Numerical methods for high-speed flows 1

Numerical methods for high-speed flows

Sergio Pirozzoli
Università di Roma ‘La Sapienza’
Dipartimento di Meccanica e Aeronautica
Via Eudossiana 18, 00184 Roma, Italy;
e-mail: [email protected]

Key Words shock waves, shock-capturing schemes, energy conservation, nu-

merical dissipation

Abstract We review numerical methods for DNS and LES of turbulent compressible flow in the
presence of shock waves. Ideal numerical methods should be accurate and free from numerical
dissipation in smooth parts of the flow, and, at the same time, they must robustly capture shock
waves without significant Gibbs ringing, which may lead to nonlinear instability. Adapting to
these conflicting goals leads to the design of strongly nonlinear numerical schemes that depend
on the geometrical properties of the solution. For low-dissipation methods for smooth flows,
numerical stability can be based on physical conservation principles for kinetic energy and/or
entropy. Shock-capturing requires addition of artificial dissipation, in more or less explicit form,
as a surrogate of physical viscosity, to obtain non-oscillatory transitions. Methods suitable for
both smooth and shocked flows are discussed, and the potential for hybridization is highlighted.
Examples of the application of advanced algorithms to DNS/LES of turbulent, compressible
flows are presented.

CONTENTS

INTRODUCTION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

METHODS FOR SMOOTH FLOWS . . . . . . . . . . . . . . . . . . . . . . . . . 5

Upwinding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
Filtering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
Energy-Consistent Schemes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
Entropy-Consistent Schemes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
Numerical Tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

METHODS FOR SHOCKED FLOWS . . . . . . . . . . . . . . . . . . . . . . . . . 10

‘Classical’ Shock-Capturing Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
Hybrid Schemes and Nonlinear Filtering . . . . . . . . . . . . . . . . . . . . . . . . . . 12
Artificial Viscosity Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
Shock-capturing through Sub-Grid Scale Models . . . . . . . . . . . . . . . . . . . . . . 16
Properties of Shock-Capturing Schemes . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

FURTHER TOPICS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
Complex Geometries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
Time Integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
Boundary Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
Discretization of Viscous Fluxes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
 Annu. Rev. Fluid Mech. 2011

APPLICATIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

SUMMARY AND CONCLUSIONS . . . . . . . . . . . . . . . . . . . . . . . . . . 22

1 INTRODUCTION
Numerical simulation of high-speed flows has a long history, dating back to the
beginning of the computer era (von Neumann & Richtmyer, 1950; Courant et al.,
1952; Godunov, 1959). Several textbooks on numerical methods have appeared
over the years (Richtmyer & Morton, 1967; Hirsch, 1988; LeVecque, 1990; Laney,
1998; Toro, 2009). Important advancements have been made, but, as illustrated
by a recent comparative study (Johnsen et al., 2010) computational gasdynamics
has not yet converged to an ‘optimal’ computational strategy. The purpose of
this review is to check the status of the discipline, and systematize the enormous
amount of material produced over the years, a large part of which was reviewed by
Ekaterinaris (2005). Given the vastness of the subject, we mostly limit ourselves
to analyze the family of finite-difference (FD) schemes that are frequently used,
especially in the academic community, for direct numerical simulation (DNS)
and large-eddy simulation (LES) of compressible turbulent flows. Resolving the
range of scales present in these flows requires numerical schemes that are accurate,
robust, and efficient in terms of CPU requirements.
The reference physical model consists in the compressible Navier-Stokes equa-
tions for a calorically perfect gas, here written in integral form in Cartesian
coordinates
Z X3 Z
d
u dV + (fi − fiv ) ni dS = 0, (1)
dt V ∂V
i=1

where
     
ρ ρui 0
u =  ρuj  , fi =  ρui uj + pδij  , fiv =  σij , j = 1, 2, 3,
ρE ρui H, σik uk − qi ,
(2)
are, respectively, the vector of conservative variables, and the vectors of the
convective and viscous fluxes in the i-th direction. Here ρ is the density, ui is
the velocity component in the i-th coordinate direction, p is the thermodynamic
pressure, E = e + ρu2 /2 is the total energy per unit mass, e = RT /(γ − 1) is
the internal energy per unit mass, H = E + p/ρ is the total enthalpy, R is the
gas constant, γ = cp /cv is the specific heat ratio, σij is the viscous stress tensor,
and qi is the heat flux vector. At high Mach numbers, the occurrence of very
strong shock waves and of severe viscous heating may bring to light additional
effects that are not incorporated in (1), such as chemical reactions and non-ideal
thermodynamic behavior (real-gas effects). These are not addressed in the present
review, and the interested reader should refer to specific publications (Bertin &
Cummings, 2006).
Under the assumption of smooth flow, the Navier-Stokes equations can be
equivalently cast in differential form
3 3
∂u X ∂fi X ∂f v
i
+ = . (3)
∂t ∂xi ∂xi
i=1 i=1

2
Numerical methods for high-speed flows 3

Neglect of molecular diffusion effects (i.e. setting µ = 0, k = 0) leads to the

Euler equations, that only incorporate the effects of macroscopic convection and
molecular collisional effects through pressure forces. The Euler equations have
several important mathematical properties that are illustrated in the books of
Lax (1973), Majda (1959), Hirsch (1988), and LeVecque (1990). Some useful
properties for the development of numerical methods are briefly recalled here.
1. Hyperbolicity. The system of Euler equations can be cast in characteristic
form, meaning that projection of the equations in any spatial direction
gives rise to a system of coupled wave-like equations (Majda, 1959). This
motivates the study of the model one-dimensional scalar conservation law
∂u ∂f (u) ∂u ∂u
+ = + a(u) = 0, (4)
∂t ∂x ∂t ∂x
which is used as a prototype for the development of numerical methods.
It can also be shown that the convective flux vectors are homogeneous
functions of order one with respect to the vector of conservative variables,
i.e. fi = dfi /du · u, a property that is useful for the design of upwind
methods based on the flux vector splitting technique (Steger & Warming,
1981).
2. Conservation properties. The Euler equations have the obvious property
(as clear from their integral form) that the integrals of ρ, ρui , and ρE over
an arbitrary control volume can only vary due to flux through the bound-
aries. Under the assumption of smooth flow, combining the continuity and
the momentum equations and integrating yields a balance equation for the
kinetic energy ρuk uk /2
Z Z Z
d ∂ui
ρuk uk /2 dV = − (ρuk uk /2 + p) ui ni dS + p dV. (5)
dt V ∂V V ∂xi

Equation (5) shows that the total kinetic energy only varies because of mo-
mentum flux through the boundary or to volumetric work of pressure forces
(which is zero for incompressible flow), whereas the convective terms do
not cause any net variation. This property has inspired numerical schemes
based on the attempt to enforce ‘kinetic energy preservation’ in the dis-
crete sense. Additional conservation laws can be derived from the Euler
equations for smooth flows, namely (Harten, 1983b)
Z Z
d
ρg(s) dV + ρg(s) ui ni dS = 0, (6)
dt V ∂V

where g(s) is an arbitrary (but differentiable) function of the thermody-

namic entropy s = log(pρ−γ ), which suggests the possibility to design
numerical schemes that discretely preserve the integral of ρg(s). In the
presence of shock waves, mechanical energy is transformed into heat (with
subsequent production of entropy) even in the limit of vanishing viscosity,
and energy/entropy conservation no longer holds.
3. Symmetrization. Harten (1983b) showed that the change of variables v =
∂ρg(s)/∂u allows to recast the Euler equations in the ‘symmetrized’ form
d
∂v X ∂v
P + Bi = 0, (7)
∂t ∂xi
i=1
4 Pirozzoli

where P = ∂u/∂v, Bi = ∂fi /∂v are symmetric matrices, and P is positive-

definite. The symmetrized form of the equations has the important property
that continuous energy estimates can be established (Olsson, 1993; Olsson
& Oliger, 1994). Indeed, referring to a 1D bounded interval [a, b], from (7)
follows that Z
d b
v · Pv dx = − [v · Bv]ba , (8)
dt a
showing that the generalized energy at the left-hand-side can only vary in
time due to boundary contributions. Tadmor (1984) showed that valid-
ity of the energy estimate (8) implies satisfaction of entropy conservation
(6). The choice g(s) = β exp (s/(β (γ − 1))) (Harten, 1983b), leads to sim-
ple transformations between conservative and entropy variables (Sandham
et al., 2002), and u, fi happen to be homogeneous functions of degree β
with respect to v,

u(θv) = θβ u(v), fi (θv) = θβ fi (v). (9)

The solutions of (1) exhibit tendency to form steep gradients (shock waves
and contact discontinuities), whose thickness is of the order of magnitude of the
mean-free-path, making shock ‘resolution’ with numerical methods infeasible in
most cases of practical interest. In the inviscid limit shock waves reduce to zero-
measure objects, across which the Rankine-Hugoniot jump conditions must be
satisfied. In order for the class of ‘weak’ solutions of the Euler equations to
coincide with that of the Navier-Stokes equations in the inviscid limit, additional
conditions (in the form of ‘entropy inequalities’) must be satisfied, to ensure
that the macroscopic effects of the small scale diffusion processes are properly
represented. For a thorough mathematical discussion of these issues the reader
may consult Lax (1973), Majda (1959), and LeVecque (1990).
To summarize, high-speed flows typically feature regions where the flow is
smooth, and the governing equations in their differential form hold, interspersed
by extremely thin regions, where the flow properties vary abruptly. Therefore,
it is not surprising that numerical methods for high-speed flows have specialized
into two classes, capable to deal with smooth flows and with shock waves, respec-
tively, and having very different properties. Indeed, it is known that standard
discretizations used for smooth flows cause (potentially dangerous) Gibbs oscilla-
tions in the presence of shock jumps, whereas typical methods used to regularize
shock calculations exhibit excessive ‘numerical’ viscosity. Schemes of the two fam-
ilies are illustrated in Sections 2 and 3, respectively. Tailored numerical methods
for high-speed flows should be able to put together the favourable properties of
the two basic families, which is frequently realized through their hybridization or
blending, as also explained in Section 3. The following discussion will mostly deal
with the discretization of the convective derivatives that appear in (3), this being
the main source of nonlinearity in the governing equations. As customary in the
computational gas-dynamics community, a method-of-lines approach is assumed,
whereby time integration is handled separately, using ODE dedicated methods.
Time integration, as well as discretization of the viscous fluxes and specification
of boundary conditions, are addressed in Section 4.
Numerical methods for high-speed flows 5

2 METHODS FOR SMOOTH FLOWS

The most obvious choice to discretize the governing equations in smooth parts of
the flow field is using high-order approximations of the derivative operators that
appear in (3). Straightforward semi-discretization of the model equation (4) on
a grid with uniform spacing h and nodes xj = j · h yields
dvj
= −Dfj , (10)
dt
where vj (t) ≈ u (xj , t), Dfj ≈ ∂f /∂x (xj , t).
A class of explicit and implicit (compact) approximations for Dfj was devised
by Lele (1992),
M
X L
1 X
am Dfj+m = bℓ fj+ℓ , (11)
h
m=−M ℓ=−L
where fj = f (vj (t)), and appropriate values of the coefficients am , bℓ , are de-
termined so as to maximize formal accuracy (the minimum attainable trun-
cation error being O(h2(L+M ) )), or to shape the spectral response of the re-
sulting scheme and improve the representation of the Fourier modes with the
highest wavenumbers supported by the computational mesh (Vichnevetsky &
Bowles, 1982). The latter idea has produced the flourishing of the so-called DRP
(Dispersion-Relation-Preserving) schemes (Tam & Webb, 1993; Zingg et al., 1993;
Lui & Lele, 2002; Bogey & Bailly, 2004). Central difference approximations, re-
covered for b−ℓ = −bℓ , a−m = am , have null dissipation error in a linear set-
ting (Lele, 1992), and are therefore ideal candidates for computations free of
spurious numerical diffusivity.
It the context of compressible flows is useful to cast the approximation (10) in
‘locally conservative’ form, i.e. look for a numerical flux fˆj+1/2 , such that
1 ˆ 
Dfj = fj+1/2 − fˆj−1/2 . (12)
h
For linear representations of the first derivative of the type (11), explicit formulas
for the corresponding numerical fluxes are easily found
M
X L
X
am fˆj+1/2+m = cℓ fj+ℓ , (13)
m=−M ℓ=−L+1

where c−L+1 = −a−L , cℓ = cℓ−1 − aℓ−1 , ℓ = −L + 2, . . . , L. Explicit (even nonlin-

ear) approximations of the numerical flux of the type fˆj+1/2 = fˆ(vj−L+1 , . . . , vj+L ),
that satisfy the consistency condition fˆ(v, . . . , v) = f (v), guarantee global dis-
crete conservation of the integral of u for the scalar model equation (and of total
mass, momentum and energy for the Euler equations) through the telescopic
property, and satisfy the hypotheses of the Lax-Wendroff theorem (Lax & Wen-
droff, 1960; LeVecque, 1990), which guarantees convergence to weak solutions,
provided the approximation itself converges.
Central derivative approximations have been widely used in the literature, es-
pecially for wave propagation problems where nonlinearities are weak (Colonius
& Lele, 2004). However, it is known that application of standard central dis-
cretizations to high-Reynolds-number fluid turbulence typically leads to numeri-
cal instability, owing to accumulation of the aliasing errors resulting from discrete
6 Pirozzoli

evaluation of the nonlinear convective terms (Phillips, 1959). Such deficiency can
also be traced back to failure to discretely preserve quadratic invariants associ-
ated with the conservation equations (Lilly, 1965; Orlandi, 2000). While finite
viscosity may help to stabilize calculations, it is usually safer to revert to al-
ternative discretization techniques capable of ensuring stability in the inviscid
limit.

