Journal of Computational Physics 386 (2019) 405–427

Contents lists available at ScienceDirect

Journal of Computational Physics

www.elsevier.com/locate/jcp

Uncertainty quantification methodology for hyperbolic

systems with application to blood flow in arteries
M. Petrella a,∗ , S. Tokareva b , E.F. Toro c
a
Seminar for Applied Mathematics, Department of Mathematics, ETH Zürich, Rämistrasse 101, 8092 Zürich, Switzerland
b
Los Alamos National Laboratory, Theoretical Division, Applied Mathematics and Plasma Physics Group (T-5), Los Alamos, NM, USA
c
Laboratory of Applied Mathematics, DICAM, University of Trento, via Mesiano 77, 38123 Mesiano (Trento), Italy

a r t i c l e i n f o a b s t r a c t

Article history: We present a Stochastic Finite Volume - ADER (SFV-ADER) methodology for Uncertainty
Received 1 August 2018 Quantification (UQ) in the general framework of systems of hyperbolic balance laws
Received in revised form 16 February 2019 with uncertainty in parameters. The resulting method is weakly intrusive, meaning that
Accepted 22 February 2019
deterministic solvers need minor modifications to include random contributions, and has
Available online 1 March 2019
no theoretical accuracy barrier. An illustration of the second-order version for the viscous
Keywords: Burgers equation with uncertain viscosity coefficient is first given in detail, under different
Uncertainty quantification deterministic initial conditions; the attainment of the theoretically expected convergence
Hyperbolic equations rate is demonstrated empirically; results and features of our scheme are discussed. We
Stochastic finite volume method then extend the SFV-ADER method to a non-linear hyperbolic system with source terms
ADER schemes that models one-dimensional blood flow in arteries, assuming uncertainties in a set of
Blood flow parameters of the problem. Results are then compared with measurements in a well-
defined 1:1 experimental replica of the network of largest arteries in the human systemic
circulation. For a 99%-confidence level, severe influence of parameter fluctuations is seen in
pressure profiles, while secondary effects on flow rate profiles become visible, particularly
in terminal branches. Results suggest that the proposed methodology could successfully be
applied to other physical problems involving hyperbolic balance laws.
© 2019 Elsevier Inc. All rights reserved.

1. Introduction

Uncertainty Quantification is the science of quantitative characterization and reduction of uncertainty in real-world appli-
cations. Sources of uncertainty may have different nature [49,54], including incompleteness of datasets, model deficiencies,
irregular boundaries or even parametric fluctuations. The quantitative assessment of parametric fluctuations impact on solu-
tions of a given Initial-Boundary Value Problem (IBVP) is commonly referred to as parametric Uncertainty Quantification (pUQ)
[13], and is the focus of this work. UQ has received increasing attention in recent years, due to its possibility to address
both parameter variability and model shortcomings. Indeed, by providing a band illustrating the variability of the quantity
of interest, as function of parameters, it identifies regions of physical space that are sensitive/insensitive to parameter fluc-
tuations. Existing methodologies can be divided essentially into two disjoint classes of methods: intrusive and non-intrusive.
A method for UQ is called non-intrusive if it allows to reuse deterministic codes and algorithms (i.e. no uncertainty/random
sources) without any structural change. However secondary this requirement may look, it becomes of practical interest in

* Corresponding author.
E-mail address: [email protected] (M. Petrella).

https://doi.org/10.1016/j.jcp.2019.02.013
0021-9991/© 2019 Elsevier Inc. All rights reserved.
406 M. Petrella et al. / Journal of Computational Physics 386 (2019) 405–427

complex situations where re-implementation is not an option. In particular, non-intrusive methodologies are often associ-
ated to the allocation of several realizations of the random inputs and a subsequent post-processing procedure to compute
statistical moments. The way every method allocates the ensemble of samples and the post-processing procedure defines
different UQ methods. In contrast, intrusive approaches expand the sought solution on a basis family, and derive algebraic
equations or deterministic partial differential equations (PDEs) for statistical moments of interest. Many numerical methods
have been proposed in the literature to treat uncertainty. Stochastic Differential Equations (SDE) have played an important
role in their development, and a first example of this is the so called perturbation method [61]. This approach expands
stochastic quantities around their mean in Taylor series and allows for an estimation of low order statistics; however, it
is limited to small perturbations and does not readily provide information on high-order statistics. Indeed, the resulting
system of equations becomes extremely complicated beyond second-order expansions, making it of very limited practical
use for general purposes. Another approach is based on expanding the inverse of the stochastic operator in a Neumann
series, but this too is limited to small fluctuations, see [13,20,21,73]. One possibility to obtain general applicability is to
use non-intrusive and sampling methods, among which, well-known schemes are the Monte Carlo (MC) based methods.
MC methods sample (by means of a pseudo-random number generator) the space of possible outcomes, and the underlying
deterministic PDEs are solved for each sample, resulting in a very robust, easily parallelizable approach. However, it is well-
known that MC based methods are characterized by a slow rate (1/2) of convergence and therefore a high-computational
cost [41,13,20,73]. In particular, accurate quantification of uncertainty requires a large number of samples and direct MC
methods combined with available methods for hyperbolic systems of conservation or balance laws in several space dimen-
sions become very costly, even with a moderately large number of random inputs. In fact, for each realization, the direct
problem simulation CPU time is independent of the number of random sources, making the overall time depend on the
number of realizations needed [60]. Therefore MC method is often considered as last resort. One attempt to address the
computational efficiency issue is the Multi-Level Monte Carlo (MLMC) version, which essentially applies the MC method
on a nested sequence of triangulations of the domain [41,42,13]. Indeed, the MLMC method is also non-intrusive and less
costly than MC, showing promising results for multivariate UQ. MLMC method is not the only way to enhance the purely MC
method, some variants are variance reduction techniques such as importance sampling, control variates, stratified sampling,
correlated sampling, and conditional Monte Carlo or quasi Monte Carlo methods, see [27]. However, the latter ones need
additional a-priori analysis, smoothness assumptions and/or knowledge about the second moment of the unknown solution,
which may rarely be available [13].
Another popular and powerful class of non-intrusive methodologies for the propagation of uncertainty is the Stochastic
Collocation (SC) Method. SC allocates a large ensemble of realizations of the random vector ξ , expands the solution on a
nodal Lagrangian basis where the coefficients of the expansion remain unknown [13]. Evaluation of the expansion at sample
points, implies that the sought solution is fully determined by the coefficients of the series. One can then reinterpret the SC
method as a non-intrusive method that requires to compute the deterministic solutions collocated at sample points. Since
statistical moments can be approximated via quadrature rules, a particularly convenient choice are the nodes for Gauss
quadratures depending on probability densities. Techniques for the propagation of uncertainty, such as stochastic collocation
methods, have proven to be powerful approaches that are now routinely used in computations [7,13,36,39,48]. However,
these methods implicitly require that outputs of interest vary smoothly with respect to uncertain input parameters. When
this is not the case, these methods may exhibit a significant deterioration in accuracy, see [13] for details. This results
in a very limiting fact for non-linear hyperbolic conservation/balance laws, where solutions typically have discontinuities.
A remarkable fact concerning regularity of statistical quantities has been shown recently [55,63]: statistical quantities ex-
hibit additional smoothness as compared to deterministic solutions resulting from each realization of the random input,
depending on the probability density function. However, when discontinuities are present in stochastic space, deterioration
in computations still arise [13]. An alternative technique utilizing localized piecewise approximations combined with local-
ized subscale recovery has been proposed to significantly improve the quality of calculated statistics when discontinuities
are present [9,10]. The success of this localized technique motivates the development of the HYbrid Global and Adaptive
Polynomial (HYGAP) method [13]. HYGAP employs a high accuracy global approximation when the solution data varies
smoothly in a random variable dimension and local adaptive polynomial approximation with local postprocessing when the
solution is non-smooth. The HYGAP method also permits an estimate of the error but a formal proof of reliability is still
lacking for non-smooth data. Other popular and spectral-representation based techniques are the generalized Polynomial
Chaos (gPC) expansion - firstly proposed by Wiener [79] - and the Stochastic Galerkin (SG) discretization [13,29,37,39,63,
73–75,77,78]. These similar approaches firstly determine a basis family, then expand the sought solution on these and then
insert the resulting series in the initial-boundary value problem (IBVP) deriving equations for the coefficients of the expan-
sion, by Galerkin projection, along each basis vector. More details of the underlying mathematical theory can be found in
[26], whereas the procedure can be found in [13,63,73–75].
The gPC method chooses basis families so that each vector is orthogonal with respect to the probability density function,
whereas SG approach favours a simpler basis functions family, typically polynomials and not necessarily orthogonal with
respect to the probability density function. In particular, if the basis family is chosen to be orthogonal with respect to the
probability density function, each polynomial family has an optimal, associated random distribution, e.g. Hermite polynomi-
als for Gaussian distributions, Charlier polynomials for Poisson distributions, Laguerre polynomials for Gamma distributions,
Jacobi polynomials for Beta distributions, etc. This results in a fast convergence rate, according to the so-called Askey scheme
 M. Petrella et al. / Journal of Computational Physics 386 (2019) 405–427 407

