Professional Documents
Culture Documents
The Carbuncle Phenomenon Is Incurable
The Carbuncle Phenomenon Is Incurable
http://actams.wipm.ac.cn
Volker Elling
Department of Mathematics, University of Michigan, MI4810-1043, USA
E-mail: velling@umich.edu
1 Introduction
as a sufficiently fine grid. Figure 1 displays two characteristic situations: the upper group of
carbuncles is located near a boundary where numerical boundary layers act; the lower group
appears in a region where the shock is almost but not exactly parallel to the grid edges. Some
authors single out the Roe scheme [21] as particularly prone to carbuncles whereas others report
bad experiences with the Godunov method [11].
As a consequence of the general confusion, there are no safe recipes for eliminating carbun-
cles. Numerical viscosity in various forms can suppress the carbuncle phenomenon, for example
by smoothing tangential to the shock, but preserving the sharp shock in the normal direction.
Since numerical viscosity smears discontinuities and decreases overall accuracy, it is desirable to
limit the amount necessary, which requires precise theoretical criteria for carbuncle avoidance.
Attempts to fix numerical schemes, especially approximate Riemann solvers, by other means
have been inconclusive. This is due to the volatile nature of carbuncles. They can be hard to
produce in one set of circumstances, but hard to suppress in another, so that absence in a small
number of test cases can lead to erroneous claims of a cure. As [13] observes: “Many have
proposed cures for the carbuncle problem, but none are universally accepted, and most papers
begin by criticizing previous work.”
With shock tracking on the other hand, suppressing carbuncles is easy: the shock is explic-
itly approximated by a smooth curve. (As we will see, this may not be desirable in cases where a
carbuncle should be present.) Although shock tracking schemes have the additional advantage
of higher accuracy even in presence of shocks [10], most practitioners prefer shock capturing
schemes, which use a fixed grid without attempts to adapt it to the present discontinuities,
because of lower complexity of implementation and claims of faster calculations.
In this article we devise a new technique to trigger carbuncles reliably and to compute their
structure in detail. All numerical results were obtained for the 2D isentropic Euler equations:
ρt + ∇ · (ρv) = 0,
p(ρ) = ργ pressure (γ = 1.4 used for the calculations). The non-isentropic (full) Euler equations
are obtained by adding an energy equation
where E = e + 12 |v|2 , e internal energy per mass, and by replacing the preessure law by p =
p(ρ, e) = (γ − 1)ρe. For full Euler we add an entropy inequality
(ρs)t + ∇ · (ρsv) ≥ 0,
where s = log(p/ργ ) is (physical) entropy per mass; for isentropic Euler we instead add an
energy inequality
(ρE)t + ∇ · ((ρE + p)v) ≤ 0.
A weak solution satisfies the equations of each system in the distributional sense; an entropy
solution satisfies the additional inequality as well.
No.6 V. Elling: THE CARBUNCLE PHENOMENON IS INCURABLE 1649
Results for full Euler and for other γ, which we do not display here, differ only in aspects
that are unimportant for our discussion. In all diagrams, flow into the domains is Mach 3.0 hor-
izontally from left to right. Unless otherwise noted, all pictures show the horizontal component
of velocity, with values increasing from blue over green and yellow to red.
In this section we describe how carbuncles can be triggered in a large variety of numerical
schemes in standard coordinates (here, “standard” means coordinates (t, x) as compared to
similarity coordinates (t, xt ).) We start with a steady vertical plane shock wave on a uniform
Cartesian 2d grid. The upper and lower domain boundaries are solid walls with slip boundary
condition; fluxes across the left resp. right domain boundary are computed with pre- resp.
post-shock state in “pseudo-cells” on the outside of the domain boundaries. In this setting, the
exact plane shock wave is reproduced faithfully by discrete shocks.