2.1 Upwinding
The ‘upwinding’ approach, commonly followed in the gas-dynamics community,
is based on the idea that solutions of the Euler equations propagate along char-
acteristics, and therefore a stable numerical method should also propagate its
information in the same characteristic direction (Moretti, 1979). The standard
practice (referred to as ‘flux vector splitting’ (Steger & Warming, 1981)) in the
FD framework is to split the flux function into ‘positive’ and ‘negative’ parts
f (u) = f + (u) + f − (u), associated, respectively, with non-negative and non-
positive propagation velocities, i.e. df + /du ≥ 0, df − /du ≤ 0, which are dis-
cretized by means of left- and right-biased approximations, respectively, thus
ensuring linear stability.
In the finite-volume (FV) framework upwinding is usually achieved through
the ‘flux difference splitting’ (or Godunov) approach. Considering a FV semi-
discretization of (4),
dv j 1

=− f (vj+1/2 ) − f (vj−1/2 ) , (14)
dt h
Rx
where v j (t) = 1/h x j+1/2 v(x, t)dx, is the spatial average of the approximate
j−1/2
solution over the cell Ij = (xj−1/2 , xj+1/2 ), a suitable reconstruction operator
is used to determine approximate left and right states at the cell interfaces,
±
vj+1/2 , and the interface flux f (vj+1/2 ) is replaced with the numerical flux re-
sulting from an exact (or approximate) Riemann solver (Roe, 1981; Toro, 2009),
formally fˆj+1/2 = R(vj+1/2
− +
, vj+1/2 ). For example, Roe’s approximate Riemann
solver (Roe, 1981) dictates

− + 1  − +
 |a
j+1/2 |

+ −

R(vj+1/2 , vj+1/2 )= f (vj+1/2 ) + f (vj+1/2 ) − vj+1/2 − vj+1/2 ,
2 2
(15)
+ − + −
where aj+1/2 = (f (vj+1/2 ) − f (vj+1/2 ))/(vj+1/2 − vj+1/2 ) is a characteristic speed
associated with the intermediate state. In the original formulation of Godunov
(1959), piece-wise constant reconstructions were assumed, leading to the general
first-order flux
fˆj+1/2 = R(v j , v j+1 ). (16)
The extension to higher order of accuracy is achieved by replacing piece-wise
constant with piece-wise polynomial reconstructions (van Leer, 1979). In the
ADER (Arbitrary accuracy DERivative Riemann problem) approach (Toro et al.,
2001), the numerical flux is evaluated by solving the generalized Riemann problem
using a semi-analytical method. The approximate solution is given as a Taylor
time expansion at the cell interface position, thus allowing the construction of
one-step schemes with arbitrary accuracy in both space and time.
Numerical methods for high-speed flows 7

Upwinding has the main effect of damping the Fourier modes with highest
supported wavenumbers, with subsequent stabilizing effect on the numerical so-
lution. Upwind schemes have often been used for DNS of shock-free compressible
turbulence (Rai & Moin, 1993; Foysi et al., 2004; Pirozzoli et al., 2008), with a
good degree of success. However, as pointed out by Park et al. (2004), the numer-
ical dissipation introduced by upwinding can be harmful for LES, where proper
resolution of marginally resolved wavenumbers is crucial, as it may hamper the
effect of sub-grid scale models.

2.2 Filtering
Filtering the computed solution is a commonly used practice to cure nonlinear
instabilities of central schemes, while retaining high-order accuracy. High-order
filters were introduced by Lele (1992), and used by Visbal & Gaitonde (1999)
and Visbal & Gaitonde (2002) for the solution of conservation laws on stretched,
curvilinear, and deforming meshes. The latter authors considered compact filters
of the form
L
X dℓ
αv̂j−1 + v̂j + αv̂j+1 = (vj+ℓ + vj−ℓ ) , (17)
2
ℓ=0
where the coefficients dℓ at the r.h.s. are constrained to be functions of the free
parameter α through Taylor- and Fourier-series analyses, so as to achieve high-
order accuracy, and selectively damp the supported Fourier modes with highest
wavenumbers. A selective filtering strategy was developed by Bogey & Bailly
(2004) based on optimized explicit filters of the type (17) with α = 0, L = 4, 5, 6.
Those authors also suggested possible use of the selective filtering to model the
effects of the unresolved (sub-grid) flow scales, replacing classical eddy-viscosity
models. As for upwinding, filtering can eliminate spurious instabilities, but it
implies artificial energy draining that adds to the physical energy draining caused
by molecular viscosity. More accurate control over diffusive effects is achieved by
reverting to numerical methods capable of conserving energy in the absence of
viscosity.

2.3 Energy-Consistent Schemes

Several attempts have been made to design nonlinearly stable numerical schemes
by replicating the energy preservation properties of the governing equations in
the discrete sense. Most efforts are based on the idea of recasting the convective
terms in skew-symmetric form, that loosely amounts to expanding the convective
derivatives found in (3) as either
∂ρui ϕ 1 ∂ρui ϕ 1 ∂ρui 1 ∂ϕ
= + ϕ + ρui , (18)
∂xi 2 ∂xi 2 ∂xi 2 ∂xi
as proposed by Feiereisen et al. (1981), or as
∂ρui ϕ 1 ∂ρui ϕ 1 ∂ρϕ 1 ∂ui
= + ui + ρϕ , (19)
∂xi 2 ∂xi 2 ∂xi 2 ∂xi
as proposed by Blaisdell et al. (1996), where ϕ stands for a generic transported
scalar property, being unity for the continuity equation, uj for the momentum
8 Pirozzoli

equation, H for the total energy equation. Discretization of the mass and mo-
mentum equations in the split form (18) implies kinetic energy preservation at the
semi-discrete level (Honein & Moin, 2004), provided the difference operators sat-
isfy the summation by parts (SBP) property (Strand, 1994) that automatically
holds in the case of periodic problems for central (both explicit and implicit)
approximations of any order (Mansour et al., 1979). From another viewpoint,
discretization of (19) guarantees minimization of the aliasing error (Blaisdell
et al., 1996). Additional robustness in the presence of strong density variations
is gained (Kennedy & Gruber, 2008) by expanding the convective derivatives in
the generalized form
 
∂ρui ϕ ∂ρui ϕ ∂ρϕ ∂ui ϕ ∂ρui
=α + β ui +ρ +ϕ
∂xi ∂xi ∂xi ∂xi ∂xi
 
∂ϕ ∂ui ∂ρ
+ (1 − α − 2β) ρui + ρϕ + ui ϕ . (20)
∂xi ∂xi ∂xi
This arrangement leads to semi-discrete energy conservation in the case α = β =
1/4. Ducros et al. (2000) showed that the skew-symmetric forms (18) and (19)
give rise to locally conservative schemes, when the derivative operators are dis-
cretized with explicit central formulas (with order up to six). Pirozzoli (2010)
has recently proved that this holds true for explicit central formulas of arbitrary
order of accuracy, and presented computationally inexpensive numerical fluxes
also for the skew-symmetric form (20). Apparently, compact derivative approxi-
mations applied to the skew-symmetric split form of the convective terms do not
lead to locally conservative schemes. For a formal mathematical formulation of
fully conservative, skew-symmetric splitting of the compressible Euler equations,
the reader may refer to Morinishi (2010).

2.4 Entropy-Consistent Schemes

Stabilization of numerical schemes can also be achieved by enforcing, at the dis-
crete level, the conservation properties associated with entropy. Tadmor (1987)
developed a second-order FV, locally conservative discretization of the Euler
equations, that discretely satisfies (6), resulting in the formulation of the nu-
merical flux Z 1
f̂j+1/2 = f (vj + ξ (vj+1 − vj )) dξ, (21)
0
which was also considered by Jameson (2008b). Note that Equation (21) is ex-
pensive, as it requires the evaluation of the entropy variables associated with
the left and right states, and the use of quadrature formulas to approximate the
integral. More general expressions for the numerical flux of entropy-consistent
schemes can be found in Tadmor (2003).
An alternative strategy to ensure entropy stability was proposed by Gerritsen
& Olsson (1996), who split the flux vector into conservative and non-conservative
parts, exploiting the homogeneity property (9)
∂f β ∂f 1 ∂v
= + B . (22)
∂x β + 1 ∂x β + 1 ∂x
This splitting, referred to in the literature as ‘canonical splitting’ or ‘entropy split-
ting’, leads to a discrete analogue of the energy estimate (8) (with consequent
Numerical methods for high-speed flows 9

benefits in terms of nonlinear stability) if the first derivatives in (22) are replaced
with SBP operators. The resulting schemes are referred to as split high-order en-
tropy conserving (SHOEC), since discrete entropy conservation in the sense (6)
follows (Tadmor, 1984). The performance of SHOEC schemes is apparently af-
fected by the choice of the splitting parameter β. Sandham et al. (2002) suggested
that a value of β ≈ 4 is appropriate for numerical simulation of compressible tur-
bulence, whereas β ≈ 2 is more appropriate for numerical simulation of isotropic
turbulence at zero viscosity (Honein & Moin, 2004). As all formulations of the
Euler equations that rely on the use of entropy variables, SHOEC schemes suf-
fer for reduced computational efficiency, owing to swapping from conservative to
entropy variables, that requires CPU-intensive transcendental operations.
Honein & Moin (2004) developed entropy-consistent schemes by applying the
skew-symmetric splitting to the Euler equations, upon replacement of the total
energy equation with the entropy equation. This approach preserves the integrals
of ρs and ρs2 . An equivalent alternative is to use the internal or total energy
equations, rearranged in such a way that the skew-symmetric form of the entropy
equation automatically follows.