[73–75]. In case of less common distributions, an optimal gPC expansion can always be found by constructing polynomials
via orthogonalization procedures, e.g. Gram-Schmidt [80].
Alternatively, the SG method relaxes the hereby described methodology decoupling the choice of the polynomial family
from the orthogonality condition. Compared to MC methods, gPC and Galerkin methods provide a more accurate approxima-
tion due to their spectral representation property, with comparative computational costs. The problem with this approach is
the choice of the family of polynomials. Indeed, the computational cost increases dramatically fast as the number of random
variables r or the maximal order of polynomials p increases, which is known as the curse of dimensionality [20]. Further-
more, although the deterministic version shows some promise, gPC and stochastic Galerkin methodologies suffer from the
disadvantage that they are highly intrusive [72]. As a consequence, non-intrusive variants have been recently developed,
see [37,38,48,11,31]. Basically two different approaches have been formulated: (i ) the projection method, which is based
on a numerical evaluation of the Galerkin integrals [37,38,48] and (ii ) the point collocation or stochastic response surfaces
method, which uses a linear system or regression based on a selected set of points [11,31,32]. A limitation of most existing
methodologies for hyperbolic equations is the exclusion of source terms. The presence of such source terms has theoreti-
cal and practical implications in both the mathematical properties of the IBVP and the numerical schemes to be used to
obtain approximate solutions. A desirable feature of numerical schemes for balance laws is the fact that in the absence of
time derivatives, they must respect the correct balance between the advective term (the flux) and the source term. These
schemes are commonly referred to as well-balanced schemes. Recently, also a blend of gPC and Galerkin based methods was
employed to modify well-balanced schemes, in order to include source terms. Examples of this are [34,22]. Another way
of circumventing this limitation is to combine the use of piecewise defined polynomials with the multi-level strategy, with
promising results, see [78]. In general, the application of intrusive PC in commercial codes is very limited [13].
Recently, emerging from work by Abgrall [1,2], Tokareva [63] and Mishra and Schwab [64], a new methodology to deal
with low parametric regularity has been proposed, the Stochastic Finite Volume (SFV) method. The original SFV approach re-
sults in a semi-discrete methodology for conservation laws and its approach to UQ is based on the application of high-order
deterministic methods in a higher dimensional approximation space which includes both physical and parametric-stochastic
dimensions. Amongst the advantages of the SFV method are the high-order approximation (through ENO/WENO recon-
struction, for example) in stochastic space, the weak intrusiveness and the possibility to reach higher convergence rates.
From a computational point of view, for a reasonable number of random inputs, the SFV method has been shown to be
more efficient, in terms of error-to-work estimates, with respect to the MLMC method -[63] Section 4.4-, the most efficient
non-intrusive method up to date, to the best of our knowledge. Moreover, when dealing with blood-flow model like the
ones proposed in [40,4] simulations must be run for sufficiently large time in order to reach periodic behaviours, which
highlights the importance of time savings for such applications. This combined with the typical large number of samples
required from the MLMC strategy will dramatically increase the overall time for multivariate uncertainty quantification. It
is still an open question, which method among MLMC and SFV provides the more efficient strategy for biomedical applica-
tions, especially if evenly parallelized. The SFV method offers the possibility to preserve the desired high-order of accuracy
in the whole computational domain, i.e. physical and stochastic, making it an attractive candidate to develop arbitrarily high
order schemes. As is well-known, high-order methods are more efficient (at the cost of complexity) as compared to their
low-order versions, making them essential ingredients for the study of many highly complex situations where low-order
methods would require prohibitively high computational costs for small-error results.
In this work we adapt the SFV method to combine it with the ADER approach [68], first proposed for physical space
discretizations. The method results in a one-step, fully discrete, finite volume scheme for hyperbolic balance laws with
uncertain parameters, with no theoretical accuracy barrier in physical space, time and stochastic space. The rest of this
paper is organized as follows: we first recall the ADER approach and introduce the methodology in its general framework
for several random parameters; in Section 3 we illustrate the SFV-ADER method in its second-order version for a model
equation, the viscous Burgers equation with uncertain viscosity coefficient. The scheme is then applied to a non-linear
hyperbolic system of balance laws that constitutes a one-dimensional mathematical model for blood flow in elastic vessels,
assuming uncertainty in parameters. Conclusions are drawn in section 5.

2. SFV-ADER finite volume method

In this section we present our approach to UQ involving hyperbolic balance laws in one space dimension. In essence, we
adopt the SFV methodology [1,2,63,64] in the frame of the finite volume ADER schemes [68].

2.1. Stochastic formulation

A common procedure in pUQ to include uncertainty is to translate it into a stochastic contribution as follows: consider
the IBVP
⎧  λ  λ
⎪
⎨∂t q (x, t ) + L(x, t , λ) q = S (x, t , λ) q , x ∈ Dx ⊂ R , t > 0,
λ d

q (x, 0) = h(x, λ),

λ x ∈ Dx ⊂ Rd , (1)
⎪
⎩ λ
q (x, t ) = g (x, t , λ), x ∈ ∂ Dx ⊂ Rd , t > 0,
408 M. Petrella et al. / Journal of Computational Physics 386 (2019) 405–427

where the space-differential operator L and the right-hand side S may depend on r ∈ N parameters λ = (λ1 , . . . , λr ). We
denote by qλ (x, t ) = q(x, t , λ) the solution of IBVP (1) corresponding to parameters λ = λ ∈ D y . Given λ1 = λ2 , one expects
the corresponding solutions to be different, in general. In many situations of practical interest it might be too ambitious to
expect exactness of λ. We are therefore led to include uncertainty in the IBVP (1): typically one assumes a certain degree
of uncertainty on parameters λ, which can be suitably described by statistical distributions [49,13]. For example, we can
model it by replacing every uncertain parameter λi by a random variable ξi (ω) = (1 + εi y (ω))λi , where εi is the maximum
value of uncertainty/error assumed for the parameter λi , y is a uniformly distributed random variable between −1 and 1
( y ∼ Unif[−1, 1]) and ω is an element of the set of events , which in turn is such that ξ () ⊂ D y . Note that in this case the
resulting random vector ξ = (ξ1 , . . . , ξr ) is composed by random variables ξi :  −→ [(1 − εi )λi , (1 + εi )λi ] ⊂ R. We denote
by ρ the probability density function of the random vector ξ . The stochastic formulation of (1) then reads
⎧
⎪
⎨∂t q(x, t , ξ (ω)) + L(x, t , ξ (ω)) [q] = S (x, t , ξ (ω)) [q] , x ∈ Dx , t > 0, ω ∈ ,
q(x, 0, ξ (ω)) = h(x, ξ (ω)), x ∈ Dx , ω ∈ , (2)
⎪
⎩
q(x, t , ξ (ω)) = g (x, t , ξ (ω)), x ∈ ∂ Dx , t > 0, ω ∈ .

Every possible outcome of the random vector ξ is called a realization, and a probabilistic profile [63] is the family of solutions
of (1) generated by all the possible realizations of the random vector ξ .
One possibility to quantify the uncertainty for solutions of (1) is offered by the Chebyshev inequality, which states that
a random variable X ∈ L 2 (, P) with expectation μ and variance σ 2 verifies the inequality

1
P[| X − μ| ≥ θ σ ] ≤ , ∀ θ > 0.
θ2
In particular, by denoting I θ = [μ − θ σ , μ + θ σ ], the following holds

1
P[ X ∈ I θ ] ∈ 1 − 2 , 1 , (3)
θ
which yields non-trivial results only in the case θ > 1. Relation (3) is connected to the notion of confidence interval: an
heuristic interpretation of (3) is that we can assume the random variable X to take values in I θ for at least (1 − θ12 )% of
the cases. Therefore, by fixing a large value of θ , one produces a band I θ which contains “the most probable” outcomes
of the random variable X . In real-world applications, random variables are represented by functionals of the random field
q(x, t , ξ ) solving (2), or even q(x, t , ξ ) itself. In the latter case, the band I θ identifies the variation interval of solutions of
(1), as parameters λ vary. Similarly, the confidence level 1 − θ12 is a measure of “reliability” of the band I θ . Hence, a major
task in UQ is to compute low-order statistics of the random variable q(x, t , ξ (ω)), like expectation

E [q] (x, t ) := q(x, t , y)ρ (y) dy, ∀ (x, t ) ∈ Dx × [0, ∞), (4)
Rr

and variance

V [q] (x, t ) := E (q − E [q])2 = E q2 (x, t ) − (E[q](x, t ))2 , ∀ (x, t ) ∈ Dx × [0, ∞). (5)

From both the practical and the theoretical point of view, attention must be put to the considered probability density func-
tion ρ . As shown in [63], regularity of statistical quantities of conservation laws depend crucially on the smoothness of
the function ρ . This highlights also the strong dependence of statistical quantities form on the corresponding probability
density function. In this work we expose the procedure using uniform distribution, but the methodology is prone to modifi-
cations to other probability density functions ρ . For example, a popular one is the one associated to the normal distribution,
which requires the evaluation of integrals of the form (4) − (5), over an unbounded domain. This is typically performed by
truncating the Gaussian probability density function and localizing the integration on a significative domain. More details
concerning the treatment of such probability density functions can be found in [63,64]. Finally, notice that SFV method
allows to consider different distributions for each random source.

2.2. Finite volume ADER schemes

Building upon the work of Toro et al. [68], there have been many further developments of the ADER approach in the
last decade, see for example [71,62,35,25,24,8]. One of the advantages of this approach is its capability to produce one-step,
fully discrete non-linear schemes of arbitrary order of accuracy in both space and time. Let us consider the system of m ∈ N
one-dimensional hyperbolic balance laws

∂t Q(x, t ) + ∂x F(Q(x, t ), x, t ) = S(Q(x, t ), x, t ), (6)

 M. Petrella et al. / Journal of Computational Physics 386 (2019) 405–427 409

Fig. 1. Schematic representation of the ADER approach to compute the numerical flux at the cell interface for the case of a scalar balance law. From cell
averages one generates reconstruction polynomials that in turn define the initial conditions for a generalised Riemann problem whose solution permits the
evaluation of the time-averaged flux at the interface.

where x ∈ [a, b] ⊂ R and t > 0. The vector F is the physical flux and S is the source term. We discretize the physical space-
time domain [a, b] × [0, T out ] into finite volumes V i = [xi − 1 , xi + 1 ] × [t n , t n+1 ], with spacial mesh width x = xi + 1 − xi − 1
2 2 2 2
and time step t = t n+1 − t n . By exact integration of (6) over the space-time control volume V i we obtain the exact formula

t
Qni +1 = Qni − F i + 1 − Fi − 1 + tSi , (7)
x 2 2

with definitions
⎫
x
i+ 1 t n +1 ⎪
⎪
2
⎪
⎪
1 1 ⎪
Qni = Q(x, t n ) dx, Fi + 1 = F(Q(xi + 1 , t ), xi + 1 , t ) dt , ⎪
⎪
⎪
x 2 t 2 2 ⎪
⎪
x tn ⎪
⎬
i− 1
2
(8)
t n +1
x
i+ 1 ⎪
⎪
2 ⎪
⎪
1 ⎪
⎪
Si = S(Q(x, t ), x, t ) dxdt . ⎪
⎪
x t ⎪
⎪
t n xi − 1
⎪
⎭
2

Numerically, formula (7) is interpreted approximately, given that integrals (8) must be computed approximately. For con-
venience, we adopt an approximate interpretation of (7) as a finite volume scheme. The ADER approach relies on accurate
approximations of integral (8) for the numerical flux Fi + 1 and the numerical source Si . The spatial integral Qni is only
2
required at the initial time, as formula (7) provides the updating procedure for later times. The ADER procedure is based
on essentially three ingredients: high-order spatial reconstruction (one per time step), solution of a Generalized Riemann
Problem (GRP) at cell interface x = xi + 1 to compute the numerical flux and evaluation of the numerical source; see Fig. 1.
2
For details on the ADER approach and the solution of GRPs see Chapters 19 and 20 of [66]. The aim of this work is to adapt
the ADER approach to hyperbolic systems of balance laws with uncertainty, preserving the theoretically arbitrary order of
accuracy.

