See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/333808016

Large Eddy Simulation of dense gas flow around a turbine cascade

Conference Paper · June 2019

DOI: 10.2514/6.2019-2843

CITATIONS READS
5 274

3 authors:

Jean-Christophe Hoarau Paola Cinnella

The French Aerospace Lab ONERA Sorbonne Université
6 PUBLICATIONS 25 CITATIONS 271 PUBLICATIONS 2,780 CITATIONS

SEE PROFILE SEE PROFILE

Xavier Gloerfelt
Ecole Nationale Supérieure d'Arts et Métiers
186 PUBLICATIONS 1,790 CITATIONS

SEE PROFILE

All content following this page was uploaded by Paola Cinnella on 20 June 2019.

The user has requested enhancement of the downloaded file.

 Large Eddy Simulation of dense gas flow
around a turbine cascade

Jean-Christophe Hoarau∗ , Paola Cinnella † and Xavier Gloerfelt ‡

DynFluid laboratory, 151 boulevard de l’hôpital, 75013 Paris, France

The flow in a transonic turbine cascade using a dense gas (a molecularly complex organic
compound) as the working fluid are investigated by means of large eddy simulations (LES).
The Martin and Hou real gas equation of state and the Chung–Lee–Starling models are used
to describe the thermodynamic behavior and the variations of transport properties with tem-
perature and density for the dense gas. A high-order spatial scheme is selected and an efficient
implicit time integration scheme is employed to relax time-step constraints arising in the rel-
atively high Reynolds number flows of interest. The numerical methodology is first validated
against experimental and numerical data from the literature for the well-documented VKI
LS89 turbine cascade, which uses a perfect gas (air) as the working fluid. Subsequently, flows
through the same cascade of a dense gas are carried out for various choices of the operating
conditions (pressure ratio and inlet thermodynamic conditions). The influence of dense gas
effects on shock wave formation and loss mechanisms are highlighted by means of the large
eddy simulations.

I. Introduction
Dense gases, characterized by complex molecules and moderate to large molecular weights, exhibit a drastically
different behavior compared to perfect gas over a range thermodynamic conditions above the saturation curve, and
precisely for pressures and temperatures of the same magnitude as the liquid/vapor critical point. The non linearity of
dense gas behavior is governed by the fundamental derivative of gas dynamics, defined as

ρ ∂c
Γ =1+ (1)
c ∂ρ s

where ρ is the density, c = ∂p/∂ ρ|s the speed of sound, s the entropy and p the pressure. Γ represents the rate of
p

change of the sound speed in isentropic transformations. According to the definition (1), Γ < 1 implies ∂c/∂ ρ|s < 0,
meaning that the sound speed grows in isentropic expansions while drops in isentropic compression. Conversely the
fundamental derivative of gas dynamics in the perfect gas case can be written as
γ+1
Γ= (2)
2
where γ = c p /cv is the ratio of specific heats at constant pressure and volume, respectively. In perfect gas case γ is
greater than the unity, hence Γ > 1. For organic molecules of moderated to high complexity, Γ becomes less than the
unity, or even negative, in the vapor region, at pressures and temperatures of the general order of magnitude of the
liquid/vapor critical point. In these regions, dense gases exhibit strong deviations from the perfect gas behavior and
possibly non-classical phenomena (see, e.g.,[1] and references cited therein).
Dense gases are of interest for several engineering applications, the most attractive one being represented by Organic
Rankine Cycle (ORC) systems. These cycles use lower maximal temperatures (less than 300◦ C) than classical gas
cycles or steam cycles and use organic working fluids instead of classical ones, such as steam or air. The low value of
the maximal temperature allows to recover energy from wasted industrial heat sources or even renewable ones, such
as geothermal or solar sources. A key enable for the design of efficient ORC systems is the availability of accurate
models and simulation tools. In the absence of accurate experimental data for this kind of flows, wall-resolved Large
Eddy Simulation (LES) constitutes an interesting way to assess Reynolds Averaged Navier–Stokes (RANS) solvers,
∗ Ph.D. student, Ecole Nationale des Arts et Métiers, [email protected]
† Professor, Ecole Nationale des Arts et Métiers, [email protected]
‡ Professor, Ecole Nationale des Arts et Métiers, [email protected]