2.5 Numerical Tests

The robustness of the numerical methods considered so far can be appreciated
from numerical simulations of isotropic compressible turbulence at zero viscos-
ity on a 323 grid, as proposed by Honein & Moin (2004). The calculations are
initialized with a field of synthetic isotropic solenoidal turbulence with assigned
wavenumber spectrum E(k) ∼ (k/k0 )4 exp −2(k/k0 )2 , peaking at k0 = 6, and
fluctuations of temperature and density are initially set to zero. Time integra-
tion is performed by means of a third-order SSP Runge-Kutta algorithm (Shu &
Osher, 1989) (see Section 4.2). This computational arrangement causes energy
to be quickly transferred to the smallest scales, posing a significant challenge to
numerical algorithms, and leading to nonlinear instability in the absence of the
aliasing control mechanisms discussed above.
The results are reported in Figure 1 for initial turbulent Mach numbers Mt0 =
0.07, 0.3. In the former case, the ‘exact’ solution, computed by means of a de-
aliased spectralR method (Honein & Moin, 2004), has nearly R constant total kinetic
energy (K = ρu2 dV ), constant total entropy (S = ρs dV ), and the r.m.s.
density levels off to ρ′ /ρ0 /Mt 20 ≈ 0.35, after an initial transient. As seen from
Figure 1, central discretization of the equations cast in divergence form (schemes
D2, D6) leads to divergence of the numerical solution on a time scale shorter
than one eddy turnover time, regardless of the order of accuracy. Divergence is
associated with increase of kinetic energy, and loss of total entropy. Upwinding
(scheme D7) curves the divergence, but dissipates most of the kinetic energy
by ten eddy turnover times, and leads to an entropy increase. The results of
simulations performed at Mt 0 = 0.07 show that skew-symmetric splitting of the
convective terms (schemes S6, K6) helps conserve kinetic energy and entropy,
although spurious growth of density fluctuations is observed. This is also true
of Tadmor entropy conserving scheme (T2) that exhibits larger errors over long
times not leading to divergence though. The entropy-consistent scheme of Honein
& Moin (2004) (H6) succeeds in correctly predicting the limiting inviscid behavior
of the system for this specific test case. The results at Mt0 = 0.3 show failure of
all methods at large times, except the skew-symmetric scheme based on (20) and
10 Pirozzoli

Tadmor scheme.

3 METHODS FOR SHOCKED FLOWS

All methods presented in the previous section suffer from spurious Gibbs oscil-
lation near shock jumps, that may lead to nonlinear instabilities. The onset of
oscillations can be avoided following two strategies. In the ‘shock-fitting’ ap-
proach, shock waves are treated (separately from the rest of the flow) as genuine
discontinuities, whose dynamics is governed by their own algebraic equations,
and the Rankine-Hugoniot relations are used as a set of nonlinear boundary con-
ditions to relate the states on the two sides of the discontinuity (Moretti, 1987).
In the ‘shock-capturing’ approach the same discretization scheme is used at all
points, and regularization is achieved through addition of numerical dissipation,
that prevents (or at least limits) the onset of Gibbs oscillations. While the for-
mer approach very often guarantees more accurate representations of shocked
flows (Johnsen et al., 2010), it is only feasible in cases where the shock topology
is extremely simple, and no shock wave forms during the calculation. In this
review, only the shock-capturing approach is discussed.

3.1 ‘Classical’ Shock-Capturing Methods

Historically, most shock-capturing schemes have been constructed having in mind
the mathematical properties of entropy solution of the model scalar conservation
law (4), including:
1. monotonicity preservation (MP), i.e. initially monotone data remain mono-
tone for all times (Godunov, 1959);
2. total variation diminishing (TVD), i.e. solutions of (4) satisfy T V (u(·, t2 )) ≤
T V (u(·, t1 )), ∀t2 ≥ t1 , where T V (u(·, t)) denotes the total variation of
u (Harten, 1983a);
3. monotonicity, i.e. let u(x, t), v(x, t) be any two entropy solutions of (4),
with u(x, 0) ≤ v(x, 0), ∀x, then u(x, t) ≤ v(x, t), ∀x, t (Harten et al., 1986).
Shock-capturing schemes implementing the above conditions in the discrete sense
are abundant in the literature, the main reason being the availability of theorems
that guarantee convergence to weak solutions for TVD schemes (LeVecque, 1990)
and to the (unique) entropy solution for monotone schemes (Harten et al., 1986).
However, monotone schemes are at most first-order accurate (Harten et al., 1986),
which explains the wide success enjoyed by TVD schemes, especially in the 1980’s.
An excellent overview of the contributions to TVD schemes was reported by Yee
(1989).
Classical TVD schemes were found to suffer from loss of accuracy at both
smooth and non-smooth extrema, which stimulated researchers to pursue alter-
natives for constructing uniformly high-order accurate shock-capturing schemes.
The very successful family of essentially-non-oscillatory (ENO) schemes (Harten
et al., 1987), was based on the idea of determining the numerical flux from
high-order (say r) reconstruction over an adaptive stencil that is selected to
avoid as much as possible interpolation across discontinuities, thus minimizing
Gibbs oscillations. ENO reconstructions of order r guarantee that T V (v n+1 ) ≤
T V (v n ) + O(hr ), which implies uniform boundedness of the total variation. ENO
Numerical methods for high-speed flows 11

schemes were found to suffer problems of convergence to steady solutions and

loss of accuracy, due to free stencil adaptation (Rogerson & Meiburg, 1990). The
problem was cured by Shu (1990) by biasing towards the ‘best’ central stencil in
the adaptation procedure. An alternative technique to circumvent the accuracy
reduction phenomenon of TVD schemes, based on the enforcement of the MP
constraint at discrete level, was proposed by Suresh & Huynh (1997).
The class of the weighted essentially-non-oscillatory (WENO) schemes, first
introduced by Liu et al. (1994), and generalized and improved by Jiang & Shu
(1996), is based on the idea of constructing a high-order numerical flux from a
convex linear combination of lower order polynomial reconstructions over a set
of staggered stencils, with weights selected to achieve maximum formal order of
accuracy in smooth regions, and assign (nearly) zero weight to reconstructions on
stencils crossed by discontinuities. As illustrated in Figure 2, L + 1 sub-stencils,
each including L points, Sℓ = {xj−L+ℓ+1 , . . . , xj+ℓ } , ℓ = 0, . . . , L, are considered
out of the overall reconstruction stencil S = {xj−L+1 , . . . , xj+L }, that includes
a total of 2L points. Assuming for now f (u) = f + (u), and suppressing the
superscripts, the FD WENO numerical flux is constructed as
L
X (ℓ)
fˆj+1/2 = ωℓ fˆj+1/2 , (23)
ℓ=0
(ℓ)
where fˆj+1/2 is the numerical flux resulting from polynomial reconstruction over
the stencil Sℓ ,
L−1
X
(ℓ)
fˆj+1/2 = cℓm fj−L+1+ℓ+m , (24)
m=0
and the weights, defined as
αℓ dℓ
ωℓ = PL , αℓ = , (25)
m=0 αm
(ǫ + βℓ )2
are functions of the ‘smoothness measurements’ associated with the sub-stencils
L−1
!2
X L−1X
βℓ = γℓmn fj−L+1+ℓ+n . (26)
m=1 n=1

The key to the success of WENO reconstructions is the proper selection of the
‘linear’ weights dℓ , so as to achieve maximum order of accuracy (2L), and the
design of the ‘nonlinear’ weights ωℓ , such that if the stencil Sℓ contains a jump,
ωℓ ≈ 0 follows. Upwinding is achieved by suitable biasing in the reconstruction
procedure. Referring to the ‘positive’ part of the flux, in the original approach of
Jiang & Shu (1996) a left-biased reconstruction was obtained discarding the right-
most sub-stencil (i.e. dL = 0), thus reducing the overall stencil width (and the
formal order of accuracy) to 2L − 1. Weirs & Candler (1997) suggested the use of
a central stencil to reduce the numerical viscosity of WENO schemes, and biased
the reconstruction by re-defining the smoothness measurement for the right-most
sub-stencil, setting βL = max0≤ℓ≤L βℓ , thus preventing the selection of the fully
downwinded stencil (SL ). Formulas for the ‘negative’ split flux are easily obtained
by symmetry about xj+1/2 . Comprehensive compilations of WENO coefficients
cℓm , dℓ , γℓmn , can be found in Weirs & Candler (1997), Balsara & Shu (2000)
and Gerolymos et al. (2009).
12 Pirozzoli

Provable properties of WENO schemes are limited to the scalar case, and
include convergence for smooth solutions (Jiang & Shu, 1996), and existence and
stability of discrete shock profiles for the third-order version of the scheme (Jiang
& Yu, 1996). Nevertheless, WENO schemes are extremely robust in practice,
and have been applied successfully (both in their FD and FV versions) to a wide
number of physical problems. For an overview of the wide-ranging ramifications
of the WENO idea the reader may consult the review paper by Shu (2009).
Many investigators have tried to improve the properties of WENO schemes,
especially in terms of their behavior in smooth regions. The mainstream of re-
search (Weirs & Candler, 1997) has been focused on the attempt to modify the
linear weights and the polynomial reconstruction coefficients to improve the be-
havior in Fourier space and extend the range of well resolved wavenumbers, in the
spirit of the DRP approach. Follow-up studies of the ‘optimized’ WENO strat-
egy have been reported by Hill & Pullin (2004), Martı́n et al. (2006), and Taylor
et al. (2007). An alternative approach to conjugate the spectral-like resolution
properties of compact schemes with the shock-capturing properties of WENO was
undertaken by Deng & Zhang (2000), who developed weighted compact nonlin-
ear schemes (WCNS) by coupling WENO interpolation of conservative variables
to the intermediate nodes and compact conservative approximation of the flux
derivatives at the grid nodes. In the author’s experience, WENO schemes and
their variants are highly accurate, and quite robust. However, they are compu-
tationally expensive, mostly because of the large number floating point opera-
tions required for the evaluation of the smoothness measurements. Furthermore,
spectral-like resolution is never achieved owing to the noxious effect of the nonlin-
ear weights, unless the weights are locally frozen in smooth regions, as suggested
by Hill & Pullin (2004) and Taylor et al. (2007).
A visual impression of the performance of WENO schemes is gained from Fig-
ure 3, where the results of numerical simulations of the test case 13 of Lax &
Liu (1998) are reported. The solution of the two-dimensional Riemann problem
involves a complex shock diffraction pattern leading to the formation of a Mach
stem, and Kelvin-Helmholtz roll-up of the vertical slip-line. Figure 3 shows that
both the TVD and the WENO methods robustly capture the shock wave pattern,
without significant spurious oscillations. Significant differences are observed in
the predicted structure of the slip line, that exhibits finer scale roll-up for the
WENO schemes, and the formation of secondary vortex cores for the seventh-
order scheme.

3.2 Hybrid Schemes and Nonlinear Filtering

This class of methods is based on the idea of endowing a baseline spectral-like
scheme with shock-capturing capability through local replacement with a classical
shock-capturing scheme, or through controlled addition of the dissipative part of
a shock-capturing scheme, that is made to act as a ‘nonlinear filter’. A key role
in this class of schemes is played by ‘shock sensors’, that must to be defined in
such a way that numerical dissipation is effectively confined in shocked regions,
not to pollute smooth parts of the flow field.
Lee et al. (1997) used a hybrid discretization made up of a sixth-order compact
scheme and a sixth-order shock-capturing ENO scheme, to analyze the interac-
tion of a (nearly) planar shock wave with a field of isotropic turbulence. Since
the shock location in that configuration is approximately known, the switch be-
Numerical methods for high-speed flows 13

tween the two schemes was established a-priori, and ENO was only activated
in the shock-normal direction, over twelve points upstream and twelve points
downstream of the mean shock location. Adams & Shariff (1996) first considered
a truly adaptive hybrid discretization, consisting of a baseline compact upwind
scheme coupled with a fifth-order ENO scheme. A simple switch, based on the
local gradient of the flux vector components, was used to mark critical cells, that
were subsequently padded with buffer cells on each side, to prevent the onset of
oscillations arising from the coupling of schemes with different properties. The
method was expanded by Pirozzoli (2002), who developed a fully conservative
formulation by hybridizing a fifth-order compact upwind numerical flux (of the
form (13)) with a seventh-order WENO flux, the switch being based on the local
density gradient. Figure 3d shows the results obtained with this method for the
Lax-Liu Riemann problem. Compared to baseline WENO schemes, the hybrid
formulation yields more accurate representation of smooth flow features, with
the formation of additional discrete vortices in the shear layer, but it also yields
some spurious wiggles near shocks. For this specific test case, the reduction of
computational cost with respect to standard seventh-order WENO is about 50%.
Further improvements to hybrid WENO schemes were introduced by Ren et al.
(2003) and Hill & Pullin (2004). In particular, the latter authors used a baseline
central DRP-like discretization (tuned for LES of compressible turbulence) and
designed a shock sensor based on the ratio of the largest to the smallest WENO
smoothness indicators λ = (maxℓ βℓ )/(minℓ βℓ + ε), that becomes large whenever
one (or more) WENO sub-stencils are crossed by discontinuities.
While hybrid schemes are now frequently used, it appears that the issues related
to coupling of the underlying schemes have not been thoroughly investigated.
A notable exception is the work of Larsson & Gustafsson (2008), who applied
Kreiss stability theory to the analysis of hybrid methods. Interestingly, it was
found that the coupling of two separately stable schemes may, in some instances,
give rise to an unstable system. However, it was also found that, if either scheme
is dissipative, the coupled system is strongly stable, thus endorsing the use of
upwind shock-capturing schemes around discontinuities.
The nonlinear filtering approach was introduced by Yee et al. (1999), who de-
signed a low-dissipative, shock-capturing algorithm based on SHOEC discretiza-
tion, augmented with the artificial dissipation of a TVD scheme. Referring to
the scalar conservation law, the following semi-discretization was proposed
dvj 1  ∗ ∗