2.3. SFV-ADER general framework

Let us consider (6), where both the physical flux and the source term depend on r ∈ N parameters λ = (λ1 , . . . , λr ),
whose values are uncertain. Following the procedure presented in Section 2.1, one can convert the hyperbolic system of
balance laws into a system of hyperbolic stochastic balance laws, by replacing every uncertain parameter λi with a random
variable ξi (ω). We are then interested in computing statistical moments of the quantity Q(x, t , ξ (ω)). To this end, let us
consider the one-dimensional system of balance laws

∂t Q(x, t , ξ (ω)) + ∂x F(Q(x, t , ξ (ω)), ξ (ω)) = S(Q(x, t , ξ (ω)), ξ (ω)), (9)

410 M. Petrella et al. / Journal of Computational Physics 386 (2019) 405–427

with r random sources ξ (ω) = [ξ1 (ω), . . . , ξr (ω)] T such that ξi :  −→ [c i , di ] ⊂ R for each i = 1, . . . , r. By parametrizing
random inputs in (9) via the random vector y = ξ (ω) and rewriting the system of stochastic balance laws (9) in parametric
form, we have

∂t Q(x, t , y) + ∂x F(Q(x, t , y), y) = S(Q(x, t , y), y). (10)

 each i = 1, . . . , r, divide (discretize) the

For i = [c i , d i ] of the random variable ξi (ω ) into N i subintervals (cells)
 target set I
(i )
y , y (i ) 1 with barycentres y j =
(i ) 1
y
(i )
+ y (i ) 1 , j ∈ {1, . . . , N i }. We therefore denote the centre of the stochastic
j − 12 j+ 2 2 j + 12 j− 2
cell as

(1 ) (r )
yj = [ y j , . . . , y j ] T , j = ( j 1 , . . . , j r ), ∀ j i ∈ {1, . . . , N i }, i = 1, . . . , r .
1 r

We then define the weighted stochastic cell volumes as

y y
j1 + 1 jr + 1
2 2

yj := ··· ρ (y) dy, (11)

y y
j1 − 1 jr − 1
2 2

where ρ is the (joint) probability density function of the random vector ξ . Furthermore, we shall make use of the following
notation
y y y
j+ 1 j1 + 1 jr + 1
2 2 2

T(x, t , y) dy := ··· T(x, t , y) dy, (12)

y y y
j− 1 j1 − 1 jr − 1
2 2 2

for any integrable function T on the stochastic cell j. A finite volume scheme to approximate solutions of (10) is given
by multiplying (10) by ρ and then integrating the resulting vector equation on the control volume V = [xi − 1 , xi + 1 ] ×
2 2

[t n , t n+1 ] × [ y
(1) (1) (r ) (r )
, y 1 ] × . . . × [ y 1 , y 1 ] to obtain the exact formula
j 1 − 12 j1 + 2 jr − 2 jr + 2

t
Qni ,+
j
1
= Qni,j − Fi + 1 ,j − Fi − 1 ,j + tSi ,j , (13)
x 2 2

with definitions
x y
i+ 1 j+ 1
2 2
1
Qni,j = Q(x, t n , y)ρ (y) dydx, (14)
x yj
x y
i− 1 j− 1
2 2
⎫
t n +1
y
j+ 1 ⎪
⎪
2
⎪
⎪
1 ⎪
⎪
Fi + 1 ,j = F(Q(xi + 1 , t , y), y)ρ (y) dydt , ⎪
⎪
2 t yj 2 ⎪
⎪
t n yj− 1 ⎪
⎬
2
(15)
t n +1
x
i+ 1
y
j+ 1 ⎪
⎪
2 2 ⎪
⎪
1 ⎪
Si ,j = S(Q(x, t , y), y)ρ (y) dydxdt . ⎪
⎪
⎪
t x yj ⎪
⎪
t n xi − 1 yj− 1
⎪
⎭
2 2

Notice that, recalling the deterministic version of the above quantities, these new formulation coincides with the stochastic-
integral average of the corresponding deterministic ones. In fact, from equations (11) and (15), we can write

y y
j+ 1 j+ 1
2 2
1 1
Fi + 1 ,j = Fi + 1 (y)ρ (y) dy, Si ,j = Si (y)ρ (y) dy, (16)
2 yj 2 yj
y y
j− 1 j− 1
2 2

where Fi + 1 (y) and Si (y) are the deterministic numerical flux and deterministic numerical source (8) collocated at stochastic
2
point y = y, respectively.
 M. Petrella et al. / Journal of Computational Physics 386 (2019) 405–427 411

In UQ we are interested in computing statistical quantities, deriving from k-th moments of the probability distribution

E Qk (x, t ) = Qk (x, t , y)ρ (y) dy, k ∈ N. (17)

ξ ()

Since ξ () = j [yj− 12 , yj+ 12 ], a formula to calculate such quantities results in

y
j+ 1
 k
2

E Qk (xi , t n ) = Qk (xi , t n , y)ρ (y) dy = Qk (xi , t n , y)ρ (y) dy ≈ Qni,j yj , ∀k ∈ N (18)
ξ () j y j
j− 1
2

where Qni,j is computed by approximating (13). Hence, we are led to draw approximations for integrals (15).

2.4. The SFV-ADER numerical flux

The numerical flux (15) is computed as the time-stochastic integral average of the physical flux at physical cell interface
x = xi + 1 , for which one seeks a solution at the cell interface. This is given by the solution of the following generalised
2
Rieman problem (GRP)

⎧
⎪ PDE : ∂t Q + ∂x F(Q, y) = S(Q, y), x ∈ D x , t > 0, y ∈ D y,
⎨ 
GRP(Q L (x, y), Q R (x, y)) = Q L (x, y), x < 0, (19)
⎪
⎩ ICs : Q(x, 0, y) = y ∈ D y,
Q R (x, y), x > 0,

where Q L (x, y) and Q R (x, y) are smooth vector-valued functions given by a polynomial reconstruction procedure. As we as-
sume Q L (x, y) and Q R (x, y) to be polynomials of degree p ∈ N, we are offered the possibility to approximate the integrals in
(16) via suitable quadrature rules, e.g. Gaussian quadrature, for which a ( p + 1)/2-points rule suffices. We then approximate
the stochastic integrals by means of a quadrature rule with nodes zk and weights w k , which are defined from the tensor
products of the corresponding one-dimensional quadrature rules, that is

y
j+ 1
2
1 1 
Fi + 1 ,j = Fi + 1 (y)ρ (y) dy ≈ Fi + 1 (zk )ρ (zk ) w k . (20)
2 yj 2 yj 2
y k
j− 1
2

We are then interested in computing the deterministic numerical flux at the quadrature nodes zk , that is Fi + 1 (zk ). The
2
numerical flux at the quadrature node Fi + 1 (zk ) is given by the ADER numerical flux obtained by solving the generalised
 2
Riemann problem GRP Q L (x, zk ), Q R (x, zk ) , according to the deterministic algorithm summarized in Section 2.2. Notice that
by applying the mid-point quadrature rule would result in a second-order method. For a higher-order approximation in both
physical and stochastic spaces, we would need a higher-order polynomial reconstruction in both spaces, and a corresponding
higher-order quandrature rule to evaluate the numerical flux. One proceeds as follows, first, recover the information in the
y direction: on each cell (i , j) provide polynomials (for example, with ENO/WENO procedure)

y
Pi ,j (y) = Pi ,j (y, Qni,j , . . . , Qni,j ), j ∈ S j ∈ {j L , . . . , j R },
L R

y
where the dimension of the stencil S j depends on the desired order of accuracy in the stochastic direction. By evaluating
these polynomials in the stochastic direction Pki,j := Pi ,j (zk ), provide polynomials (for example, with ENO/WENO procedure)

 
Rki,j (x) = Rki,j x, PkiL ,j , . . . , PkiR ,j , i ∈ S ix = {i L , . . . , i R },

where the dimension of the stencil S ix depends on the desired order of accuracy in the physical direction. By solving the
 
GRP Rki,j (x), Rki+1,j (x) at each interface x = xi + 1 one finds Fk 1 := Fi + 1 (zk ). Once the flux is computed at each quadrature
2 i+ 2 2
node along the cell interface, the numerical flux Fi + 1 ,j is then computed according to (20). See Fig. 2.
2
412 M. Petrella et al. / Journal of Computational Physics 386 (2019) 405–427

Fig. 2. Schematic representation of the GRPk imposed at quadrature nodes, in a one-random-input problem.

2.5. The SFV-ADER numerical source

We approximate the integral in (16) via a quadrature rule with nodes zk and weights w k , i.e.

y
j+ 1
2
1 1 
Si ,j = Si (y)ρ (y) dy ≈ Si (zk )ρ (zk ) w k . (21)
yj yj
y k
j− 1
2

Again, we are interested in the quantity Si (zk ). The numerical source Si (zk ) is then a deterministic quantity which can
be computed by applying the standard ADER method with polynomial approximation Rki,j (x) at quadrature node zk , inside
cell (i , j). That is, one expands the sought solution inside the cell in an explicit time Taylor expansion and converts time
derivatives into space derivatives and the source term (also known as Cauchy-Kowalewskaya procedure) [66,68]: using (10)
(l) (l)
and recalling that ∂x Q(x, 0, zk ) = ∂x Rki,j (x) for all l ∈ N0 , one gets the time-expansion of the solution and approximates
space-time integrals according to suitable quadrature rules. Finally, the numerical source is computed according to (21).

2.6. Summary of the method

The SFV-ADER method can be summarized in the following steps for each time level t = t n :

1. For all i and j, provide a high-order polynomial reconstruction Pi ,j (y) in stochastic space;
2. At each quadrature node zk in stochastic space, do:
2a. Evaluate polynomials Pki,j = Pi ,j (zk ), for all i , j;
2b. Provide a high-order polynomial reconstruction Rki,j (x) in physical space by means of the previously defined values
Pki,j ;
2c. Apply the deterministic ADER solver with initial data Rki,j (xi + 1 ) and Rki+1,j (xi + 1 ) to compute the quantity Fk 1 ;
2 2 i + 2 ,j
2d. Apply the deterministic ADER solver to compute the quantity Ski with known relations

(l) (l)
∂x Q(x, t n , zk ) = ∂x Rki,j (x), ∀ l ∈ N0 .

3. Compute numerical flux and numerical source according to formulae (20) and (21), respectively, and evolve quantity
Qni,j according to formula (13).
4. Compute statistics of interest, by means of k-th moments according to (17).