Now the initial and boundary data is modified as follows: we choose an edge in the center
of the left boundary; in its pseudo-cell and in all cells in a horizontal one-cell-high “filament”
from this edge to the shock, the horizontal velocity is set to zero [7]. This filament immediately
triggers a structure (see Figure 3) that grows at a constant rate (and eventually reaches the
domain boundaries). This structure, somewhat reminiscent of a stingray, is very similar to the
small discrete carbuncles shown in Figure 2, except for the filament. The latter carbuncles are
triggered by some unknown discrete mechanism which is probably different for each numerical
scheme and perhaps even Mach number and shock-grid angle. The larger carbuncle produced
by the original blunt-body calculation in [17] is even more clearly similar to Figure 3.
Figure 3 has been computed with the Godunov scheme [11]; similar carbuncles can be
observed in the local Lax-Friedrichs [22] or the Osher-Solomon [16] schemes. Higher-order
1650 ACTA MATHEMATICA SCIENTIA Vol.29 Ser.B
schemes are likewise affected, which is not particularly surprising because the creation of the
carbuncle takes place in a nonsmooth region where higher-order schemes revert to first order.
Moreover, adding generous amounts of numerical viscosity anywhere except near the filament
and the carbuncle tip does not destroy the carbuncle. This suggests that the tip is the unphysical
part of the carbuncle — except in cases where carbuncles are indeed physically meaningful (see
below).
Fig.3 A stingray-like structure resembling carbuncles. The filament is one grid cell tall while the
carbuncle grows at a constant rate
The reader may suspect that a pronounced disturbance like the carbuncle in Figure 3 is
inevitable given the reduced momentum in the filament. However, the filament is too thin to
support this reasoning. When the grid is refined and the filament height (and hence the L1
norm perturbation to the initial and boundary data) decreases, the carbuncle does not shrink.
In particular the carbuncle grows at a constant rate (until it starts to interact with the domain
boundaries) while the filament has a fixed size. The filament is an arbitrarily small perturbation
that triggers a large-scale effect due to an intrinsic instability.
The new method for triggering carbuncles is most effective in standard coordinates where
it can be successfully applied to many numerical methods. It is observed that the carbuncle pat-
tern grows at a constant rate and that its inner structure settles when the carbuncle has grown
to occupy many grid cells. Hence carbuncles are most naturally represented and computed as
self-similar flow patterns. Moreover, to avoid interaction with numerical domain boundaries,
for detailed structure calculations similarity coordinates (t, x/t) are more appropriate.
A major drawback is that similarity coordinates add dissipation because grid cells move and
smoothen stationary features by averaging. For most numerical schemes tested, this decreases
carbuncles to a point where it cannot be said for sure whether the carbuncle is uniformly large
as the filament height is reduced. More importantly, it is not possible to create carbuncles of
sufficient size to compute significant detail. However, the Godunov scheme does produce robust
and large carbuncles in the self-similar setting.
Here we compute on a Cartesian grid with adaptive refinement to achieve better resolution.
Similarity coordinates are treated by the moving-edge modifications discussed in [6]. To save
No.6 V. Elling: THE CARBUNCLE PHENOMENON IS INCURABLE 1651
time, only the upper halfplane of the (obviously symmetric) carbuncles is computed. The
trigger mechanism remains the same: in the lowest vertical edge at the left domain boundary,
the flux of horizontal momentum is reduced to 0. In similarity coordinates the resulting contact
discontinuity is not preserved exactly; additional waves propagate upwards and the filament
weakens noticeably (see Figure 4). However, this unelegant sideeffect is small enough for our
purposes.
Fig.4 Filament emanating from the lowest edge in the left boundary
Fig.5 Filament and carbuncle tip (contact smeared due to poor grid alignment)
In this new setting, carbuncles can be computed in great detail (see Figure 6 for overview
and Figures 7 and 8 for more detail). Better resolution could be achieved by using higher-order
modifications, but here we choose the Godunov scheme because of its theoretical appeal (it
manifestly satisfies all discrete entropy inequalities).