= −Dfj − fj+1/2 − fj−1/2 , (27)
dt h
where the filter numerical flux is defined as
∗
fj+1/2 = k Ψj+1/2 dj+1/2 . (28)

Here k is a (problem-dependent) constant, dj+1/2 is the artificial flux of a second-

order TVD scheme, and Ψj+1/2 is the Harten switch (Harten, 1978), which is
designed to be near unity near shocks, and near zero in smooth parts of the
flow. As a consequence, the algorithm automatically localizes numerical dissi-
pation around discontinuities. The extension of the methodology to the multi-
dimensional Euler equations involves application of the nonlinear filter to each
characteristic field, hence the frequently used terminology of ‘characteristic-based
filters’ also given to this class of schemes. An advantage of the nonlinear filtering
14 Pirozzoli

approach over hybrid methods is that, in the case of explicit multi-stage time in-
tegration, the filter numerical flux can be applied at the end of the full time step,
rather than at each sub-step, with substantial saving in terms of CPU time (but
with some loss in terms of robustness). Ducros et al. (1999) modified the original
nonlinear filtering approach by adopting the artificial flux of the JST scheme (as
from Equation (30)), and replacing in Equation (28) the Harten switch with the
product of the Jameson sensor (32) and another shock indicator
(∇ · u)2
Θ= , 0 ≤ Θ ≤ 1, (29)
(∇ · u)2 + (∇ × u)2 + ǫ
frequently referred to as the Ducros sensor. Further modifications of the method
were introduced by Garnier et al. (2001), who replaced the TVD artificial flux
with the dissipative part of ENO and WENO schemes, and showed that the
Ducros sensor is capable to distinguish turbulent fluctuations from shocks better
than the Harten switch. The idea of adaptively filtering the numerical solution for
shock-capturing also underlies the work of Visbal & Gaitonde (2005) and Bogey
et al. (2009), who developed computational strategies based on high-accuracy
central discretization of spatial derivatives, supplemented with selective filtering
of the computed solution. Shock-capturing is achieved by locally decreasing the
filter bandwidth in critical regions.
A key ingredient in the formulation of hybrid and nonlinear filtering schemes
consists in the proper specification of the shock sensor. The Ducros sensor rep-
resents a simple and frequently used choice. Alternative simple shock sensors
include those proposed by Hill & Pullin (2004), Visbal & Gaitonde (2005), and
Larsson et al. (2007) (the latter being a modification of the Ducros sensor). More
elaborate sensors for shock waves and shears, based on multi-resolution wavelet
analysis were proposed by Sjögreen & Yee (2000) and Yee & Sjögreen (2007).
The performance of shock sensors in practical computations can be appreciated
from inspection of Figure 4, where we report results of application of some of the
indicators previously mentioned to the instantaneous flow field obtained from
DNS of transonic shock/boundary layer interaction (Pirozzoli et al., 2010). The
threshold values for the various sensors have been selected in such a way that
similar representation of the shock system is obtained. Note that the Ducros
sensor in its original formulation does not perform properly outside of the wall
layer (where ∇ × u ≈ 0), due to excessive sensitivity to dilatational fluctuations.
The sensor can be conveniently adapted to this case by setting ǫ = (u∞ /δ0 )2
in Equation (29), so that it is only activated when the local dilatation becomes
larger than a typical large-scale velocity gradient. Also note that, for evaluation
purposes, all other sensors are based on the pressure field. The data reported in
Figure 4 suggest that the modified Ducros sensor is capable to selectively isolate
‘genuine’ shocks, whereas other simple sensors also mark as critical zones smooth
regions populated by vortical structures.

3.3 Artificial Viscosity Methods

The basic idea of this class of methods is to explicitly introduce the amount of nu-
merical dissipation needed to stabilize shock computations through the addition
of diffusive terms that adaptively adjust to the local regularity of the solution.
In the original approach of von Neumann & Richtmyer (1950), artificial diffu-
sive terms with viscosity coefficient µ∗ = cµ ρh2 |∂u/∂x| (where cµ is a tunable
Numerical methods for high-speed flows 15

constant), were added to the momentum and energy equations to spread shock
waves to a thickness comparable to the local mesh spacing. The artificial vis-
cosity idea was expanded by Jameson et al. (1981), who designed a high-order
non-oscillatory numerical flux by augmenting the numerical flux of a second-order
central scheme (fˆC ), with explicit diffusive fluxes resulting from discretization of
second- and fourth-derivative operators, as follows
fˆj+1/2 = fˆj+1/2
C
− dj+1/2 ,
h i
(2) (4)
dj+1/2 = aj+1/2 ǫj+1/2 (vj+1 − vj ) − ǫj+1/2 (vj+2 − 3vj+1 + 3vj − vj−1 ) , (30)

where in the case of the Euler equations the characteristic speed aj+1/2 is replaced
with the spectral radius of the Jacobian matrix of the inviscid flux at Roe’s
average state (Roe, 1981). The diffusion coefficients , defined as
 
(2) (4) (2)
ǫj+1/2 = k (2) ψj+1/2 , ǫj+1/2 = max 0, k (4) − ǫj+1/2 , (31)

are functions of the local regularity of the solution through

|vj+1 − 2vj + vj−1 |
ψj+1/2 = max (ψj , ψj+1 ) , ψj = , 0 ≤ ψj ≤ 1, (32)
|vj+1 + 2vj + vj−1 |

and k (2) ≈ 1, k (4) ≈ 0.01 − 0.05. The Jameson-Schmidt-Turkel (JST) artificial

flux is designed in such a way that second-order dissipation is activated only near
shocks, whereas fourth-order dissipation only acts in smooth regions to suppress
high-frequency oscillations. In the case of the Euler equations, regularity of the
solution is gauged by applying (32) to the pressure field.
von Neumann’s original idea was extended by Tadmor (1989) and Guo et al.
(2001), who introduced a spectrally accurate vanishing viscosity to augment the
approximation of nonlinear conservation laws in spectral space. It was shown
that a suitable choice of the viscosity kernel in Fourier space recovers spectral
convergence for weak solutions of Burgers equation. Cook & Cabot (2004, 2005)
carried over the idea of selectively damping the highest resolved wavenumbers to
physical space, through addition of an artificial stress tensor
∗
σij = 2µ∗ Sij + (β ∗ − 2/3 µ∗ ) uk,k δij , (33)
where Sij = 1/2 (ui,j + uj,i ). The artificial shear and bulk viscosity coefficients
are defined as
µ∗ = crµ ηr , β = crβ ηr , ηr = ρ ∆2(r+1) |∇d
2r S|, (34)
where crµ , crβ are user-specified constants, ∆ is a measure of the local mesh spac-
ing, S = (2Sij Sij )1/2 , ∇2r denotes the polyharmonic operator (i.e. the r-th power
of the Laplace operator), and the caret denotes a truncated Gaussian filter, de-
signed to have smooth ηr . A set of coefficients used by most investigators is r = 2,
c2β = 1, c2µ = 0.002. Compared to the von Neumann formulation (that is recovered
for r = 1), the choice (34) imparts on the numerical viscosity a higher wavenum-
ber bias. Inclusion of the bulk viscosity term is the key to capture shock waves
without affecting the vorticity field, whereas the shear viscosity has the effect of
a Smagorinsky-like sub-grid scale model for unresolved vortical fluctuations. As
observed by Mani et al. (2009), the definition (34) introduces unnecessary damp-
ing into resolved dilatational motions (acoustic waves). To cure such deficiency,
16 Pirozzoli

Mani et al. (2009) replaced S with uk,k in (34), in the definition of crβ , and set
crβ to zero whenever uk,k > 0 (i.e. in expansion zones). Additional selectivity in
the definition of crβ was introduced by Bhagatwala & Lele (2009), in order that
it effectively becomes zero in the presence of resolved acoustic motions.
Further improvements of the artificial viscosity method have led to the inclusion
of the effects of strong temperature gradients (contact discontinuities), as well as
species diffusion in multi-component mixtures (Fiorina & Lele, 2007; Cook, 2007).
The extension of the method to generalized curvilinear coordinates was presented
by Kawai & Lele (2007).

3.4 Shock-capturing through Sub-Grid Scale Models

Methods of this class are based on the attempt to regularize weak solutions of
the conservation equations through the addition of sub-grid scale models that
drain energy from the unresolved scales of motion, in analogy to what is done for
LES of smooth flows. This idea is related to the artificial viscosity method, and
was elaborated by Adams & Stolz (2002). Those authors developed a method
based on explicit filtering of the flow variables, whereby the closure terms are
derived from an approximate deconvolution of the flow scales represented on
the computational grid, and entropy regularization is achieved by addition of a
relaxation term driven by the amount of energy found at the smallest resolved
scales. Referring to the scalar conservation law (4), the method of Adams & Stolz
(2002) can be formulated as an evolution equation

∂ û ∂f (û)
+ = G1 + R, (35)
∂t ∂x
for the filtered variable û = G ⋆ u, where G is a (primary) low-pass filtering
operator. The term G1 accounts for the effect of the computed flow scales (i.e.
those represented on the computational grid) on the resolved ones (i.e. on the
filtered field), and is approximated as

∂f (û) ∂f (ũ)
G1 = − , (36)
∂x ∂x
where ũ = Q ⋆ û, Q being a regularized approximation of the exact deconvolution
operator G−1 . The effect of the interaction between the non-represented scales
and the resolved scales is further modeled through the relaxation term

R = −χ (û − G2 ⋆ û) , (37)

where G2 is a secondary (more selective) low-pass filter, and χ is an adjustable

relaxation constant. The modified equation (35) is then discretized through high-
order compact differencing. A weak point of the formulation is the specification
of a suitable value for the relaxation constant χ, that apparently cannot be es-
tablished a-priori. A dynamic procedure to estimate χ as a function of space and
time was described by von Kaenel et al. (2004).