The ENO reconstruction procedure for second-order schemes used here will be briefly explained in the following section.
 M. Petrella et al. / Journal of Computational Physics 386 (2019) 405–427 413

3. A second order scheme for a model equation

3.1. Algorithm

We discuss features of the SFV-ADER approach applied to a model equation, the viscous Burgers equation (vBE)
2
∂t q(x, t ) + ∂x f (q(x, t )) = α ∂xx q(x, t ), (22)
where f (q) = 12 q2 and uncertain viscosity coefficient α = α ± ε%, where the uncertain value α ∈ R is assumed to be affected
by an error ε ∈ [0, 100]. We convert (22) into
2
∂t q(x, t , ω) + ∂x f (q(x, t , ω)) = α (ω)∂xx q(x, t , ω), (23)
   
ε α and d := 1 + ε α , i.e. α (ω) ∼ Unif [c , d]. By
where α is a uniformly distributed random variable between c := 1 − 100 100
parametrizing (23) via the random variable y = α (ω), which takes values in the interval [c , d] ⊂ R with probability density
function

1
d−c
, if y ∈ [c , d]
ρ ( y) = ,
0, otherwise

we obtain
2
∂t q(x, t , y ) + ∂x f (q(x, t , y )) = y ∂xx q(x, t , y ). (24)
For simplicity in this viscous example, we follow the splitting procedure proposed by Toro and Brown [67]: the original PDE
(24) is replaced by two IVPs for the PDEs
2
∂t q(x, t , y ) + ∂x f (q(x, t , y )) = 0, ∂t q(x, t , y ) = y ∂xx q(x, t , y ). (25)
At each time level t = t , one first approximates the solution of the first equation in (25) and uses it as the initial value
n

for an implicit discretization of the diffusive term, i.e. second equation in (25). First we apply the SFV-ADER approach to
approximate the solution of the first PDE in (25). A finite volume formula to compute it is the scalar version of (13), where
an expression for the numerical flux has to be defined. Following Section 2.4, for each cell (i , j ) ∈ {1, . . . , M } × {1, . . . , N } the
numerical flux is given by the scalar version of (20). An ENO reconstruction in stochastic coordinate y yields the following
polynomials

y y j − 12 , | j − 12 | ≤| j + 12 |
qni, j +1 − qni, j
P i, j ( y) = qni, j + (y − y j ), = , j + 12 := , (26)
j j
j + 12 , | j − 12 | ≥| j + 12 |
y

where qni, j is the physical-stochastic integral average given by (14) of the sought solution q at time t = t n inside cell (i , j )
and y = y j +1 − y j is the stochastic mesh width. By evaluating this quantity at quadrature nodes zk , one gets the quantities
P ik, j = P i , j ( zk ), k ∈ {1, 2}. These are then used to provide polynomial reconstructions in physical space. Define polynomials
⎧
⎨ k
, | i − 12 | ≤| k
| P ik+1, j − P ik, j
i − 12 i + 12
R ki, j (x) = P ik, j + x
i ,k (x − xi ),
x
= , k
= . (27)
i ,k ⎩ k
, | k
| ≥| k
| i + 12 x
i + 12 i − 12 i + 12

We are then interested in computing f i + 1 ( zk ), for each quadrature node zk in stochastic space: apply the ADER approach
2
by solving the following GRP
⎧
  ⎪⎨∂t q + ∂x f (q) 
= 0,
k k
GRP Ri,j (x), Ri+1,j (x) = R ki, j (x) x < 0, (28)
⎪
⎩q(x, 0, yk ) =
R ki+1, j (x) x > 0.

The global numerical flux f i + 1 , j is then computed according to the scalar version of (20).
2
Now consider the second equation in (25): by integrating it into the finite volume V = [xi − 1 , xi + 1 ] × [t n , t n+1 ] ×
2 2
[ y j − 1 , y j + 1 ] and discretizing implicitly, one gets the approximate formula
2 2

y y
j+ 1 j+ 1
2 2
μj t
qni ,+j 1 = qni, j + qni ++11, j − 2qni ,+j 1 + qni −+11, j , μ j := y ρ ( y ) dy , y j := ρ ( y ) dy . (29)
y j x2
y y
j− 1 j− 1
2 2
414 M. Petrella et al. / Journal of Computational Physics 386 (2019) 405–427

The first-order splitting scheme to solve (24) reads

n+ 12 t
qi , j = qni, j − f i+ 1 , j − f i− 1 , j ,
x 2 2
(30)
n+ 12 μj t
qni ,+j 1 = qi , j + qni ++11, j − 2qni ,+j 1 + qni −+11, j ,
y j x2
which can be expressed as qni ,+j 1 = D t ( A t (qni, j )), where A stands for the advection operator and D stands for the diffusion
operator. As proven in [59] and in Chapter 15 of [66], this might result in a loss of accuracy even if both the operators A
and D are second-order accurate. The fully second-order version used here reads qni ,+j 1 = D t /2 ( A t ( D t /2 (qni, j ))).

3.2. Numerical results

Convergence of the SFV method for statistical quantities such as expectation and variance was proven for homogeneous
hyperbolic equations, see Section 4.5.2 of [63]. For clarity, we recall some conclusions: let us denote by q the exact solution
y xy
of (23), by qh the numerical solution which is exact in the x variable and discretized in the y and by qh the numerical
solution discretized in both variables. Assume that the numerical solution converges with rate p in the x variable and rate
s in the y variable, that is
y xy
qh − qh L1 (D ) ≤ C1 xp , ∀ y ∈ [c , d].
y s
(31)
q − qh L 1 ( Y ) ≤ C2 y , ∀x ∈ [a, b].
Then one can verify that the following holds
y xy y xy
E[qh ] − E[qh ] L1 (D ) ≤ C 1E x p , V[qh ] − V[qh ] L1 (D ) ≤ C 1V xp ,
y y
(32)
E[q] − E[qh ] L1 (D ) ≤ C 2E ys, V[q] − V[qh ] L1 (D ) ≤ C 2V ys.
In turn, (32) implies
xy  
E[q] − E[qh ] L1 (D ) ≤ C̃ E xp + ys ,
xy   (33)
V[q] − V[qh ] L1 (D ) ≤ C̃ V xp + ys .
Now returning to the ADER method of the present paper, this was tested with two initial conditions: smooth (i.e. Gaussian)
data and piecewise constant data. Results are shown in Figs. 3, 4, 5 and 6, for which exact solutions have been computed
using a fine mesh. Results show the ability of the SFV methodology combined with the ADER approach to accurately capture
the main features of statistical curves, such as expectation and variance. Notice the differences in approximation quality of
expectation and variance for both test cases; due to the higher oscillatory trend of variance profiles, expectations are more
accurately captured. To quantify the influence of mesh resolutions we conducted a mesh refinement study, reported
in Fig. 3. First we keep the stochastic mesh constant and refine the physical one. Top figures report results for a 2 cell
scheme in stochastic space, while a 10 cell scheme is reported at the bottom. Notice that when the stochastic space is not
sufficiently resolved, the scheme is not able to produce accurate solutions even by physical mesh refinement. Notice that a
few stochastic cells are sufficient to ensure convergence. To further investigate such empirical deduction, we conducted a
mesh convergence study (Fig. 4) in the stochastic direction, keeping constant the physical mesh. Fig. 4 shows no substantial
improvement of the numerical solution by stochastic mesh refinement, meaning that the stochastic space is resolved with
just a few stochastic cells compared to the number of physical ones. Notice that the same conclusions hold for the test
case with smooth initial data, as shown in Fig. 5. We note that in order to ensure convergence to the exact solution one
has to calibrate the stochastic and physical mesh sizes. Indeed, for more general problem, one needs to propagate each
random input with uncertainty and perform stochastic mesh refinement studies to find a suitable stochastic mesh size
for a prescribed physical mesh size. By “suitable” we mean that the contribution of the stochastic mesh size y in (33)
results comparable to the physical mesh size x one. Fig. 6 shows the norm of the difference between two successive
approximations computed with N = n and N = 2n stochastic cells. The decay of the error demonstrates that starting from a
certain threshold, further refinement of the stochastic space does not increase the overall accuracy of numerical solutions.
Indeed, for a sufficiently small physical mesh size x, the error in approximation statistical quantities is dominated by the
stochastic mesh width y, as stated in (33). Due to the typical small size of the stochastic space to subdivide into volumes,
the stochastic space results resolved with already a moderate number of stochastic cells.
Finally, we have also verified empirically the expected order of convergence, reported in Table 1, for the manufactured
solution q(x, t , ω) = α (ω) sin(π (x − st )), with characteristic velocity s = 1, Courant number C = 0.9, at time t out = 5 and
distribution α ∼ Unif[0.18, 0.22]. For the considered case, the exact statistical quantities of interest are the expectation E [q]
and the variance V [q], given by

(0.22 − 0.18)2
E [q] = 0.2 sin(π (x − st )), V [q] = sin2 (π (x − st )). (34)
12
 M. Petrella et al. / Journal of Computational Physics 386 (2019) 405–427 415

Fig. 3. Mesh refinement study in physical space. Initial condition: piecewise constant data. Column-wise: Expectation (Eq) and Variance (Vq) for different
physical mesh sizes (M) and stochastic mesh sizes (N) at time tout = 4 units.

Fig. 4. Mesh Refinement study in stochastic space. Initial condition: Riemann data. Column-wise: Expectation (Eq) and Variance (Vq) keeping constant the
physical mesh sizes (M) and for several stochastic mesh sizes (N) at time tout = 4 units.
416 M. Petrella et al. / Journal of Computational Physics 386 (2019) 405–427

Fig. 5. Mesh refinement study in both physical and stochastic direction for a test case with smooth initial profile (Gaussian). Column-wise: Expectation (Eq)
and Variance (Vq) at to ut = 4 units. Corresponding numbers of physical (M) and stochastic (N) cells are reported in labels.

Table 1 reports results for the SFV-ADER method computed using ENO reconstruction procedure in both physical and
stochastic space. Virtually for each norm, results suggest lower convergence for expectation as compared to variance. Such
behaviour has to be attributed to the ENO slope selection criterion: due to the oscillatory behaviour of statistical quantities
(34), the ADER method in combination with the ENO reconstruction results in a over-diffusive method, even for deter-
ministic problems. However, if approximate expectation exhibit typical features of numerical solutions under the effect of
numerical diffusion (e.g. clipping of extrema), variance bottoms results accurately captured for each mesh. Indeed, regions in
which variance vanishes consist of points where the pathwise solution has virtually no distortion from expectation. There-
fore, any acceptable approximation will result essentially fixed where variance has minima, explaining the difference in
approximation between expectation and variance bottoms. This in turn affects the order of convergence.
Nevertheless, statistical quantities present an empirical order of convergence lower than two in L 2 and L ∞ . As previ-
ously mentioned, this is a typical feature of ENO-reconstruction procedure combined with ADER schemes, which has been
shown to exhibit high diffusivity, [66]. Such observation motivated convergence studies with different slope selection cri-
teria, demonstrating the achievement of empirical second order. As an example, Table 2 shows results for a fixed central
choice of slopes in both stochastic and physical spaces.
The absence of fluxes in stochastic directions combined with the relatively low number of stochastic cells needed to
resolve the features of the solutions make the SFV-ADER method significantly cheaper than a full higher dimensional prob-
lem. This becomes even more important in the case of UQ for multiple random variables. The SFV-ADER scheme shows
promising results, and allows to consider multi-dimensional uncertainty at a non-excessive computational cost. The pro-
posed methodology results in a semi-intrusive, or weakly intrusive, method: the existing (deterministic) ADER solver is just
applied at quadrature nodes in stochastic space to calculate the corresponding numerical fluxes and numerical sources. The
 M. Petrella et al. / Journal of Computational Physics 386 (2019) 405–427 417

Fig. 6. Mesh convergence study for the two test cases of Fig. 5. Total number of physical cells M = 80.

approach results in a simple, flexible and generic deterministic method for UQ problems. This turns out to be crucial for the
computational cost of much more complex problems, as for biomedical simulations in which the physical setting is already
highly complex without any assumption of uncertainty.

4. UQ for a 1D blood flow model

In what follows we apply the SFV-ADER method to a non-linear hyperbolic system of balance laws of relevance to
biological fluid dynamics. The system of interest constitutes a one-dimensional mathematical model for arterial blood flow
in elastic vessels.

