 and ERKJ314; the U.S. Air Cliffe, K. A., M. B. Giles, R. Scheichl, and A. L. Teckentrup (2011), Multilevel Monte Carlo methods and applications to elliptic PDEs with ran-
Force Office of Scientific Research dom coefficients, Comput. Visualization Sci., 14, 3–15.
under grants 1854-V521-12; the U.S. Deutsch, C. V., and A. G. Journel (1998), GSLIB: Geostatistical Software Library and User’s Guide, 2nd ed., Oxford Univ. Press, New York.
National Science Foundation, Dodwell, T. J., C. Ketelsen, R. Scheichl, and A. L. Teckentrup (2015), A hierarchical multilevel Markov Chain Monte Carlo algorithm with
Computational Mathematics program applications to uncertainty quantification in subsurface flow, SIAM/ASA J. Uncertainty Quantification, 3(1), 1075–1108.
under awards 1620280 and 1620027; Efendiev, Y., B. Jin, M. Presho, and X. Tan (2014), Multilevel Markov Chain Monte Carlo method for high-contrast single-phase flow prob-
and by the Laboratory Directed lems, Commun. Comput. Phys., 17(1), 259–286.
Research and Development program Franssen, H. J. H, A. Alcolea, M. Riva, M. Bakr, N. van der Wiel, and F. Stauffer (2009), A comparison of seven methods for the inverse model-
at the Oak Ridge National Laboratory, ing of groundwater flow: Application to the characterization of well catchments, Adv. Water Resour., 32(6), 851–72.
which is operated by UT-Battelle, LLC., Ghanem, R. (1998), Scales of fluctuation and the propagation of uncertainty in random porous media, Water Resour. Res., 34, 2123–2136.
for the U.S. Department of Energy, Ghanem, R. (1999), Ingredients for a general purpose stochastic finite elements implementation, Comput. Methods Appl. Mech. Eng., 168,
under contract DE-AC05-00OR22725. 19–34.