3.5 Properties of Shock-Capturing Schemes

The analysis of shock-capturing schemes is made difficult by their inherent non-
linearity, which prevents straightforward application of the tools extensively used
Numerical methods for high-speed flows 17

to evaluate linear schemes, especially Fourier analysis (Vichnevetsky & Bowles,

1982). Nonlinearity also makes comparison of the relative quality of different
methods, other than a case-by-case check, almost impossible. Very often, classi-
cal shock-capturing schemes have linear counterparts, that are used as a guidance
to improve their spectral behavior. For example, freezing the nonlinear weights
turns WENO schemes into linear upwind schemes, whose approximate disper-
sion relation has been used in the development of the optimized-WENO schemes.
However, numerical tests show that the nonlinear features have a dramatic impact
on the performance of numerical schemes, and their actual behavior may be very
different from linear predictions. For instance, going back to Figure 1, one can see
a substantially poorer performance (in terms of increased numerical dissipation)
of the seventh-order WENO scheme (W7) compared to its ‘linearized’ seventh-
order upwind version (D7). For a discussion of the impact of nonlinearities on the
behavior of shock-capturing schemes in Fourier space see Pirozzoli (2006). Useful
information related to the performance of shock-capturing schemes for shock-free
turbulent flows can be found in Garnier et al. (1999) and Larsson et al. (2007).
A major flaw of shock-capturing schemes, often disregarded, is the reduction
of accuracy near shocks. Indeed, even (nominally) uniformly high-order schemes
yield first-order accurate solutions downstream of moving shocks, the main rea-
son being that shocks are objects with zero measure, and as such, their location is
only known to O(h) on a finite grid. The loss of accuracy phenomenon was high-
lighted in model scalar shock-sound interaction problems (Casper & Carpenter,
1998; Engquist & Sjogreen, 1998), and it is likely to be the cause of slow conver-
gence of apparently simple shock/turbulence interaction calculations (Larsson &
Lele, 2009). For an interpretation of the pathology in terms of numerical viscosity
the reader may refer to Efraimsson & Kreiss (1999). Shock-capturing schemes
applied to systems of conservation laws also suffer from spurious post-shock os-
cillations, especially in the case of slowly moving shock (Arora & Roe, 1997),
which may prevent the accurate prediction of shock/sound and shock/turbulence
interactions, as observed by Johnsen et al. (2010). These pathologies are ap-
parently unavoidable, unless one reverts to special techniques, such as sub-cell
resolution (Harten, 1989), or even to shock-fitting.

4 FURTHER TOPICS
All methods presented so far are focused on the discretization of the spatial
derivatives that appear in the governing equations. Other issues, related to time
integration and specification of numerical boundary conditions, as well as the
extension to complex geometries, are briefly addressed in the present Section.

4.1 Complex Geometries

The extension of FD schemes to multi-dimensional systems of conservations laws
in Cartesian domains is easily obtained through component-by-component (or
characteristic-wise), line-wise application of one-dimensional algorithms. Rela-
tively complex geometries, that can be smoothly mapped to a Cartesian mesh,
can be handled exploiting the ‘strong conservation form’ of the equations, (Vi-
nokur, 1974) that allows extension of locally conservative approximations derived
for uniformly spaced Cartesian meshes, thus allowing convergence to weak solu-
tions. The ‘immersed boundary’ approach (Mittal & Iaccarino, 2005) promises
18 Pirozzoli

simple handling of complex geometries in the FD framework, representing a vi-

able alternative to the adoption of body-fitted grids. However, relatively few
applications of the method to compressible flows have appeared so far (Hu et al.,
2006; de Tullio et al., 2007; Ghias et al., 2007).
Finite-volume methods provide greater flexibility than FD in dealing with com-
plex geometries, since locally conservative FV schemes can be easily designed
for both structured and unstructured meshes. The FV framework also allows
the construction of one-step methods in time (as opposed to multi-stage Runge-
Kutta time integration commonly used for FD, see Section 4.2) having high-
order accuracy in both space and time, which makes ADER schemes competitive
with typical shock-capturing FD schemes having comparable accuracy (Titarev
& Toro, 2005). Furthermore, it appears that a FV formulation, with use of
suitable positivity-preserving approximate Riemann solvers, is necessary in some
instances, such as the computation of compressible multi-component flows, to
avoid oscillations in the presence of material interfaces (Johnsen & Colonius,
2006). With regard to FV methods for structured meshes, it is known that
straightforward dimensional splitting gives rise to second-order errors (regard-
less of the accuracy of the underlying reconstructions), unless computationally
expensive quadratures are used to evaluate the flux integrals transverse to the di-
rection being reconstructed. High-order accurate, quadrature-based FV schemes
have been developed both in the classical method-of-lines approach with stan-
dard Riemann solvers (Titarev & Toro, 2004), and in the one-step ADER frame-
work (Titarev & Toro, 2005). However, as observed by Ducros et al. (2000) and
Gerolymos et al. (2009), the splitting error is usually small, and line-wise ap-
plication of 1D reconstructions is quite successful in practice, yielding accurate
representation of smooth flow features and good shock-capturing properties. Un-
structured meshes mandate the use of both high-order flux quadratures at cell
interfaces and of genuinely multi-dimensional reconstructions (Hu & Shu, 1999),
thus making high-order FV schemes highly expensive. An important step in the
direction of improving the computational efficiency of high-order FV schemes
was accomplished by Dumbser et al. (2007), who succeeded in designing one-step
non-oscillatory FV schemes for unstructured tetrahedral meshes with arbitrary
order of accuracy, without the need of quadratures. Their strategy exploits a
characteristic WENO reconstruction yielding the whole polynomial information
in each cell, and a Cauchy-Kovalewski procedure to provide a space-time Taylor
series for the conserved quantities and the physical fluxes. This information is
used to construct highly accurate upwind numerical fluxes, that are subsequently
integrated analytically in space and time. A comprehensive review of modern FV
methods is reported in the book of Toro (2009).
Alternative strategies to achieve uniform high accuracy for flows in complex ge-
ometries, especially in an unstructured mesh framework, include the classes of the
spectral volumes (SV) and discontinuous Galerkin (DG) methods (Wang, 2007).
However, these methods require the storage of several degrees-of-freedom per cell,
and therefore seem to be currently too demanding in terms of memory require-
ments for large-scale turbulence computations. A novel approach was proposed
by Dumbser et al. (2008), who developed a unified framework for the construction
of fully-discrete and very high-order accurate quadrature-free one-step FV and
DG schemes on unstructured tetrahedral meshes (the so-called PN PM schemes),
where the data are represented in each cell by piecewise polynomials of degree N ,
and the fluxes are computed in using piecewise polynomials of degree M ≥ N .
Numerical methods for high-speed flows 19

The class of FV methods is then recovered for N = 0, whereas DG methods are

recovered for N = M . The intermediate classes of schemes with M > N and
N 6= 0 were found to produce computationally more efficient algorithms than
conventional FV and DG methods.
Some recent studies have addressed the possibility of constructing low-dissipative,
nonlinearly stable schemes for complex geometries. Kok (2009) designed fourth-
order accurate schemes suitable for smooth structured meshes, that preserve the
mathematical skew symmetry of the convective terms. Jameson (2008a) and
Subbareddy & Candler (2009) have developed second-order FV methods suitable
for unstructured meshes, that discretely preserve kinetic energy. In particular,
in the latter study, fully discrete energy conservation was obtained through a
density-weighted Crank-Nicholson-like time integration, and shock-capturing was
incorporated in the method in the form of a TVD filter controlled by the Ducros
sensor.

4.2 Time Integration

Time integration in the method-of-lines framework is usually performed by means
of explicit Runge-Kutta multi-stage schemes (Kennedy et al., 2000). Special types
of so-called strong-stability-preserving (SSP) Runge-Kutta schemes are frequently
used (Shu & Osher, 1989; Gottlieb et al., 2001), having the remarkable property
that if a semi-discretization of the scalar conservation law (4) is stable in a certain
norm (e.g. the TV norm) with first-order Euler forward time stepping, the multi-
stage time discretization will maintain stability in that norm, under a suitable
CFL restriction. It has been argued (Shu, 1997) that the use of SSP Runge-
Kutta time discretizations is safer for solving hyperbolic problems in the presence
of shock waves. In the author’s experience, and as confirmed by inspection of
the literature on the subject, standard Runge-Kutta time integration will work
out just as fine for most of typical DNS and LES applications. Optimized low-
dissipation and low-dispersion Runge-Kutta algorithms, originally introduced for
computational aeroacoustics (Hu et al., 1996), especially in their low-storage
implementation (Stanescu & Habashi, 1998), can also be fruitfully adopted for
the purpose. Recent developments on the subject are reported in Bogey & Bailly
(2004), Berland et al. (2006) and Bernardini & Pirozzoli (2009).

4.3 Boundary Conditions

The correct enforcement of boundary conditions in compressible flow simulations
is of utmost importance for accuracy and stability of numerical algorithms. An
exhaustive account of the subject can be found in Colonius (2004). A critical
issue in the simulation of spatially evolving wall-bounded compressible flows,
not covered in that review, is the specification of suitable inlet conditions, that
must guarantee a fully developed turbulent state of the boundary layer at the
shortest possible distance from the inflow. A frequently used technique is the
rescaling-recycling procedure developed by Lund et al. (1998), and extended to
the compressible case by Sagaut et al. (2004) and Xu & Martı́n (2004). The basic
idea consists of extracting a cross-stream slice of the flow field and recycling it to
the inflow, after suitable rescaling. This approach produces a realistic turbulent
boundary layer within a short distance from the inflow, and it allows to control
the skin friction and the thickness of the simulated boundary layer. Alterna-
20 Pirozzoli

tive techniques rely on the specification of synthetic disturbances at the inlet

constructed so as to mimic coherent boundary layer structures (Sandham et al.,
2003; Pamiès et al., 2009), or on the attempt to synthesize quasi-deterministic
disturbances having desired first- and second-order velocity statistics at the in-
flow (Klein et al., 2003).

4.4 Discretization of Viscous Fluxes

Discretization of the viscous terms in the compressible Navier-Stokes equations
involves consecutive derivative operations. As observed by Lele (1992), Sandham
et al. (2002) and Nagarajan et al. (2003), better accuracy and robustness is
achieved by expanding the diffusive terms as
 
∂ ∂u ∂ 2 u ∂µ ∂u
µ =µ 2 + , (38)
∂x ∂x ∂x ∂x ∂x
and applying dedicated second derivative discretization to the first term at the
r.h.s. This treatment is justified by spectral analysis of the approximation, show-
ing significantly improved representation of the diffusive effects at the highest
resolved wavenumbers, compared to straightforward repeated application of dis-
crete first derivative operators, that may result in odd-even decoupling.

5 APPLICATIONS
Applications of the methods of computational gasdynamics to the analysis of
flow physics are innumerable. We limit ourselves to highlight the most recent
applications to high-fidelity simulation of flows involving the interaction of shock
waves with turbulence.
The simplest setting consists of the interaction of a (nominally) normal shock
wave with a field of isotropic turbulence. Canonical shock/turbulence interac-
tions were first investigated through DNS by Lee et al. (1997), using a hybrid
compact/ENO scheme. LES of the same problem was performed by Ducros
et al. (1999), by means of a characteristic-based nonlinear filtering scheme. The
problem was revisited by Larsson & Lele (2009), who carried out high-resolution
calculations using a hybrid central skew-symmetric/WENO discretization.
Substantial efforts in the last decades have been devoted to the analysis of
shock / boundary layer interactions (SBLI) (Délery & Marvin, 1986). The first
DNS study of SBLI was reported by Adams (2000), who investigated the flow
over a 18◦ ramp at free-stream Mach number M∞ = 3, using the hybrid com-
pact/ENO method of Adams & Shariff (1996). DNS of a 24◦ compression ramp
configuration at M∞ = 2.9, was performed by Wu & Martı́n (2007), using the
bandwidth-optimized WENO algorithm of Taylor et al. (2007). LES of the su-
personic ramp flow was carried out by Rizzetta et al. (2001), using the adaptive
filtering technique of Visbal & Gaitonde (2005), and by von Kaenel et al. (2004),
using the regularization method of Adams & Stolz (2002).
Another frequently used prototype SBLI consists of the reflection of an oblique
shock wave from a flat plate where a boundary layer is developing. The first LES
of impinging shock interaction was performed by Garnier et al. (2002), who used
a baseline central fourth-order discretization augmented with a nonlinear WENO
filter, whose local activation was controlled by the Ducros sensor. Pirozzoli &
Grasso (2006) performed a DNS study with flow conditions similar to Garnier
Numerical methods for high-speed flows 21