4.1. Mathematical model

The one-dimensional blood flow equations for a compliant vessel are the following

∂t A + ∂x q = 0,
  (35)
∂t q + ∂x β qA + ρA ∂x p = − R qA ,

where A (x, t ) is the cross-sectional area at space point x and at time t, q = Au is the flow of blood through the area A with
μ
axial velocity u, p is the pressure, ρ is the blood density (set to 1050 kg/m3 ), R = 22π ρ is the friction force with kinematic
viscosity μ = 2.5 mPa s and α is the Coriolis coefficient assumed here to be β = 1. The two governing partial differential
equations have three unknowns, namely the area A, the flow q and the pressure p and hence a closure condition, commonly
referred to as Tube Law, is required. Here we assume the elastic tube law
418 M. Petrella et al. / Journal of Computational Physics 386 (2019) 405–427

Table 1
Error estimates and empirical orders of accuracy for the discussed manufactured solution, using ENO procedure in both stochastic and physical space.

Manufactured solution
M N L1 L2 L∞ O (L1 ) O (L2 ) O (L∞ )
E V E V E V E V E V E V
20 2 9.24e-01 3.75e-01 8.62e-01 3.47e-01 1.22e+00 4.42e-01 0.00 0.00 0.00 0.00 0.00 0.00
40 4 2.64e-01 9.50e-02 2.69e-01 9.29e-02 5.16e-01 1.46e-01 1.81 1.98 1.68 1.90 1.24 1.59
80 8 7.13e-02 2.37e-02 8.52e-02 2.44e-02 2.12e-01 4.84e-02 1.89 2.00 1.66 1.93 1.29 1.60
160 16 1.89e-02 5.92e-03 2.67e-02 6.48e-03 8.57e-02 1.66e-02 1.92 2.00 1.67 1.92 1.30 1.55
320 32 4.99e-03 1.48e-03 8.53e-03 1.77e-03 3.50e-02 5.95e-03 1.92 2.00 1.65 1.87 1.29 1.48

Table 2
i− 1
+ i+ 1
Error estimates and empirical orders of accuracy for the discussed manufactured solution, using fixed central slope = 2
2
2
in both stochastic and
physical space.

Manufactured solution
M N L1 L2 L∞ O (L1 ) O (L2 ) O (L∞ )
E V E V E V E V E V E V
20 2 2.30e-01 3.48e-01 1.80e-01 3.01e-01 1.78e-01 3.41e-01 0.00 0.00 0.00 0.00 0.00 0.00
40 4 5.04e-02 8.68e-02 3.95e-02 7.52e-02 3.94e-02 8.66e-02 2.19 2.00 2.19 2.00 2.18 1.98
80 8 1.20e-02 2.16e-02 9.43e-03 1.87e-02 9.42e-03 2.16e-02 2.07 2.01 2.07 2.01 2.06 2.00
160 16 2.91e-03 5.39e-03 2.29e-03 4.67e-03 2.29e-03 5.40e-03 2.04 2.00 2.04 2.00 2.04 2.00
320 32 7.18e-04 1.35e-03 5.64e-04 1.17e-03 5.64e-04 1.35e-03 2.02 2.00 2.02 2.00 2.02 2.00

p (x, t ) = p ext (x, t ) + K (x)ψ( A (x, t ), A 0 (x)), (36)

where p ext (x, t ) is the external pressure, A 0 (x) is the cross-sectional area at equilibrium, ptrans = p − p ext is the transmural
pressure and
 m  n
A (x, t ) A (x, t )
ψ( A (x, t ), A 0 (x)) = − . (37)
A 0 (x) A 0 (x)
Allowing for arteries and veins in the tube law one has
⎧
⎨ E (x)h(x)
, m = 12 , n = 0 for arteries,
(1−ν 2 )r0 (x)
K (x) =  3 (38)
⎩ E (x) 2 h(x) , m ≈ 10, n = − 32 , for veins,
12(1−ν ) r 0 ( x)

where E (x) is the Young’s modulus, h(x) is the wall-thickness, ν is the Poisson coefficient and r0 (x) the radius of the vessel
at equilibrium.
We note that (36) for arteries models small deformations in a thin, homogeneous incompressible and purely elastic
arterial wall [6,4]; for this reason, the present formulation is commonly referred to as elastic formulation. For more details
on the mathematical structure of the proposed equations, see for example [28,70]. In the following we restrict the discussion
to arterial circulation, though the proposed UQ methodology can also be applied to veins.

4.2. Modelling arterial networks

The modelling of arterial blood flow has made significant advances in recent years, see [4,5,3,14,15,17,19,43,45,47,51,56,
58,65,69]. Arterial pulse wavelengths are sufficiently long to mathematically justify the use of a one-dimensional (1D) rather
than a three-dimensional (3D) approach when global assessment of blood flow in the cardiovascular system is required. The
assumption of 1D equations becomes less appropriate with decreasing calibers of vessels. Blood flow in large arteries is
pulsatile and dominated by inertia, whereas blood flow in smaller vessels is quasi steady and dominated by viscosity [6].
Consequently, any 1D model has to be truncated after a relatively small number of generations of bifurcations; haemody-
namic effects of vessels beyond 1D models for arteries is typically simulated using lumped parameter or zero-dimensional
(0D) models governed by systems of ordinary differential equations that relate pressure to the flow at the outflow of each
1D terminal vessel.
Matthys et al. [40] constructed a well-defined experimental 1:1 replica of the largest conduict arteries in the human
systemic circulation that has become very useful for assessing models and methods for blood flow in large systemic arteries.
Fig. 7 depicts a schematic representation of the experimental setup: the network of vessels consists of 37 silicone tubes
connected to a pulsatile pump delivering periodic output, analogous to the aortic flow. Terminal branches end in simple
resistive models consisting of stiff capillary tubes leading to an overflow reservoir reflecting constant venous pressure. All
the parameters required by the mathematical model were directly measured in the in vitro setup so no data fitting is
involved. Parameters are set according to Table 1 proposed in [40].
 M. Petrella et al. / Journal of Computational Physics 386 (2019) 405–427 419

Fig. 7. Sketch of the in vitro model proposed in [40,4] (by permission).

In this paper we propose a numerical simulation of such experiment, assuming constant parameters. The system of
governing equations (35)-(36) can easily be expressed in conservation-law form with an algebraic source term as follows

∂t Q(x, t ) + ∂x F(Q(x, t )) = S(Q(x, t )), (39)

with definitions
   
A (x, t ) q 0
Q(x, t ) = , F(Q) = 2 , S(Q) = . (40)
− R qA
q 3
q(x, t )
A
+γ A2

The model then computes pressure according to the elastic arterial tube law
 
A (x, t )
p ( A (x, t )) = p ext + K −1 , (41)
A0

with parameters defined as

1 K hE
γ= √ , K= , (42)
3ρ A0 (1 − ν 2 )r0
where the Poisson coefficient is assumed ν = 0.5 and the Young modulus is E = 1.2 MPa, as for silicone. Note that mod-
elling a multitude of vessels composing a network requires to provide at least three different types of boundary/interface
conditions, namely the heart flow at the inlet boundary of the network, the imposition of resistances at the outlet boundary
and a model for segments joining at junctions. From a numerical point of view, this is translated into determining values
for the sought solution at the boundary cell interface.

4.3. Boundary conditions and junctions

At the inlet of the ascending aorta, the periodic flow rate is imposed by means of a time-Fourier series of the flow
rate measured during the experiment. Thus, for each time level t = t n one provides the flow q∗ at the left interface of the
boundary, posing the problem of computing the corresponding area A ∗ . One way to compute it is to exploit generalized
Riemann invariants of the governing equations. These are relations that apply across the wave structure and lead to Ordinary
Differential Equations (ODE) in phase space. These relations for the blood flow equations (35)-(41) leads to constancy of the
quantity [65]

q 3 √
− (Q) = − 4c ( A ), c( A) = γ A, (43)
A 2
420 M. Petrella et al. / Journal of Computational Physics 386 (2019) 405–427

Fig. 8. Schematic representation of imposed boundary conditions at the inlet of the network.

where c is the wave speed, across the wave structure associated with the eigenvalue λ2 = u + c. Imposing constancy of −
between the star region and the right domain one finds

q1 q∗
− (Q∗ ) = − (Q1 ) ⇔ F ( A∗) = − 4c ( A 1 ) − + 4c ( A ∗ ) = 0. (44)
A1 A∗
Fig. 8 shows a schematic representation of the procedure. Due to the non-linearity of the resulting function F , we solve
equation (44) numerically for which many numerical methods are available. Here we use a Newton-Raphson method for
which a good initial guess is A (0) = A 1 . See Fig. 8.
The rigid tube attached to the outlet of each terminal branch and the overflow reservoir at constant height are simulated
as outflow boundary conditions [6] by enforcing the relation

p ( A ∗ ) − p out
q∗ = . (45)
Rp
In particular, indirect measurements for peripheral resistances are provided in Table 1 in [40], so the values of resistance
R p are known. It is worth recalling that peripheral resistances have been computed via relation (45) by measuring mean
pressure and flow close to the outlet of each terminal branch. Similarly, the hydrostatic pressure measured at the overflow
reservoir representing the venous pressure is p out = 3.2 mmHg. Hence, A ∗ at the right boundary remains unknown. Again,
by exploiting generalized Riemann invariants one imposes constancy of + (Q) = A + 4c ( A ) across the wave associate with
q

the eigenvalue λ1 = u − c, and connects the right boundary of ending cells in the computational domain by imposing the
relation

+ (Q M ) = + (Q∗ ), (46)

where + (Q M ) is the evaluation of + at ending cells. Notice that in (46) both A ∗ and q∗ are unknown: by inserting (45)
into (46) and after some manipulations one is led to finding the roots of the following equation

F ( A ∗ ) =  M A ∗ R p − 4c ( A ∗ ) A ∗ R p − p ( A ∗ ) + p out = 0. (47)

Applying a Newton-Raphson method, for example, with initial guess A (0) = A M (i.e. the first component of the datum at
ending cells of the computational vessels), one finds A ∗ and, thus, q∗ , according to (45).
Finally, boundary conditions of arterial segments joining at junctions are determined by enforcing constancy of Riemann
invariants connecting the boundary of each vessel involved in the junction to the interior of the vessel; conservation of mass
and continuity of the total pressure p + 12 ρ u 2 . This procedure leads to a non-linear system of algebraic equations which can
be solved by means of a suitable Newton method [57,3]; see also [46] and [23].
It is worth outlining that in [46] and [23] it is demonstrated that the described junction methodology to treat junctions,
as originally proposed, results in a loss of accuracy, even if high-order schemes are used in the interior of vessels. Such
problem is remedied by matching the accuracy at junctions to that in the interior of the domain. One way to do so is
by reinterpreting the initial-value problems at junctions as particular Generalized Riemann Problems augmented by cou-
pling conditions. For second-order accuracy one proceeds as follows: first, a high-order polynomial reconstruction around
the boundary (typically accomplished by means of one-sided reconstruction procedures) is implemented, one then evalu-
ates such polynomials at the appropriate boundaries (extrapolated values), and applies the methodology presented so far
to extrapolated values, instead of cell averages. This preserves the desired order of accuracy in the entire computational
domain.
 M. Petrella et al. / Journal of Computational Physics 386 (2019) 405–427 421