LU ET AL. AN IMPROVED MLMC METHOD 9659


Water Resources Research 10.1002/2016WR019475

Giles, M. (2008), Multilevel Monte Carlo path simulation, Oper. Res., 56(3), 607–617.
Giles, M., and C. Reisinger (2012), Stochastic finite difference and multilevel Monte Carlo for a class of SPEDs in finance, SIAM J. Finan.
Math., 1(3), 575–592.
Giles, M., and B. J. Waterhouse (2009), Multilevel quasi-Monte Carlo path simulation, Radon Ser. Comput. Appl. Math., 8, 1–18.
Giles, M., T. Nagapetyan, and K. Ritter (2015), Multilevel Monte Carlo approximation of distribution functions and densities, SIAM/ASA J.
Uncertainty Quantification, 3, 267–295.
Graham, I. G., F. Y. Kuo, D. Nuyens, R. Scheichl, and I. H. Sloan (2011), Quasi-Monte Carlo methods for elliptic PDEs with random coefficients
and applications, J. Comput. Phys., 230(10), 3668–3694.
Goovaerts, P. (2001), Geostatistical modeling of uncertainty in soil science, Geoderma, 103, 3–26.
Guadagnini, A, and S. P. Neuman (1999a), Nonlocal and localized analyses of conditional mean steady state flow in bounded, randomly
nonuniform domains: 1. Theory and computational approach, Water Resour. Res., 35, 2999–3018.
Guadagnini, A, and S. P. Neuman (1999b), Nonlocal and localized analyses of conditional mean steady state flow in bounded, randomly
nonuniform domains: 2. Computational examples, Water Resour. Res., 35, 3019–3039.
Gunzburger, M., C. G. Webster, and G. Zhang (2014), An adaptive wavelet stochastic collocation method for irregular solutions of partial dif-
ferential equations with random input data, in Sparse Grids and Applications—Munich, pp.137–170, Springer, Switzerland.
Heinrich, S. (2001), Multilevel Monte Carlo methods, in Lecture Notes Computer Science, vol. 2179, pp. 3624–3651, Springer, Philadelphia, Pa.
Helmig, R. (1997), Multiphase Flow and Transport Processes in the Subsurface, 367 pp., Springer, Berlin.
Hill, M. C., and C. Tiedeman (2007), Effective Groundwater Model Calibration: With Analysis of Data, Sensitivities, Predictions, and Uncertainty,
480 pp., John Wiley, Hoboken, N. J.
Laloy, E., B. Rogiers, J. A. Vrugt, D. Mallants, and D. Jacques (2013), Efficient posterior exploration of a high-dimensional groundwater model
from two-stage Markov Chain Monte Carlo simulation and polynomial chaos expansion, Water Resour. Res., 49, 2664–2682, doi:10.1002/
wrcr.20226.
Li, H., and D. Zhang (2007), Probabilistic collocation method for flow in porous media: Comparisons with other stochastic methods, Water
Resour. Res., 43, W09409, doi:10.1029/2006WR005673.
Li, L., H. Zhou, and J. J. Gomez-Hernandez, (2011a), A comparative study of three-dimensional hydraulic conductivity upscaling at the mac-
rodispersion experiment (MADE) site, Columbus air force base, Mississippi (USA), J. Hydrol., 404(3-4), 278–293.
Li, L., H. Zhou, and J. J. Gomez-Hernandez (2011b), Transport upscaling using multi-rate mass transfer in three-dimensional highly hetero-
geneous porous media, Adv. Water Resour., 34(4), 478–489.
Lu, D., M. Ye, and S. P. Neuman (2011), Dependence of Bayesian model selection criteria and fisher information matrix on sample size,
Math. Geosci., 43, 971–993.
Ma, X., and N. Zabaras (2009), An adaptive hierarchical sparse grid collocation algorithm for the solution of stochastic differential equa-
tions, J. Comput. Phys., 228(8), 3084–3113.
Mill, R. T., G. E. Hammond, P. C. Lichtner, V. Sripathi, G. Mahinthakumar, and B. F. Smith (2007), Modeling subsurface reactive flows using
leadership-class computing, J. Phys., 180, 1–10, doi:10.1088/1742-6596/180/1/012062.
Mukherjee, M. (2013), Instrumented permeable blankets for estimating subsurface hydraulic conductivity and confirming numerical mod-
els used for subsurface liquid injection, PhD thesis, Mich. State Univ., East Lansing.
Nobile, F., and F. Tesei (2015), A multilevel Monte Carlo method with control variate for elliptic PDEs with log-normal coefficients, Stochas-
tic Partial Differential Equation Anal. Comput., 3(3), 398–444.
Nobile, F., R. Tempone, and C. G. Webster (2008a), A sparse grid stochastic collocation method for partial differential equations with ran-
dom input data, SIAM J. Numer. Anal., 46(5), 2309–2345.
Nobile, F., R. Tempone, and C. G. Webster (2008b), An anisotropic sparse grid stochastic collocation method for partial differential equa-
tions with random input data, SIAM J Numer. Anal., 46(5), 2411–2442.
Pasetto, D., A. Guadagnini, and M. Putti (2014), A reduced-order model for Monte Carlo simulations of stochastic groundwater flow, Com-
put. Geosci., 18, 157–169, doi:10.1007/s10596-013-9389-4.
Smith, R. (2013), Uncertainty Quantification: Theory, Implementation, and Applications, 400 pp., Soc. for Ind. and Appl. Math., Philadelphia, Pa.
Tartakovsky, D. M., (2013), Assessment and management of risk in subsurface hydrology: A review and perspective, Adv. Water Resour., 51,
247–260.
Teckentrup, A. L. (2013), Multilevel Monte Carlo methods and uncertainty quantification, PhD thesis, Univ. of Bath, Bath, U. K.
Teckentrup, A. L., R. Scheichl, M. B. Giles, and E. Ullmann (2013), Further analysis of multilevel Monte Carlo methods for elliptic PDEs with
random coefficients, Numer. Math., 125, 569–600.
Tversky, A., and D. Kahneman (1992), Advances in prospect theory: Cumulative representation of uncertainty, J. Risk Uncertainty, 5,
297–323.
Wen, X., and J. G omez-Hernandez (1996), Upscaling hydraulic conductivities in heterogeneous media: An overview, J. Hydrol., 183, ix–xxxii.
Zhang, D. (2002), Stochastic Method for Flow in Porous Media: Coping with Uncertainties, Academic, San Diego, Calif.
Zhang, D., and Z. Lu (2004), An efficient, high-order perturbation approach for flow in random porous media via Karhunen-Loeve and
polynomial expansions, J. Comput. Phys., 194(2), 773–794, doi:10.1016/j.jcp.2003.09.015.
Zhang, G., D. Lu, M. Ye, M. Gunzburger, and C. Webster (2013), An adaptive sparse-grid high-order stochastic collocation method for
Bayesian inference in groundwater reactive transport modeling, Water Resour. Res., 49, 6871–6892, doi:10.1002/wrcr.20467.
Zhang, G., C. Webster, M. Gunzburger, and J. Burkardt (2016), Hyperspherical sparse approximation techniques for high-dimensional
discontinuity detection, SIAM Rev., 58, 517–551.

LU ET AL. AN IMPROVED MLMC METHOD 9660

You might also like