et al. (2002), using a seventh-order WENO scheme. LES of the impinging shock
interaction have also been performed by Touber & Sandham (2009). Their nu-
merical method relied on a baseline SHOEC scheme, and shock-capturing was
achieved through a TVD filter based controlled by the Ducros sensor. Sample
results of the calculations of Touber & Sandham (2009) are reported in Figure 5,
showing a comparison between the mean LES field (in terms of streamwise veloc-
ity and turbulent shear stress), and reference experimental data (Dupont et al.,
2008). Overall, the LES results are in very good agreement with experimental
PIV data, the most apparent difference being the size of the separation bubble.
However, the boundary layer thickening is very well captured, as well as the am-
plification of the Reynolds shear stress past the interaction zone, associated with
the shedding of vortices.
Shock/boundary layer interactions also occur under transonic conditions. The
first LES of transonic SBLI over a circular-arc bump was performed by Sandham
et al. (2003), using the SHOEC scheme with nonlinear TVD filtering. Pirozzoli
et al. (2010) have recently reported DNS results of transonic SBLI at M∞ =
1.3 over a flat plate using a hybrid approach, whereby smooth flow regions are
handled by means of conservative sixth-order central discretization of the skew-
symmetric split form (20), and shock waves are captured through a seventh-order
WENO scheme, the switch being based on the Ducros sensor. An illustrative
result of that calculation is reported in Figure 6, where the shock system is
captured through the modified Ducros sensor, and vortical structures through
iso-surfaces of the ‘swirling strength’ (Zhou et al., 1999). The figure highlights the
three-dimensional nature of the lambda-like shock pattern, whereby the vortical
structures in the incoming boundary layer cause the spanwise wrinkling of the
upstream compression fan. Numerous hairpin-shaped vortex loops, resembling
those found in incompressible boundary layer DNS, are observed both in the
upstream boundary layer and past the interacting shock.
The turbulent mixing resulting from the injection of an under-expanded sonic
jet in a supersonic cross-flow was studied by Kawai & Lele (2010) by means of
LES. Those authors used the artificial viscosity method in the version of Kawai &
Lele (2007), employing sixth-order compact approximations of the spatial deriva-
tives, coupled with eight-order low-pass filtering. The flow visualizations reported
in Figure 7 demonstrate the capability of the numerical method to capture the
front bow shock, the upstream separation shock, the barrel shock, and the Mach
stem, all without spurious wiggles, and at the same time to accurately resolve a
broad range of turbulence scales.
Low-dissipative shock-capturing methods have also been used by Hill et al.
(2006) to analyze Richtmyer-Meshkov instability with reshock. Those authors
used a DRP-like central approximation of the skew-symmetric split equations
in smooth regions, and switched to the tuned-WENO scheme of Hill & Pullin
(2004) near shock waves, the switch being controlled by the local ‘curvature’ of
the pressure and density field, in a fashion similar to the Jameson sensor.
Many other applications of low-dissipative shock-capturing algorithms are col-
lected in the review paper of Ekaterinaris (2005).
22 Pirozzoli

6 SUMMARY AND CONCLUSIONS

We have presented (some of the) currently available numerical methods for high-
speed flows, with emphasis on FD schemes of current use for LES and DNS of
compressible flows with shocks. Owing to their great computational efficiency,
FD methods are preferable for large-scale computations whenever the physical
problem allows use of uniform Cartesian or smooth curvilinear meshes.
Two classes of methods have been discussed, specifically designed for appli-
cation to smooth and shocked flows, respectively. With regard to the former
class, several mathematical principles can be exploited for the design of stable,
low-dissipative schemes that are superior to classical upwinding or filtering. In
particular, nonlinearly stable schemes can be designed by enforcing at the discrete
level properties inherent to the Euler equations in the absence of shocks, namely
preservation of the total kinetic energy and conservation properties related to the
thermodynamic entropy. The results of several investigators, and numerical tests
herein reported, support the effectiveness of these methods in providing nonlinear
stability for the calculation of high-Reynolds-number flows, without addition of
numerical viscosity or filtering. A survey of the current literature leads me to give
some favour to energy-preserving schemes that rely on expansion of convective
derivatives to skew-symmetric form. Numerical schemes stemming from central
explicit discretization of equations in skew-symmetric split form are computation-
ally efficient, have good nonlinear stability properties, and can be formulated in
locally-conservative form. In particular, the variant of skew-symmetric splitting
proposed by Kennedy & Gruber (2008) seems to be promising for smooth flows
with strong density variations.
Regarding methods for shocked flows, the class of the WENO schemes (and
their variants) seem to have superseded other shock-capturing methods in the
last decade, having proven to be extremely accurate and robust in the pres-
ence of strong shock waves and complex shock interactions. WENO schemes,
however, are quite CPU intensive, and suffer from excessive numerical damp-
ing in smooth zones of the flow field. Therefore their application to LES and
DNS is suggested only in hybrid form, i.e. in conjunction with a non-dissipative
algorithm to treat smooth flow zones. Effective strategies for coupling shock-
capturing and non-dissipative methods include hybridization and filtering. Both
strategies rely on shock sensors that have to be as simple and effective as possi-
ble. One such choice is the Ducros sensor, that was found to perform reasonably
well in many shock/turbulence interaction problems, particularly with some mi-
nor case-by-case modification and tuning of the threshold value for activation of
shock-capturing. With regard to the hybrid methods, a suggested best practice
is marking critical nodes using a shock sensor, and then pad a sufficient number
of nodes around the critical ones to make sure that the stencil of the underlying
non-dissipative scheme does not cross shocked zones.
An attractive alternative to hybrid shock-capturing methods is offered by
the nonlinear artificial viscosity methods, especially in their most recent vari-
ants (Mani et al., 2009). These methods have the advantage to be applicable
in the same form throughout the computational domain, without the need of
hybridization, and that their implementation in existing compressible Navier-
Stokes solvers for shock-free flows is straightforward. Applications of the method
are recent and further testing is required to demonstrate comparable reliability
to WENO for a range of applications.
Numerical methods for high-speed flows 23

In all cases, a locally conservative discretization is encouraged. Although it

is frequently stated that non-conservative methods are capable to handle shock
waves properly, a locally conservative formulation (based on numerical fluxes)
automatically guarantees global conservation of integral properties through the
telescopic property, and applicability of the Lax-Wendroff theorem. Note there
is no extra cost incurred with the use of a locally conservative discretization, and
numerical fluxes are available for skew-symmetric and entropy-consistent schemes
(see Sections 2.3,2.4). Furthermore, hybridization with shock-capturing schemes
(always formulated in conservation form) is more natural.
Some open issues remain. First, one must be aware that the global order of
accuracy of shock-capturing schemes in unsteady problems is always reduced to
unity, and shock-capturing is the cause of spurious oscillations, especially down-
stream of slowly moving shocks. These limitations, related to the misrepresen-
tation of discontinuities on a mesh with finite spacing, can only be overcome by
some form of shock-fitting. A detailed study of the effect of shock-capturing os-
cillations on the prediction of shock/sound and shock/turbulence interactions is
lacking, which would be highly desirable. Second, even though hybrid schemes are
frequently used, a systematic quantitative analysis of the coupling between shock-
capturing and non-dissipative schemes has not been carried out so far (however,
hints are found in the work of Larsson & Gustafsson (2008)). Third, a compar-
ative efficiency analysis of numerical algorithms (in terms of CPU cost for given
error tolerance) for problems involving shock waves is not available at present,
and cost figures are seldom reported in computational studies. Fourth, it ap-
pears that efficient, low-dissipative methods suitable for compressible turbulence
simulation on unstructured meshes are lacking in the literature, one notable ex-
ception being the recent work of Subbareddy & Candler (2009). In this respect,
alternative avenues to FV are also worth being explored (Dumbser et al., 2008).
Further efforts are needed before computational gasdynamics can reach a fully
mature stage, and cope with the growing demand for DNS and LES of high-speed
turbulent flows for configurations of technological relevance.

NUMBERED LITERATURE CITED

References
Adams NA. 2000. Direct simulation of the turbulent boundary layer along a
compression ramp at M = 3 and Reθ = 1685. J. Fluid Mech. 420:47–83
Adams NA, Shariff K. 1996. A high-resolution hybrid compact-ENO scheme for
shock-turbulence interaction problems. J. Comput. Phys. 127:27–51
Adams NA, Stolz S. 2002. A subgrid-scale deconvolution approach for shock-
capturing. J. Comput. Phys. 178:391–426
Arora M, Roe PL. 1997. On postshock oscillations due to shock capturing schemes
in unsteady flows. J. Comput. Phys. 130:25–40
Balsara D, Shu CW. 2000. Monotonicity preserving weighted essentially non-
oscillatory schemes with increasingly high order of accuracy. J. Comput. Phys.
160:405–452
Berland J, Bogey C, Bailly C. 2006. Low-dissipation and low-dispersion fourth-
order Runge-Kutta algorithm. Comp. & Fluids 35:1459–1463
24 Pirozzoli

Bernardini M, Pirozzoli S. 2009. A general strategy for the optimization of Runge-

Kutta schemes for wave propagation phenomena. J. Comput. Phys. 228:4182–
4199
Bertin JJ, Cummings RM. 2006. Critical hypersonic aerothermodynamic phe-
nomena. Annu. Rev. Fluid Mech. 38:129–157
Bhagatwala A, Lele SK. 2009. A modified artificial viscosity approach for com-
pressible turbulence simulations. J. Comput. Phys. 228:4965–4969
Blaisdell GA, Spyropoulos ET, Qin JH. 1996. The effect of the formulation of
non-linear terms on aliasing errors in spectral methods. Appl. Numer. Math.
21:207–219
Bogey C, Bailly C. 2004. A family of low dispersive and low dissipative explicit
schemes for flow and noise computations. J. Comput. Phys. 194:481–491
Bogey C, de Cacqueray N, Bailly C. 2009. A shock-capturing methodology based
on adaptive spatial filtering for high-order non-linear computations. J. Comput.
Phys. 228:1447–1465
Casper J, Carpenter MH. 1998. Computational considerations for the simulation
of shock-induced sound. SIAM J. Sci. Comput. 19:813–828
Colonius T. 2004. Modeling artificial boundary conditions for compressible flow.
Annu. Rev. Fluid Mech. 36:315–345
Colonius T, Lele SK. 2004. Computational aeroacoustics: progress on nonlinear
problems of sound generation. Progr. Aero. Sci. 40:345–416
Cook AW. 2007. Artificial fluid properties for large-eddy simulation of compress-
ible turbulent mixing. Phys. Fluids 19:055103
Cook AW, Cabot WH. 2004. A high-wavenumber viscosity for high-resolution
numerical methods. J. Comput. Phys. 195:594–601
Cook AW, Cabot WH. 2005. Hyperviscosity for shock-turbulence interactions.
J. Comput. Phys. 203:379–385
Courant R, Isaacson E, Rees M. 1952. On the solution of non-linear hyperbolic
differential equations. Comm. Pure Appl. Math. 5:243–255
de Tullio MD, Palma PD, Iaccarino G, Pascazio G, Napolitano M. 2007. An
immersed boundary method for compressible flows using local grid refinement.
J. Comput. Phys. 225:2098–2117
Délery J, Marvin JG. 1986. Shock-wave boundary layer interactions. AGARDo-
graph 280
Deng X, Zhang H. 2000. Developing high-order weighted compact nonlinear
schemes. J. Comput. Phys. 165:22–44
Ducros F, Ferrand V, Nicoud F, Weber C, Darracq D, et al. 1999. Large-eddy
simulation of the shock/turbulence interaction. J. Comput. Phys. 152:517–549
Ducros F, Laporte F, Soulères T, Guinot V, Moinat P, Caruelle B. 2000. High-
order fluxes for conservative skew-symmetric-like schemes in structures meshes:
application to compressible flows. J. Comput. Phys. 161:114–139
Dumbser M, Balsara DS, Toro EF, Munz CD. 2008. A unified framework for the
construction of one-step finite-volume and discontinuous Galerkin schemes on
unstructured meshes. J. Comput. Phys. 227:8209–8253
Numerical methods for high-speed flows 25