However, it is likely that the carbuncle structure varies depending on upstream Mach
number, γ, and other parameters. Indeed we now argue that the essential common feature is
the tip area.
In [8] a numerical example is presented that suggests that entropy solutions of the Cauchy
problem for the compressible Euler equations are not always unique. More precisely, Figure 9
constitutes an exact steady and self-similar entropy solution; however, using it as initial data,
numerical schemes produce another solution, Figure 10, which appears to be self-similar and
(being computed by the Godunov scheme) an entropy solution, but clearly unsteady. All other
numerical scheme that were tested yield Figure 10 as well.
No.6 V. Elling: THE CARBUNCLE PHENOMENON IS INCURABLE 1653
Fig.10 A second self-similar (but unsteady) entropy solution, for Figure 9 as initial data
The initial data in this example was motivated by research on the problem of supersonic
flow onto a solid wedge [9]. For thin wedges (small α in Figure 9), there are two possible shocks
attached to the tip, a stronger and a weaker one (the latter was chosen for computing Figure
10). Replacing the solid wedge by gas with post-shock density and zero velocity yields Figure
9.
In Figure 5 the tip of the carbuncle in Figure 6 is shown in detail. Clearly it resembles
Figure 9 (the post-shock area in Figure 5 is supersonic which means the shock corresponds
to the weak shock in the wedge problem). This analogy suggests that carbuncles are another
instance of the same problem: two different entropy solutions for identical initial data. In this
case, the theoretical solution (the plane shock wave) is known to be physically correct as well as
a vanishing viscosity and hence entropy solution; the carbuncle appears to be a second entropy
solution which is probably unphysical in absence of the filament.
This paper is not the first to conjecture that carbuncles are phenomena of the continuum
equations rather than purely numerical artefacts. Other references are [20] where a linear
stability analysis of plane shocks in the exact Euler equations seems to reveal a new unstable
1654 ACTA MATHEMATICA SCIENTIA Vol.29 Ser.B
mode related to carbuncles ([2] have raised objections to this analysis), as well as [15] and [5],
both of which explicitly state suspicions that carbuncles and other anomalies have a physical
background.
It should be emphasized that, although the carbuncle in Figure 6 does not appear to vanish
as the grid is refined, it shrinks and grows and does not seem to converge to a single pattern
either. Indeed, if the comparison to the nonuniqueness example is appropriate, then there may
be a parametrized family of carbuncle patterns rather than a single one. There are at least
two parameters that are currently not under our control: the tip angle β and the location of
the carbuncle tip. These are determined in some unknown and probably rather sensitive way
by the numerical method, the filament and the inner structure of the carbuncle pattern. If
one or both parameters are free (i.e., if there is a different carbuncle pattern for each choice of
parameter in some interval), the carbuncle calculation should not be expected to converge to a
single function as the grid is refined.
More generally, we should stress that even on a single fixed grid Figure 6 is rather un-
stable; small modifications to the numerical scheme can have drastic effects. For example, the
Godunov scheme requires solving a nonlinear equation for every edge in every time step; solv-
ing this equation with less than machine accuracy can either trigger oscillations that abort the
computation or produce additional smaller carbuncles along the shock that grow to spoil the
overall pattern.
It is possible to stabilize the calculation to some degree by adding numerical viscosity every-
where except near the lower boundary; this suppresses the appearance of additional carbuncles,
but smears the carbuncle pattern even more.
5 Vanishing Viscosity
The Euler equations arise from the vanishing viscosity limit: the limit as ↓ 0 when various
parabolic terms of the form ∂i (B(U )∂i U ) (where U = (ρ, ρv, ρE)) are added to the right-hand
side of the system. Examples are uniform viscosity (B = I), Navier-Stokes viscosity, or Navier-
Stokes-Fourier viscosity which has several small parameters, two viscosity coefficients and a
heat conduction coefficient, that may go to zero at different rates, depending on the situation.