Notice that the same procedure can be applied to inlet and outlet boundary conditions, namely the imposition of heart
flow and peripheral resistances. Indeed, this increases the convergence rate of the approximate solution providing a uniform
order in accuracy, and it is the procedure we have adopted for our simulations reported here. A more detailed discussion
on GRPs and numerical methodologies to solve them can be found in [66].
The proposed 1D mathematical model has been tested against in vitro experiments [12,33] and in vivo measurements [50,
52,53,44,16] demonstrating its ability to capture the main features of pressure and flow wave propagation in large arteries
and hence our adoption of this mathematical model to simulate clinically relevant problems. In [40] the authors noted that,
as a consequence of the simple resistive boundary conditions used to reduce the uncertainty of the parameters involved in
the simulation, the experimental set-up generates waveforms at terminal branches with additional non-physiological oscil-
lations. The frequencies of these oscillations are well captured by the 1D model, even though amplitudes are overestimated.
Indeed, in both pressure and flow rates the elastic formulation produces abnormally large amplitude fluctuations as
compared to experiments, that is, the hereby considered elastic model yields the paradoxical situation in which, given a
mesh size, it produces better results with low-order schemes, due to their inherent numerical viscosity, than higher-order
methods. By better we mean as compared to measurements. On the other hand, considering errors, the 1D model (39)-(41)
produces acceptable approximations to measurements, highlighting the presence of model shortcomings. Indeed, Fig. 3
and Fig. 4 in [40] show that in several locations studied, numerical results do not fit into the range of uncertainty of
measurements, posing the uncertain situation in which either measurements are inaccurate to be considered reliable, or
numerical results are unreliable due to model shortcomings or deficient numerical approximations. This discussion motivates
the contents of the next section of this paper.

4.4. A 1D model with uncertainty

Considering the heterogenous nature of arterial vessels, constancy of mechanical and geometric parameters character-
izing the vessel throughout its length may be a strong assumption. In light of this, it becomes also unrealistic to aim for
knowledge of the exact values of such parameters along each vessel that constitutes the arterial tree; this is specially so
in clinically relevant situations. Assuming inexactness of parameters, one has then to quantify the associated uncertainty of
computational results. As a first step in uncertainty quantification for the 0D-1D model (39)-(41), in this work we assume
the elastic formulation with uncertainty in several parameters. That is, we assume uncertainty in Young’s modulus E = E (ω),
wall-thickness h = h(ω) and equilibrium radius r0 = r0 (ω). Besides these parameters, friction is also another important un-
certain parameter which could play a role on the flow characteristics. In this work we aim to perform a sensitivity analysis
with respect to wall parameters, therefore we will assume constant friction force R, as well as no uncertainty on heart-flow
and on terminal boundary conditions. We rewrite (39) as

∂t Q(x, t , ω) + ∂x F(Q(x, t , ω), ω) = S(Q(x, t , ω)), ω ∈ , (48)

T
q(x,t ,ω)
where  is the set of events, Q(x, t , ω) = [ A (x, t , ω), q(x, t , ω)] T is the conserved variable vector, S(Q) = 0, − R A (x,t ,ω)
is the source term and the (stochastic) flux vector F depends explicitly on the random vector ξ (ω) = [ E (ω), h(ω), r0 (ω)] T
according to
 
q(x, t , ω)
F(Q(x, t , ω), ω) = q2 (x,t ,ω) 3 ,
A (x,t ,ω)
+ γ (ω) A 2 (x, t , ω)
(49)
1 K (ω) h(ω) E (ω)
γ (ω) = √ , K (ω) = ,
3ρ A 0 (ω) (1 − ν 2 )r0 (ω)
where A 0 (ω) := A 0 (r0 (ω)) = π r0 (ω)2 . We parametrize (48) by means of the random vector [ y 1 , y 2 , y 3 ] T = y = ξ (ω), i.e.

∂t Q(x, t , y) + ∂x F(Q(x, t , y), y) = S(Q(x, t , y)), (50)

with definitions
   
A (x, t , y) q 0
Q(x, t , y) = , F(Q, y) = q2 , S(Q) = , (51)
− R qA
3
q(x, t , y)
A
+ γ (y) A 2

and parameters
K (y) y1 y2
γ (y) = √ , K (y) = , A 0 (y) = π y 3 2 . (52)
3ρ A 0 (y) (1 − ν 2 ) y 3
The system of stochastic balance laws (50) computes pressures according to the stochastic-elastic tube law
 
A (x, t , y)
p (x, t , y) = K (y) −1 . (53)
A 0 (y)
422 M. Petrella et al. / Journal of Computational Physics 386 (2019) 405–427

In the next section we show computational results from our proposed UQ methodology as applied to blood flow in
arteries assuming uncertainty in three parameters. For the purpose of our study, the physical model and associated mea-
surements of Matthys and collaborators [40] have been indeed very useful.

4.5. Results

The system of stochastic balance laws (50) was solved with the SFV-ADER approach described in subsection 2.3, as-
suming first-degree polynomial reconstruction in both stochastic and physical space. The scheme was implemented to test
uncertainty in the Young’s modulus (assumed to be affected by an error of about ±5%) and the uncertainties provided in
Table 1 of [40]: we assume an error of the order of 2.5% and 3.5% for wall-thickness h and equilibrium radius r0 , respec-
tively, and uniform distribution for every random input. The joint probability density function ρi (y) of the random vector
ξ i = ( E i (ω), hi (ω), r0i (ω)) in vessel i ∈ {1, . . . , 37} is given (by independence) by
 1
 
, yk ∈ cki , dki
dki −cki
ρi (y) = k3=1 ki ( yk ),
ρ i
ρ
k ( yk ) = (54)
0 otherwise,
 
where the interval [cki , dki ] = (1 − εk )λki , (1 + εk )λki , εk is the error assumed on parameter k, and λki is equal to 1.2 MPa
for the Young’s modulus E, whereas is the value provided in Table 1 of [40] for the parameters h i and r0i . The stochastic
space was discretized with an isotropic mesh of N 1 = 20, N 2 = 8 and N 3 = 12 stochastic cells for variables y 1 , y 2 and
y 3 , respectively. This particular choice was motivated by a mesh convergence study in stochastic space: by Section 3.2 one
needs to consider a sufficiently large number of stochastic cell for each random input to ensure convergence. This was
accomplished by propagating the model (39)-(42) with uncertainty in each parameter and keeping constant all the other
uncertain parameters. For example, to find a sufficiently large number of stochastic cells for the Young’s modulus E, we
considered the model (50)-(53) taking h i (ω) = h i and r0i (ω) = r0 for each ω ∈ , where h i and r0i are the wall-thickness
and the equilibrium radius provided in Table 1 on [40]. Afterwards, we performed a mesh convergence study in stochastic
space, finding N 1 = 20 as the threshold after which no substantial improvements in refining the stochastic mesh have been
found. The same procedure was then applied for the random inputs h, keeping constant parameters E and r0 , and for r0 ,
keeping constant parameters E and h.
The physical discretization was imposed assuming a cell width of the order of 2 cm in each vessel, requiring a minimum
of 2 cells per vessel (due to the ENO reconstruction procedure used). The solver used for solving local GRPs was the
Harten-Engquist-Osher-Chakravarthy (HEOC) solver [18], [30]. For each realization y = ( y 1 , y 2 , y 3 ) of each random input
y, a network consisting of 37 vessels with parameters y = y is allocated, for a total of 20 × 8 × 12 = 1920 networks and
1920 × 37 = 71040 vessels. The global time-step is computed as the minimum among all time steps associated to each
network.
Before discussing results a remark concerning boundary conditions is needed. In this work we assume deterministic
boundary conditions at the inlet and outlet of the network, whereas we use uncertain junction conditions. By uncertain
junction conditions we mean the following: as pointed out in section 4.3, the imposition of junction conditions in the de-
terministic model can be enforced by imposing suitable numerical fluxes at the boundaries of each computational vessel.
Thus, in case of uncertain parameters, one imposes the numerical flux by following the procedure exposed in subsection
2.4: reconstruct the information in stochastic space, apply the deterministic procedure to compute junction condition at
each quadrature node in stochastic space, and finally do computations to recover the overall numerical flux. Notice that the
same procedure can be easily extended to heart-flow or terminal boundary conditions.
Fig. 9 shows expectations and variances for pressure and flow at the midpoint of the left renal artery from the 8-th to
the 16-th cardiac cycle, noting that we look for periodic solutions. Each cardiac cycle has a period T = 0.827 s. Appreciable
differences in cycle plots can be seen in variance peaks and pressure variance ending segments, while expectations have
already assumed a periodic behaviour at the 8-th cardiac cycle. For this reason, all the results of the following figures have
been obtained running the code up to the 14-th cardiac cycle in order to observe the sought periodic solutions.
Fig. 10 shows results for a 99% confidence interval (blue area) for pressure and flow rates at the mid point of some stud-
ied locations for one cardiac cycle. Considerable difference in variability of pressure and flow rates emerges when changes in
parameter moduli are considered, highlighting a strong sensitivity of pressure to geometric and mechanical characteristics.
On the one hand, moderately small changes in parameter values can cause large changes in the corresponding pressures.
On the other hand, flow rates show an essentially insensitive trend to parameter fluctuations. By the confidence level con-
sidered and in virtue of the Chebyschev Inequality (3), it holds that if we slightly perturbe the values for the parameters E,
h and r0 the results of the model produced by such choice of parameters will lay inside the blue area depicted in Fig. 10
with probability P ∈ [0.99, 1], that is, almost surely. We may then conclude that the elastic formulation of the 1D model
(39)-(41) associates similar flow rate predictions to different values of parameters E, h and r0 .
Given the above observation, a comparison between standard deviation for pressure and flow rates throughout the whole
network was conducted in order to identify more/less sensitive areas to parameter fluctuations. Recall that by definition,
variance is the mean squared deviation of a random variable from its mean. Thus, heuristically, its square root (i.e. the
standard deviation) is a measure that quantifies the amount of dispersion of the random variable from its mean. A low
 M. Petrella et al. / Journal of Computational Physics 386 (2019) 405–427 423

Fig. 9. Expectation (left) and variance (right) of pressure (top) and flow (bottom) at the midpoint of the left renal artery from the 8-th cardiac cycle (Cyl:
8) to the 16-th cardiac cycle (Cyl: 16). Note different scales for variance.

standard deviation indicates that data points tend to be close to the mean, while a high standard deviation indicates that
the outcomes are spread out over a wider range of values.
Fig. 11 shows the absolute values of standard deviation at the midpoint of each vessel in the network. Terminal vessels
ending in resistive models have been highlighted by adopting a different colour (purple). Note that bigger moduli of standard
deviations in flow results are associated to larger and less resistive arteries, i.e. the ones composing the aortic path. Similarly,
the most robustly predicted values (i.e. the ones connected to lowest standard deviation moduli) are thinner and more
resistive ones. Indeed, flow in peripheral vessels is less pulsatile and more dominated by inertia than in more central ones,
making flow waveforms less sensitive to parameters fluctuations, and explaining the decreasing variability as we move
distally. Furthermore, in such areas, terminal resistances generate backward travelling waves, which propagate in the whole
computational domain. The reflections at terminal branches cause a kind of resonant effect, where backward travelling
waves are added to forward travelling ones, due to the less pulsativity nature of such compartments. Thus, in particular,
flow waveforms result shaped by terminal resistances, demonstrating the impact of simple models for resistive terminal
boundary conditions on the whole network results.