Dumbser M, Käser M, Titarev VA, Toro EF. 2007. Quadrature-free non-

oscillatory finite volume schemes on unstructured meshes for nonlinear hy-
perbolic systems. J. Comput. Phys. 226:204–243
Dupont P, Piponniau S, Sidorenko A, Debieve JF. 2008. Investigation by particle
image velocimetry measurements of obliques shock reflection with separation.
AIAA J. 46:1365–1370
Efraimsson G, Kreiss G. 1999. A remark on numerical errors downstream of
slightly viscous shocks. SIAM J. Numer. Anal. 36:853–863
Ekaterinaris JA. 2005. High-order accurate, low numerical diffusion methods for
aerodynamics. Progr. Aero. Sci. 41:192–300
Engquist B, Sjogreen B. 1998. The convergence rate of finite difference schemes
in the presence of discontinuities. SIAM J. Numer. Anal. 35:2464–2485
Feiereisen WJ, Reynolds WC, Ferziger JH. 1981. Numerical simulation of a
compressible, homogeneous, turbulent shear flow. Report TF 13, Thermosci.
Div., Mech. Eng., Stanford University
Fiorina B, Lele SK. 2007. An artificial nonlinear diffusivity method for supersonic
reacting flows with shocks. J. Comput. Phys. 222:246–264
Foysi H, Sarkar S, Friedrich R. 2004. Compressibility effects and turbulence
scalings in supersonic channel flow. J. Fluid Mech. 509:207–216
Garnier E, Mossi M, Sagaut P, Comte P, Deville M. 1999. On the use of shock-
capturing for large-eddy simulation. J. Comput. Phys. 153:273–311
Garnier E, Sagaut P, Deville M. 2001. A class of explicit ENO filters with appli-
cation to unsteady flows. J. Comput. Phys. 170:184–204
Garnier E, Sagaut P, Deville M. 2002. Large-Eddy Simulation of shock/boundary-
layer interaction. AIAA J. 40:1935–1944
Gerolymos GA, Sénéchal D, Vallet I. 2009. Very high-order WENO schemes. J.
Comput. Phys. 228:8481–8524
Gerritsen M, Olsson P. 1996. A stable high order finite difference scheme and
sharp shock resolution for the Euler equations. J. Comput. Phys. 129:245–262
Ghias R, Mittal R, Dong H. 2007. A sharp interface immersed boundary method
for compressible viscous flows. J. Comput. Phys. 225:528–553
Godunov SK. 1959. A difference scheme for numerical computation of discontin-
uous solution of hydrodynamic equations. Math. Sbornik 47:271–306
Gottlieb S, Shu CW, Tadmor E. 2001. Strong stability-preserving high-order time
discretization methods. SIAM Review 43:89–112
Guo BY, Ma HP, Tadmor E. 2001. Spectral vanishing viscosity method for
nonlinear conservation laws. SIAM J. Numer. Anal. 39:1254–1268
Harten A. 1978. The artificial compression method for computation of shocks and
contact discontinuities. III. Self-adjusting schemes. Math. Comp. 32:363–389
Harten A. 1983a. High resolution schemes for hyperbolic conservation laws. J.
Comput. Phys. 49:357–393
Harten A. 1983b. On the symmetric form of systems of conservation laws with
entropy. J. Comput. Phys. 49:151–164
26 Pirozzoli

Harten A. 1989. ENO schemes with subcell resolution. J. Comput. Phys. 83:148–
184
Harten A, Engquist B, Osher S, Chakravarthy SR. 1987. Uniformly high order
accurate essentially non-oscillatory schemes, III. J. Comput. Phys. 71:231–303
Harten A, Hyman JM, Lax PD. 1986. On finite-difference approximations and
entropy conditions for shocks. Comm. Pure Appl. Math. 29:297–322
Hill DJ, Pantano C, Pullin DI. 2006. Large-eddy simulation and multiscale mod-
elling of a Richtmyer-Meshkov instability with reshock. J. Fluid Mech. 557:29–
61
Hill DJ, Pullin DI. 2004. Hybrid tuned center-difference-WENO method for large
eddy simulations in the presence of strong shocks. J. Comput. Phys. 194:435–
450
Hirsch C. 1988. Numerical computation of internal and external flows. New York:
Wiley
Honein AE, Moin P. 2004. Higher entropy conservation and numerical stability
of compressible turbulence simulations. J. Comput. Phys. 201:531–545
Hu C, Shu CW. 1999. Weighted essentially non-oscillatory schemes on triangular
meshes. J. Comput. Phys. 150:97–127
Hu FQ, Hussaini MY, Manthey JL. 1996. Low-dissipation and low-dispersion
Runge-Kutta schemes for computational acoustics. J. Comput. Phys. 124:177–
191
Hu XY, Khoo BC, Adams NA, Huang FL. 2006. A conservative interface method
for compressible flows. J. Comput. Phys. 219:553–578
Jameson A. 2008a. Formulation of kinetic energy preserving conservative schemes
for gas dynamics and direct numerical simulation of one-dimensional viscous
compressible flow in a shock tube using entropy and kinetic energy preserving
schemes. J. Sci. Comp. 34:188–208
Jameson A. 2008b. The construction of discretely conservative finite volume
schemes that also globally conserve energy or entropy. J. Sci. Comp. 34:152–
187
Jameson A, Schmidt W, Turkel E. 1981. Numerical simulation of the Euler
equations by finite volume methods using Runge-Kutta time stepping schemes.
Paper 81-1259, NASA
Jiang GS, Shu CW. 1996. Efficient implementation of weighted ENO schemes.
J. Comput. Phys. 126:202–228
Jiang GS, Yu SH. 1996. Discrete shocks for finite difference approximations to
scalar conservation laws. SIAM J. Numer. Anal. 35:749–772
Johnsen E, Colonius T. 2006. Implementation of WENO schemes in compressible
multicomponent flow problems. J. Comput. Phys. 219:715–732
Johnsen E, Larsson J, Bhagatwala AV, Cabot WH, Moin P, et al. 2010. As-
sessment of high-resolution methods for numerical simulations of compressible
turbulence with shock waves. J. Comput. Phys. 229:1213–1237
Kawai S, Lele SK. 2007. Localized artificial diffusivity scheme for discontinuity
capturing on curvilinear meshes. J. Comput. Phys. 227:9498–9526
Numerical methods for high-speed flows 27

Kawai S, Lele SK. 2010. Large-eddy simulation of jet mixing in supersonic cross-
flows. AIAA J. Revision submitted
Kennedy CA, Carpenter MH, Lewis RM. 2000. Low-storage, explicit Runge-
Kutta schemes for the compressible Navier-Stokes equations. Appl. Numer.
Math. 35:177–219
Kennedy CA, Gruber A. 2008. Reduced aliasing formulations of the convective
terms within the Navier-Stokes equations. J. Comput. Phys. 227:1676–1700
Klein M, Sadiki A, Janicka J. 2003. A digital filter based generation of inflow
data for spatially developing direct numerical or large eddy simulations. J.
Comput. Phys. 186:652–665
Kok JC. 2009. A high-order low-dispersion symmetry-preserving finite-volume
method for compressible flow on curvilinear grids. J. Comput. Phys. 228:6811–
6832
Laney CB. 1998. Computational gasdynamics. Cambridge University Press
Larsson J, Gustafsson B. 2008. Stability criteria for hybrid difference methods.
J. Comput. Phys. 227:2886–2898
Larsson J, Lele SK. 2009. Direct numerical simulation of canonical
shock/turbulence interaction. Phys. Fluids 21:126101
Larsson J, Lele SK, Moin P. 2007. Effect of numerical dissipation on the predicted
spectra for compressible turbulence. Annu. Res. Briefs, Center for Turbulence
Research, Stanford University
Lax PD. 1973. Hyperbolic systems of conservation laws and the mathematical
theory of shock waves. In Regional Conference Series in Applied Mathematics.
SIAM
Lax PD, Liu XD. 1998. Solution of two-dimensional Riemann problems of gas
dynamics by positive schemes. SIAM J. Sci. Comput. 19:319–340
Lax PD, Wendroff B. 1960. Systems of conservation laws. Comm. Pure Appl.
Math. 13:217–237
Lee S, Lele SK, Moin P. 1997. Interaction of isotropic turbulence with shock
waves: effect of shock strength. J. Fluid Mech. 340:225–247
Lele SK. 1992. Compact finite difference schemes with spectral-like resolution.
J. Comput. Phys. 103:16
LeVecque R. 1990. Numerical methods for conservation laws. Basel: Birkhauser-
Verlag
Lilly DK. 1965. On the computational stability of numerical solutions of time-
dependent non-linear geophysical fluid dynamics problems. J. Comput. Phys.
93:11–26
Liu XD, Osher S, Chan T. 1994. Weighted essentially non-oscillatory schemes.
J. Comput. Phys. 115:200–212
Lui C, Lele SK. 2002. A numerical study of shock-associated noise. AIAA Paper
2002-2530, AIAA
Lund T, Wu X, Squires K. 1998. Generation of turbulent inflow data for spatially-
developing boundayr layer simulations. J. Fluid Mech. 140:233–258
28 Pirozzoli

Majda A. 1959. Compressible fluid flow and systems of conservation laws in

several space variables. In Applied Mathematical Sciences, vol. 53. Springer,
501–504
Mani A, Larsson J, Moin P. 2009. Suitability of artificial bulk viscosity for large-
eddy simulation of turbulent flows with shocks. J. Comput. Phys. 228:7368–
7374
Mansour NN, Moin P, Reynolds WC, Ferziger JH. 1979. Improved methods
for large eddy simulations of turbulence. In Turbulent Shear Flows I, eds.
BF Launder, FW Schmidt, HH Whitelaw. Springer-Verlag
Martı́n MP, Taylor EM, Wu M, Weirs VG. 2006. Optimization of nonlinear error
for weighted essentially non-oscillatory methods in direct numerical simulations
of compressible turbulence. J. Comput. Phys. 220:270–289
Mittal R, Iaccarino G. 2005. Immersed boundary methods. Annu. Rev. Fluid
Mech. 37:239–261
Moretti G. 1979. The λ-scheme. Comput. Fluids 7:191–205
Moretti G. 1987. Computation of flows with shocks. Annu. Rev. Fluid Mech.
19:313–337
Morinishi Y. 2010. Skew-symmetric form of convective terms and fully conserva-
tive finite difference schemes for variable density low-Mach number flows. J.
Comput. Phys. 229:276–300
Nagarajan S, Lele SK, Ferziger JH. 2003. A robust high-order compact method
for large eddy simulation. J. Comput. Phys. 191:392–419
Olsson P. 1993. Summation by parts, projections, and stability. RIACS Technical
Report TR-93-04, NASA Ames research center
Olsson P, Oliger J. 1994. Energy and maximum norm estimates for nonlinear
conservation laws. RIACS Technical Report TR-94-01, NASA Ames research
center
Orlandi P. 2000. Fluid Flow Phenomena: A Numerical Toolkit. Kluwer, Dor-
drecht
Pamiès M, Weiss PE, Garnier E, Deck S, , Sagaut P. 2009. Generation of syn-
thetic turbulent inflow data for large eddy simulation of spatially evolving
wall-bounded flows. Phys. Fluids 21:045103
Park N, Yoo JY, Choi H. 2004. Discretization errors in large eddy simulation:
on the suitability of centered and upwind-biased compact difference schemes.
J. Comput. Phys. 198:580–616
Phillips NA. 1959. An example of nonlinear computational instability. In The
atmosphere and the sea in motion. Rockefeller Institute Press and Oxford Uni-
versity Press, 501–504
Pirozzoli S. 2002. Conservative hybrid compact-WENO schemes for shock-
turbulence interaction. J. Comput. Phys. 178:81–117
Pirozzoli S. 2006. On the spectral properties of shock-capturing schemes. J.
Comput. Phys. 219:489–497
Pirozzoli S. 2010. Generalized numerical fluxes for skew-symmetric approxima-
tions of convective derivatives. J. Comput. Phys. Submitted
Numerical methods for high-speed flows 29