Any physically meaningful viscosity term has to preserve the entropy inequality.
Naturally each choice of viscosity, topology and relative convergence rates may yield a
different Euler solution; moreover the same limit may have two converging subsequences that
yield different weak solutions. However, in each case,
1. either the weak solution is the same for all “reasonable” limits,
2. or the choice of limit matters, so that the Euler equations, as well as “pure Euler”
numerical schemes, cannot be expected to be useful on their own.
The filament in Figure 4 has similarities with of viscous boundary layers. Instability of
boundary layers is of course well-known in fluid mechanics [1]. In the presence of shock waves,
so-called lambda shocks occur [12]. On inspection it is obvious that the local structure of
lambda shocks and the carbuncles computed in this article is essentiall the same (although the
downstream features may differ in each setting).
At a boundary, lambda shocks/carbuncles are unavoidable because they are physical phe-
No.6 V. Elling: THE CARBUNCLE PHENOMENON IS INCURABLE 1655
nomena. It is well-known that a pure Euler scheme is inherently unable to compute the correct
flow because it ignores the viscous scales that decide separation or attachment. In fact, regard-
less of the virtues of the actual numerical scheme, the mere step of approximating the initial
data on a grid that is coarser than the viscous scale already destroys information needed to
determine the unique flow forward in time. For example turbulence in the boundary layer can
cause the flow to stay attached to the boundary longer, which has a large-scale effect.
Any attempt to “cure” carbuncles in flow away from boundaries would assume that the
upstream flow is smooth, free of filaments and other disturbances. This need not be the case;
for example the wake of any object farther upstream contains a street of vorticity resembling
our filament. Indeed, [14, 19] show physical results of experiments that use wall corners to
generate a vortex sheet that creates carbuncle-like features when impinging on a shock.
Therefore we have to conclude that
in the sense that a physical assumption, namely smoothness of upstream flow, is needed to
justify the deliberate suppression of carbuncles in numerical schemes. This is unsatisfactory for
two reasons: first, it is far from obvious that this assumption alone is sufficient. Downstream
phenomena may influence the formation of carbuncles as well. Second, the assumption may be
violated in many practical applications. Then a scheme has to resolve any upstream disturbance
at viscous scales; schemes that solve the Euler equations without modelling the viscous scales
can be useless or at least unreliable.
If there is a large number of disturbances, viscous scales may have to be resolved in any
part of the domain. (Figuratively speaking, we may have to model every butterfly to predict
the presence of a hurricane.) This, of course, is precisely what the scientific community hopes
to avoid. In many applications the viscous scales are too small to be resolved at a reasonable
cost.
As [4, p. 553] observe: “For an aircraft, subject to laminar flow over its wing section, the
thickness of the boundary layer is seldom more than 1 mm.” The large scales in reasonable
models are at least on the order of 10 m. A grid resolution of 1 mm in the entire domain would
require 108 grid cells merely for 2d calculations, 1012 for 3d. Worse, there is a time dimension
as well: the CFL condition imposes a restriction on the numerical time-step proportional to the
diameter of grid cells. (However, for calculations of stationary flow, iteration methods that are
not time-accurate can exceed the CFL limit to some degree.)
The only hope to avoid this cost is a new subgrid model, similar to the common turbulence
models, but taking laminar instabilities into account as well. The author would surprised,
however, if a moderate number of additional fields can represent all viscous-scale features that
can make a large-scale difference.
In any case, much of the research on numerics for the pure1) Euler equations appears to
rest on rather shaky theoretical and physical foundations. Numerical schemes for conservation
laws may well turn out to find their true calling in potential flow [9] or propagation of waves in
nonlinear elastic solids [3].
1) without any form of viscosity modelling
1656 ACTA MATHEMATICA SCIENTIA Vol.29 Ser.B
References