5. Concluding remarks

We have introduced a fully discrete, general framework for a blend of SFV and ADER methodologies for systems of
non-linear hyperbolic balance laws with uncertainty. The proposed scheme is simple, flexible, applicable to the general IBVP
and has no theoretical accuracy barrier in space, time and stochastic space. The resulting practical scheme is semi-intrusive
[1,63]: the deterministic solver is applied at quadrature points in stochastic dimension, followed by a quadrature rule.
The procedure has been applied in its second order mode, first to the viscous Burgers equation with uncertain viscosity
424 M. Petrella et al. / Journal of Computational Physics 386 (2019) 405–427

Fig. 10. Results for 99% confidence interval for pressure (left) and flow (right) waveforms at the midpoint of the first segment of the thoracic aorta artery,
left renal artery, ending segment of the right-iliac femoral artery and right carotid artery proposed in [4,40]. (For interpretation of the colours in the
figure(s), the reader is referred to the web version of this article.)
 M. Petrella et al. / Journal of Computational Physics 386 (2019) 405–427 425

Fig. 11. Standard deviation comparison. Absolute values of standard deviation for pressure (top) and flow (bottom) rates at the midpoint of each vessel in
the computational network within a cardiac cycle. Purple columns identify vessels with resistance as right boundary condition.

coefficient; for this example we have demonstrated the attainment of the theoretically expected order of accuracy. A few
examples with initial data without uncertainty were also considered in order to show advantages and limitations of the
approach.
The SFV-ADER method was then applied to a non-linear one-dimensional hyperbolic system of balance laws for blood
flow in elastic arteries with uncertainty in Young’s modulus, wall-thickness and equilibrium radius. Heart-flow and terminal
boundary conditions have been kept deterministic, in order to assess the impact of mechanical properties uncertainty on
the overall results, and uniform distribution of random inputs has been assumed. Pressure waves predicted by the elastic
formulation have been shown to be highly sensitive to uncertain parameter fluctuations, while no considerable sensitivity
has been found for flow waves.
By means of a standard deviation study throughout the network, we have demonstrated the strong impact of the sim-
ple resistive boundary conditions applied here, on the entire network results. Such boundary conditions are typical in
biomedical simulations, in which a particular fluid compartment is separated artificially from the remaining parts. Indeed,
such approach simplifies the modelling procedure and reduces computational cost. However, as demonstrated in this study
through uncertainty quantification, such boundary procedures are prone to errors. This confirms the importance of consid-
ering the arterial network as a district of a larger network of fluid compartments, as proposed, for example, in the work of
Müller and Toro [44] for the entire human circulation.
The range of possible developments of the proposed SFV-ADER methodology is extensive. These include numerical as-
pects, optimization of codes and application to other physical/biological problems. Regarding accuracy, for example, the
methodology could be extended to include schemes of higher order of accuracy (greater than two); this will result in
significant improvements on the efficiency of the methodology for computations aiming at small errors. For multidimen-
sional conservation laws the scheme can also be implemented on unstructured meshes. The current SFV-ADER methodology
has been formulated in the finite volume framework, but this can potentially be done also in the Discontinuous Galerkin
framework [24].
Obviously, the SFV-ADER computational cost increases dramatically as the number of random inputs increases. This poses
the problem of looking for efficiency gains, apart from that resulting from purely higher order of accuracy of the numerical
scheme. One way would be to use an adaptive parametrization in stochastic space, as proposed by Tokareva et al. [63]
or the ANOVA decomposition [76,81]. These are developments that are relevant to ambitious engineering and scientific
applications.
426 M. Petrella et al. / Journal of Computational Physics 386 (2019) 405–427

Acknowledgements

The Los Alamos unlimited release number for this paper is LA-UR-18-27216.

References

[1] R. Abgrall, A Simple, Flexible and Generic Deterministic Approach to Uncertainty Quantifications in Non Linear Problems: Application to Fluid Flow
Problems, Rapport de Recherche INRIA, 2007.
[2] R. Abgrall, P. Congedo, A semi-intrusive deterministic approach to uncertainty quantification in non-linear fluid flow problems, J. Comput. Phys. 235
(2013) 828–845.
[3] J. Alastruey, Numerical Modelling of Pulse Wave Propagation in the Cardiovascular System: Development, Validation and Clinical Applications, PhD
thesis, Imperial College London, University of London, 2006.
[4] J. Alastruey, A.W. Khir, K.S. Matthys, P. Segers, S.J. Sherwin, Pulse wave propagation in a model human arterial network: assessment of 1-D visco-elastic
simulations against in vitro measurements, J. Biomech. 44 (2011) 2250–2258.
[5] J. Alastruey, S.M. More, K.H. Parker, T. David, J. Peirò, S.J. Sherwin, Reduced modelling of blood flow in the cerebral circulation: coupling 1D, 0D and
cerebral auto-regulation models, Int. J. Numer. Methods Fluids 56 (2008) 1061–1067.
[6] J. Alastruey, K.H. Parker, J. Peirò, S.J. Sherwin, Lumped parameter outflow models for 1-D blood flow simulations: effect on pulse waves and parameter
estimation, Commun. Comput. Phys. 4 (2008) 317–336.
[7] I. Babuska, F. Nobile, R. Tempone, A stochastic collocation method for elliptic partial differential equations with random input data, SIAM J. Numer.
Anal. (2007) 1005–1034.
[8] D. Balsara, T. Rumpf, M. Dumbser, C.D. Munz, Efficient, high accuracy ADER-WENO schemes for hydrodynamics and divergence-free magnetohydrody-
namics, J. Comput. Phys. 228 (7) (2009) 2480–2516.
[9] T.J. Barth, On the propagation of statistical model parameter uncertainty in CFD calculations, Theor. Comput. Dyn. 26 (2012) 435–457.
[10] T.J. Barth, Non-intrusive uncertainty propagation with error bounds for conservation laws containing discontinuities, in: Uncertainty Quantification, in:
Lecture Notes in Computational Science and Engineering, vol. 92, Springer-Verlag, Heidelberg, 2013, pp. 1–55.
[11] M. Berveiller, B. Sudret, M. Lemaire, Stochastic finite element: a non-intrusive approach by regression, Rev. Eur. Mec. Numer. 15 (2006) 81–92.
[12] D. Bessems, C.G. Giannopapa, M.C. Rutten, F.N. van de Vosse, Experimental validation of a time-domain-based wave propagation model of blood flow
in viscoelastic vessels, J. Biomech. 41 (2) (2008) 284–291.
[13] H. Bijl, D. Lucor, S. Mishra, C. Schwab, Uncertainty Quantification in Computational Fluid Dynamics, Lecture Notes in Computational Science and
Engineering, vol. 92, Springer Verlag, 2013.
[14] P.J. Blanco, R.A.B. de Queiroz, R.A. Feijóo, A computational approach to generate concurrent arterial networks in vascular territories, Int. J. Numer.
Methods Biomed. Eng. 29 (2013) 601–614.
[15] P.J. Blanco, R.A. Feijóo, S.A. Urquiza, A unified variational approach for coupling 3D-1D models and its blood flow applications, Comput. Methods Appl.
Mech. Eng. 196 (2007) 4391–4410.
[16] P.J. Blanco, S.M. Watanabe, E.A. Dari, M.R.F. Passos, R.A. Feijóo, Blood flow distribution in an anatomically detailed arterial network model: criteria and
algorithms, Biomech. Model. Mechanobiol. 13 (6) (2014) 1303–1330.
[17] A. Caiazzo, F. Caforio, G. Montecinos, L.O. Muller, P.J. Blanco, E.F. Toro, Assessment of reduced-order unscented Kalman filter for parameter identification
in 1-dimensional blood flow models using experimental data, Int. J. Numer. Methods Biomed. Eng. (2016) e2843, https://doi.org/10.1002/cnm.2843.
[18] C.E. Castro, E.F. Toro, Solvers for the high-order Riemann problem for hyperbolic balance laws, J. Comput. Phys. 227 (2008) 2481–2513.
[19] V. Casulli, M. Dumbser, E.F. Toro, Semi-implicit numerical modelling of axially symmetric flows in compliant arterial systems, Int. J. Numer. Methods
Biomed. Eng. 28 (2) (2012) 257–272.
[20] M. Cheng, T.Y. Hou, Z. Zhang, A dynamically bi-orthogonal method for time-dependent stochastic partial differential equations I: derivation and algo-
rithms, J. Comput. Phys. (2013), https://doi.org/10.1016/j.jcp.2013.02.033.
[21] M. Cheng, T.Y. Hou, Z. Zhang, A dynamically bi-orthogonal method for time-dependent stochastic partial differential equations II: adaptivity and
generalizations, J. Comput. Phys. (2013).
[22] A. Chertock, S. Jin, A. Kurganov, A well-balanced operator splitting based stochastic Galerkin method for the one-dimensional Saint-Venant system with
uncertainty, Conference Proceedings Preprint, 2015.
[23] C. Contarino, E.F. Toro, G.I. Montecinos, R. Borsche, J. Kall, Junction-generalized Riemann problem for stiff hyperbolic balance laws in networks: an
implicit solver and ADER schemes, J. Comput. Phys. 315 (2016) 409–433.
[24] M. Dumbser, D. Balsara, E.F. Toro, C.D. Munz, A unified framework for the construction of one-step finite-volume and discontinuous Galerkin schemes
on unstructured meshes, J. Comput. Phys. 227 (2008) 8209–8253.
[25] M. Dumbser, C.D. Munz, ADER discontinuous Galerkin schemes for aeroacoustics, C. R., Méc. 333 (9) (2005) 683–687.
[26] O.G. Ernst, A. Mugler, H.J. Starkloff, E. Ulmann, On the convergence of generalized polynomial chaos expansions, ESAIM: M2AN 46 (2012) 317–339.
[27] G. Fishman, Monte Carlo, Springer, 1996.
[28] I. Formaggia, A. Quarteroni, A. Veneziani, Cardiovascular Mathematics. Modelling and Simulation of the Circulatory System, Springer, 2009.
[29] R. Ghamen, P. Spanos, Stochastic Finite Elements, Dover Pub. Inc., Mineola, New York, 1991.
[30] A. Harten, B. Engquist, S. Osher, S.R. Chakravarthy, Uniformly high order accuracy essentially non-oscillatory schemes III, J. Comput. Phys. 71 (1987)
231–303.
[31] S. Hosder, R.W. Walters, R. Perez, A non-intrusive polynomial chaos method for uncertainty propagation in CFD simulations, in: 44th AIAA Aerospace
Sciences Meeting and Exhibit, 2006.
[32] S. Hosder, R.W. Walters, M. Balch, Efficient uncertainty quantification applied to the aeroelastic analysis of a transonic wing, in: 46th AIAA Aerospace
Sciences Meeting and Exhibit, 2008.
[33] W. Huberts, K.V. Canneyt, P. Segers, S. Eloot, J. Tordoir, P. Verdonck, F. van de Vosse, E. Bosboom, Experimental validation of a pulse wave propagation
model for predicting hemodynamics after vascular access surgery, J. Biomech. 45 (9) (2012) 1684–1691.
[34] S. Jin, D. Xiu, X. Zhu, A well-balanced stochastic Galerkin method for scalar hyperbolic balance laws with random inputs, J. Sci. Comput. 67 (2015).
[35] M. Käser, A. Iske, ADER schemes on adaptive triangular meshes for scalar conservation laws, J. Comput. Phys. 205 (2005) 486–508.
[36] G. Loeven, H. Bijl, Probabilistic collocation used in a two-step approach for efficient uncertainty quantification in computational fluid dynamics, Comput.
Model. Eng. Sci. 36 (3) (2008) 193–212.
[37] O. Le Maitre, O. Knio, H. Najm, R. Ghanem, A stochastic projection method for fluid flow I. Basic formulation, J. Comput. Phys. 173 (2001) 481–511.
[38] O. Le Maitre, M. Reagan, H. Najm, R. Ghanem, O. Knio, A stochastic projection method for fluid flow II. Random process, J. Comput. Phys. 181 (2002)
9–44.
[39] L. Mathelin, M.Y. Hussaini, T.A. Zang, Stochastic approaches to uncertainty quantification in CFD simulations, Numer. Algorithms 38 (2005) 209–236.
[40] K.S. Matthys, J. Alastruey, J. Peirò, A.W. Khir, P. Sergers, P.R. Verdonck, K.H. Parker, S.J. Sherwin, Pulse wave propagation in a model human arterial
network: assessment of 1-D numerical simulation against in vitro measurements, J. Biomech. 40 (2007) 3476–3486.
 M. Petrella et al. / Journal of Computational Physics 386 (2019) 405–427 427