Pirozzoli S, Bernardini M, Grasso F. 2008. Characterization of coherent vortical

structures in a supersonic turbulent boundary layer. J. Fluid Mech. 613:205–
231
Pirozzoli S, Bernardini M, Grasso F. 2010. Direct numerical simulation of tran-
sonic shock/boundary layer interaction under conditions of incipient separa-
tion. J. Fluid Mech. Revision submitted
Pirozzoli S, Grasso F. 2006. Direct numerical simulation of impinging shock wave
/ turbulent boundary layer interaction at M = 2.25. Phys. Fluids 18:065113–17
Rai MM, Moin P. 1993. Direct numerical simulation of transition and turbulence
in a spatially evolving boundary layer. J. Comput. Phys. 109:169–192
Ren YX, Liu M, Zhang H. 2003. A characteristic-wise hybrid compact-WENO
scheme for solving hyperbolic conservation laws. J. Comput. Phys. 192:365–386
Richtmyer RD, Morton KW. 1967. Difference methods for initial value problems.
New York: Interscience-Wiley
Rizzetta DP, R.Visbal M, Gaitonde DV. 2001. Large-eddy simulation of super-
sonic compression-ramp flow by high-order method. AIAA J. 39:2283–2292
Roe PL. 1981. Approximate Riemann solvers, parameter vectors, and difference
schemes. J. Comput. Phys. 34:357–372
Rogerson AM, Meiburg E. 1990. A numerical study of the convergence properties
of ENO schemes. J. Sci. Comp. 5:151–167
Sagaut P, Garnier E, Tromeur E, Larcheveque L, Labourasse E. 2004. Turbulent
inflow conditions for large eddy simulation of compressible wall-bounded flows.
AIAA J 42:469–477
Sandham ND, Li Q, Yee HC. 2002. Entropy splitting for high-order numerical
simulation of compressible turbulence. J. Comput. Phys. 178:307–322
Sandham ND, Yao YF, Lawal AA. 2003. Large-eddy simulation of transonic flow
over a bump. International Journal of Heat and Fluid Flow 24:584–595
Shu CW. 1990. Numerical experiments on the accuracy of ENO and modified
ENO schemes. J. Sci. Comp. 5:127–149
Shu CW. 1997. Essentially non-oscillatory and weighted essentially non-
oscillatory schemes for hyperbolic conservation laws. Contractor Report 97-65,
NASA
Shu CW. 2009. High order weighted essentially nonoscillatory schemes for con-
vection dominated schemes. SIAM Review 51:82–126
Shu CW, Osher S. 1989. Efficient implementation of essentially non-oscillatory
shock-capturing schemes II. J. Comput. Phys. 83:32–78
Sjögreen B, Yee HC. 2000. Multiresolution wavelet based adaptive numerical
dissipation control for shock-turbulence computation. RIACS Technical Report
TR-01-01, NASA Ames research center
Stanescu D, Habashi WG. 1998. 2N-storage low dissipation and dispersion Runge-
Kutta schemes for computational acoustics. J. Comput. Phys. 143:674–681
Steger JL, Warming RF. 1981. Flux vector splitting of the inviscid gasdynamic
equations with application to finite-difference methods. J. Comput. Phys.
40:263–293
30 Pirozzoli

Strand B. 1994. Summation by parts for finite difference approximations for

d/dx. J. Comput. Phys. 110:47–67
Subbareddy PK, Candler GV. 2009. A fully discrete, kinetic energy consistent
finite-volume scheme for compressible flows. J. Comput. Phys. 228:1347–1364
Suresh A, Huynh HT. 1997. Accurate monotonicity-preserving schemes with
Runge-Kutta time stepping. J. Comput. Phys. 136:83–99
Tadmor E. 1984. Skew-selfadjoint form for systems of conservation laws. J. Math.
Anal. Appl. 103:428–442
Tadmor E. 1987. The numerical viscosity of entropy stable schemes for systems
of conservation laws. i. Math. Comp. 49:91–103
Tadmor E. 1989. Convergence of spectral methods for nonlinear conservation
laws. SIAM J. Numer. Anal. 26:30–44
Tadmor E. 2003. Entropy stability theory for difference approximations of non-
linear conservation laws and related time dependent problems. Acta Numerica
12:451–512
Tam CKW, Webb JC. 1993. Dispersion-relation-preserving finite difference
schemes for computational acoustics. J. Comput. Phys. 107:262–281
Taylor EM, Wu M, Martı́n MP. 2007. Optimization of nonlinear error for weighted
essentially non-oscillatory methods in direct numerical simulations of compress-
ible turbulence. J. Comput. Phys. 223:384–397
Titarev VA, Toro EF. 2004. Finite-volume WENO schemes for three-dimensional
conservation laws. J. Comput. Phys. 201:238–260
Titarev VA, Toro EF. 2005. ADER schemes for three-dimensional non-linear
hyperbolic systems. J. Comput. Phys. 204:715–736
Toro EF. 2009. Riemann solvers and numerical methods for fluid dynamics.
Springer
Toro EF, Millington RC, Nejad LAM. 2001. Towards very high order go-
dunov schemes. In Godunov Methods. Theory and Applications, ed. EF Toro.
Kluwer/Plenum Academic Publishers, 905–938
Touber E, Sandham ND. 2009. Large-eddy simulation of low-frequency unsteadi-
ness in a turbulent shock-induced separation bubble. Theoretical and Compu-
tational Fluid Dynamics 23:79–107
van Leer B. 1979. Towards the ultimate conservative difference scheme V. A
second order sequel to Godunov’s method. J. Comput. Phys. 32:101–136
Vichnevetsky R, Bowles JB. 1982. Fourier analysis of numerical approximations
of hyperbolic equations. Philadelphia: SIAM
Vinokur M. 1974. Conservation equations of gasdynamics in curvilinear coordi-
nate systems. J. Comput. Phys. 14:105–125
Visbal MR, Gaitonde DV. 1999. High-order accurate methods for complex un-
steady subsonic flows. AIAA J. 37:1231–1239
Visbal MR, Gaitonde DV. 2002. On the use of higher-order finite-difference
schemes on curvilinear and deforming meshes. J. Comput. Phys. 181:155–185
Visbal MR, Gaitonde DV. 2005. Shock capturig using compact-differencing-based
methods. Paper 2005-1265, AIAA
Numerical methods for high-speed flows 31

von Kaenel R, Kleiser L, Adams NA, Vos JB. 2004. Large-eddy simulation of
shock-turbulence interaction. AIAA J. 42:2516–2528
von Neumann J, Richtmyer RD. 1950. A method for the calculation of hydrody-
namical shocks. J. Appl. Phys. 21:232–237
Wang ZJ. 2007. High-order methods for the Euler and Navier-Stokes equations
on unstructured grids. Progr. Aero. Sci. 43:1–41
Weirs VG, Candler GV. 1997. Optimization of weighted ENO schemes for DNS
of compressible turbulence. AIAA Paper 97-1940, AIAA
Wu M, Martı́n M. 2007. Direct numerical simulation of supersonic turbulent
boundary layer over a compression ramp. AIAA J. 45:879–889
Xu S, Martı́n MP. 2004. Assessment of inflow boundary conditions for compress-
ible turbulent boundary layers. Phys. Fluids 16:2623–2639
Yee HC. 1989. A class of high-resolution explicit and implicit shock-capturing
methods. Technical Memorandum 101088, NASA
Yee HC, Sandham ND, Djomehri MJ. 1999. Low-dissipative high-order shock-
capturing methods using characteristic-based filters. J. Comput. Phys.
150:199–238
Yee HC, Sjögreen B. 2007. Development of low dissipative high order fil-
ter schemes for multiscale Navier-Stokes/MHD systems. J. Comput. Phys.
225:910–934
Zhou J, Adrian RJ, Balachandar S, Kendall TM. 1999. Mechanisms for generating
coherent packets of hairpin vortices in channel flow. J. Fluid Mech. 387:353–396
Zingg DW, Lomax H, Jurgens HM. 1993. Optimized finite-difference schemes for
wave propagation. AIAA Paper 93-0459, AIAA
 32 Pirozzoli
2 2

ρ′ /ρ0 /Mt 20
1.5 1.5

1 1

0.5 0.5

0 0
0 10 20 30 40 0 20 40 60 80

2 2

1.5 1.5
K/K0

1 1

0.5 0.5

0 0
0 10 20 30 40 0 20 40 60 80

0.2 0.2
S/cp /Mt 20

0 0

-0.2 -0.2

-0.4 -0.4
0 10 20 30 40 0 20 40 60 80

t/τ t/τ
D2 D6 D7 W7 S6 K6 H6 T2

Figure 1: Simulations of isotropic turbulence at zero viscosity at Mt0 = 0.07 (left

column) and Mt0 = 0.3 (right column): time evolution of r.m.s. density fluctua-
tions, total kinetic energy and total entropy (τ is the eddy turnover time). Dn: n-
th order discretization of divergence form; W7: seventh-order WENO scheme; S6:
sixth-order discretization of skew-symmetric split form (19); K6: sixth-order dis-
cretization of skew-symmetric split form (20); H6: sixth-order entropy-consistent
scheme of Honein & Moin (2004); T2: second-order entropy-conserving scheme
of Tadmor (1984) (β = 4). Interrupted lines indicate numerical divergence.
S3
S2
S1
S0
h

j−2 j−1 j j+1 j+2 j+3

j + 1/2
S
Figure 2: Definition of stencils for WENO reconstruction (L = 3).
Numerical methods for high-speed flows 33

Figure 3: Numerical simulation of test case 13 of Lax & Liu (1998) on a mesh
with hx = hy = 1/1200 (only the central part of the domain is shown). Time
integration is performed with a third-order SSP Runge-Kutta algorithm (the
Courant number was set to 0.6). Fourteen equally spaced density contours are
shown, from 0.6 to 2.3. (a) second-order TVD scheme with Van Leer limiter; (b)
fifth-order WENO scheme; (c) seventh-order WENO scheme; (d) hybrid com-
pact/WENO scheme (Pirozzoli, 2002). In the latter case, WENO is activated at
the intermediate node xj+1/2 (and two neighboring nodes left and right of it),
whenever |ρj+1 − ρj | ≥ 0.2.
34 Pirozzoli

25 25

20
(a) 20
(b)

15 15
y/δ0

10 10

5 5

0 0
15 20 25 30 35 40 15 20 25 30 35 40
25 25

20
(c) 20
(d)

15 15
y/δ0

10 10

5 5

0 0
15 20 25 30 35 40 15 20 25 30 35 40

x/δ0 x/δ0
Figure 4: Performance of shock sensors for DNS of transonic shock/boundary
layer interaction. Black lines denote pressure iso-lines (24 equally spaced levels,
1 ≤ p/p∞ ≤ 1.85). Red lines indicate iso-lines of shock sensors. (a) Ducros
sensor (Ducros et al., 1999); (b) Jameson sensor (Jameson et al., 1981); (c) Hill
& Pullin sensor (Hill & Pullin, 2004); (d) Visbal & Gaitonde sensor (Visbal &
Gaitonde, 2005). δ0 denotes the incoming boundary layer thickness.
Numerical methods for high-speed flows 35

12

10
y(mm)

240 260 280 300 320 340 360 380 400

12

10
y(mm)

240 260 280 300 320 340 360 380 400

x(mm)
Figure 5: LES of shock/boundary layer interaction (Touber & Sandham, 2009):
comparison with PIV experimental data (Dupont et al., 2008). Top: mean
streamwise velocity; bottom: Reynolds shear stress. The same contour levels
are used for PIV (filled contours) and LES (solid lines). The yellow/red lines
denote the mean separation line as from PIV (dashed) and LES (solid). Figure
courtesy of E. Touber and N. D. Sandham.
36 Pirozzoli

Figure 6: DNS of transonic shock/boundary layer interaction (Pirozzoli et al.,

2010). The shock system (in gray shades) is identified as an iso-surface of the
modified Ducros sensor (Θ = 0.0025). Vortical structures are educed as iso-
surfaces of the swirling strength (λci = 1.6 u∞ /δ0 ), and are coloured with the
local Mach number (levels from 0 to 1.3, colour scale from blue to violet). Axes
are scaled with respect to the incoming boundary layer thickness (δ0 ).
Numerical methods for high-speed flows 37

Figure 7: LES of jet in supersonic cross-stream (Kawai & Lele, 2010). Visualiza-
tions in streamwise/wall-normal plane are shown in the top row: density gradient
(left), and jet fluid (right). Visualizations in wall-parallel plane are shown in the
bottom row: streamwise velocity (left), and jet fluid (right). D is the jet diameter.
Figure courtesy of S. Kawai and S.K. Lele.