[41] S. Mishra, C. Schwab, Sparse tensor multi-level Monte Carlo finite volume methods for hyperbolic conservation laws with random initial data, Math.
Comput. 81 (2012) 1979–2018.
[42] S. Mishra, C. Schwab, J. Sukys, Multilevel Monte Carlo finite volume methods for shallow water equations with uncertain topography in multi-
dimensions, SIAM J. Sci. Comput. 34 (6) (2012) B761–B784.
[43] G.I. Montecinos, L.O. Müller, E.F. Toro, Hyperbolic reformulation of a 1D viscoelastic blood flow model and ADER finite volume schemes, J. Comput.
Phys. 266 (2014) 101–123.
[44] L.O. Müller, E.F. Toro, A global multiscale mathematical model for the human circulation with emphasis on the venous system, Int. J. Numer. Methods
Biomed. Eng. 30 (2014) 681–725.
[45] L.O. Müller, E.F. Toro, Well-balanced high-order solver for blood flow in networks of vessels with variable properties, Int. J. Numer. Methods Biomed.
Eng. 29 (2013) 1388–1411.
[46] L.O. Müller, P.J. Blanco, A high order approximation of hyperbolic conservation laws in networks: application to one-dimensional blood flow, J. Comput.
Phys. 300 (2015) 423–437.
[47] J.P. Mynard, M.R. Davidson, D.J. Penny, J.J. Smolich, A simple, versatile valve model for use in lumped parameter and one-dimensional cardiovascular
models, Int. J. Numer. Methods Biomed. Eng. 28 (2012) 626–641.
[48] F. Nobile, R. Tempone, C. Webster, A sparse grid stochastic collocation method for partial differential equations with random input data, SIAM J. Numer.
Anal. 46 (2009) 2309–2345.
[49] W.L. Oberkampf, T.G. Truncano, Verification and validation in computational fluid dynamics, Prog. Aerosp. Sci. 38 (2002) 209–272.
[50] M.S. Olufsen, C.S. Peskin, W.Y. Kim, E.M. Pedersen, A. Nadim, J. Larsen, Numerical simulation and experimental validation of blood flow in arteries with
structured-tree outflow conditions, Ann. Biomed. Eng. 28 (2000) 1281–1299.
[51] A. Quarteroni, M. Tuveri, A. Veneziani, Computational vascular fluid dynamics: problems, models and methods, Comput. Vis. Sci. 2 (2000) 163–197.
[52] P. Reymond, F. Merenda, F. Perren, D. Rüfenacht, N. Stergiopulos, Validation of a one-dimensional model of the systemic arterial tree, Am. J. Physiol.,
Heart Circ. Physiol. 297 (2009) H208.
[53] P. Reymond, Y. Bohraus, F. Perren, F. Lazeyras, N. Stergiopulos, Validation of a patient-specific one-dimensional model of the systemic arterial tree, Am.
J. Physiol., Heart Circ. Physiol. 301 (3) (2011) H1173.
[54] C.J. Roy, W.L. Oberkampf, A comprehensive framework for verification, validation, and uncertainty quantification in scientific computing, Comput.
Methods Appl. Mech. Eng. 200 (2011) 2131–2144.
[55] C. Schwab, S. Tokareva, High order approximation of probabilistic shock profiles in hyperbolic conservation laws with uncertain initial data, ESAIM:
Math. Model. Numer. Anal. 47 (2013) 807–835.
[56] S.J. Sherwin, L. Formaggia, J. Peirò, V. Franke, Computational modelling of 1D blood flow with variable mechanical properties and its application to the
simulation of wave propagation in the human arterial system, Int. J. Numer. Methods Fluids 43 (2003) 673–700.
[57] S.J. Sherwin, V. Franke, J. Peirò, K.H. Parker, One-dimensional modelling of a vascular network in space-time variables, J. Eng. Math. 47 (2003) 217–250.
[58] N. Stergiopulos, D.F. Young, T.R. Rogge, Computer simulation of arterial flow with applications to arterial and aortic stenoses, J. Biomech. 25 (1992)
1477–1488.
[59] G. Strang, On the construction and comparison of difference schemes, SIAM J. Numer. Anal. 5 (3) (1968) 506–517.
[60] J. Sukys, Robust Multi-Level Monte Carlo Finite Volume Methods for Systems of Hyperbolic Conservation Laws with Random Input Data, PhD Th., Diss.
ETH No. 21990, 2014.
[61] A.C. Taylor, L.L. Green, P.A. Newman, M.M. Putko, Some advanced concepts in discrete aerodynamic sensitivity analysis, AIAA J. (2001) 2001–2529.
[62] V.A. Titarev, E.F. Toro, ADER: arbitrary high order Godunov approach, J. Sci. Comput. 17 (2002) 609–618.
[63] S. Tokareva, Stochastic Finite Volume Methods for Computational Uncertainty Quantification in Hyperbolic Conservation Laws, PhD Dissertation, ETH
Zürich, Nr. 21498, 2013.
[64] S. Tokareva, C. Schwab, S. Mishra, High-Order SFV and Mixed SDG/FV Methods for the Uncertainty Quantification in Multidimensional Conservation
Laws. High Order Nonlinear Numerical Schemes for Evolutionary PDEs, Lecture Notes in Computational Science and Engineering, vol. 99, 2014.
[65] E.F. Toro, Brain venous haemodynamics, neurological diseases and mathematical modelling. A review, Appl. Math. Comput. 272 (2016) 542–579.
[66] E.F. Toro, Riemann Solvers and Numerical Methods for Fluid Dynamics, Springer-Verlag Berlin Heidelberg, 2009.
[67] E.F. Toro, R.E. Brown, The WAF method and splitting procedures for viscous shocked flows, in: Proceedings of the 18th International Symposium,
Sendai, Japan, July 21–26, 1991, vol. 2, 1992, pp. 1119–1126.
[68] E.F. Toro, R.C. Millington, L.A. Nejad, Towards very high-order Godunov schemes, in: E.F. Toro (Ed.), Godunov Methods: Theory and Applications. Edited
Review, Kluwer Academic/Plenum Publishers, 2001, pp. 905–937.
[69] E.F. Toro, G. Montecinos, Advection-diffusion-reaction equations: hyperbolization and high-order ADER discretizations, SIAM J. Sci. Comput. 36 (5)
(2014) A2423–A2457.
[70] E.F. Toro, A. Siviglia, Flow in collapsible tubes with discontinuous mechanical properties: mathematical model and exact solutions, Commun. Comput.
Phys. 13 (2013) 361–385.
[71] E.F. Toro, V.A. Titarev, Solution of the generalised Riemann problem for advection-reaction equations, Proc. R. Soc. Lond. A 458 (2002) 271–281.
[72] J. Tryoen, O. Le Maitre, M. Ndjinga, A. Ern, Intrusive projection methods with upwinding for uncertain nonlinear hyperbolic systems, J. Comput. Phys.,
Elsevier 229 (18) (2010) 6485–6511.
[73] D. Xiu, G.E. Karniadakis, Modeling uncertainty in flow simulations via generalized polynomial chaos, J. Comput. Phys. 187 (2003) 137–167.
[74] D. Xiu, G.E. Karniadakis, Modeling uncertainty in steady state diffusion problems via generalized polynomial chaos, Comput. Methods Appl. Mech. Eng.
191 (2002) 4927–4948.
[75] D. Xiu, G.E. Karniadakis, The Wiener-Askey polynomial chaos for stochastic differential equations, SIAM J. Sci. Comput. 24 (2) (2002) 619–644.
[76] X. Yang, M. Choi, G. Lin, G.E. Karniadakis, Adaptive ANOVA decomposition of stochastic incompressible and compressible flows, J. Comput. Phys. 231
(2012) 1587–1614.
[77] R.W. Walters, L. Huyse, Uncertainty Analysis for Fluid Mechanics with Applications, ICASE Rep. no. 2002-1, 2002.
[78] X. Wan, G. Karniadakis, An adaptive multi-element generalized polynomial chaos method for stochastic differential equations, J. Comput. Phys. (2005)
617–642.
[79] N. Wiener, The homogeneous chaos, Am. J. Math. (1938) 897–936.
[80] J.A.S. Witteveen, H. Bijl, Efficient quantification of the effect of uncertainties in advection-diffusion problems using polynomial chaos, Numer. Heat
Transf., Part B 53 (2008) 437–465.
[81] Z. Zhang, X. Yang, I.V. Oseledets, G.E. Karniadakis, L. Daniel, Enabling high-dimensional hierarchical uncertainty quantification by ANOVA and tensor-
train decomposition, IEEE Trans. Comput.-Aided Des. Integr. Circuits Syst. 20 (2015).