1
which is still the standard tool for ORC turbine design. High-fidelity simulations can also be helpful to gain a better
understanding of physical mechanisms generating losses, such as laminar-to-turbulent transition, shock wave formation
and interaction with boundary layers or wake formation and mixing.
The paper is organized as follows. The numerical methodology and governing equations are described in Section
II. Section III presents the chosen flow configuration and preliminary 2D dense gas calculations used to set up the
upcoming LES dense gas simulations. The analyses of LES results are discussed in Section IV and first conclusions are
drawn in Section V.

II. Numerical Methodology

A. Governing Equations
Dense-gas flows are governed by the single-phase Navier–Stokes equation written in integral form as:
Z I
d
w dΩ + φ.n dΓ = 0 (3)
dt Ω ∂Ω

where t is the time, w = ρ, ρ U, ρ V, ρ W, ρ E T is the vector of conservative variables (with ρ the fluid density, U,V,W
 
the Cartesian components of the velocity vector and E the specific total energy), Ω is a closed control volume with
boundary ∂Ω, φ is the physical flux density and n is the unit outward normal. The flux density contains the contributions
of both the convective and viscous fluxes, namely φ = φ c − φ v . φ c and φ v are smooth functions of the variables w, and
of the variables and their spatial gradient (w, ∇w), respectively.
The viscous fluxes may contain the contribution of the subgrid stresses for LES calculations. Note that, in practice,
present LES are conducted without the introduction of an explicit model for the subgrid terms. Instead, the damping of
unresolved subgrid scales is ensured implicitly by the numerical dissipation term associated with the spatial scheme,
described later. The spatial resolution of the computational grid close to the wall is fine enough to resolve the viscous
sublayer, so that no wall functions are used.

B. Thermodynamic models
Equation (3) has to be supplemented by a suitable thermodynamic model (equation of state) and by models for
transport properties. The thermodynamic model comprises a thermal and a caloric equation of state, written for instance
under the form:

p = p( ρ,T ), e = e( ρ,T ) (4)

where p is the static pressure, ρ the density, e the internal energy and T the temperature. For perfect gas cases, the gas
is considered as thermally and calorically perfect, and equation (4) reads:

p = ρRT, e = cv T (5)
where cv = R/(γ − 1) with γ = 1.4 for perfect gas, and R = R/M with R = 8.314 J−1 K−1 mol−1 , the universal gas
constant, and M the molecular weight. Concerning the transport properties, the classical Sutherland law is used to
model the dependency of the dynamic viscosity on the temperature. A constant Prandtl number hypothesis is adopted to
determine the thermal conductivity.
For modeling dense gases, we consider the equation of state proposed by Martin and Hou [2], considered as a
reasonably accurate model for fluorocarbons:

RT X Ai + Bi T + Ci e−kT /Tc
5
p= + (6)
v − b i=2 (v − b) i

where b and k are constants, Tc the critical temperature of the gas and Ai , Bi ,Ci are experimentally calibrated coefficients.
Then the compatibility formula between the caloric and thermal equation of states supplemented by a power law is used
to determine the caloric equation of state. Finally, the transport properties are represented by means of the model by
Chung et al. [3]. For a dense gas, these properties depend both on the fluid temperature and density.

2
C. Spatial integration scheme
The spatial terms in the governing equations are discretized using a structured finite volume scheme, previously
described in [4]. The convective terms are approximated by means of a nominally fourth-order centred scheme,
supplemented with a non-linear artificial viscosity term of third-order accuracy in smooth flow regions. The latter allows
to damp grid-to-grid oscillations and prevents spurious oscillations close to flow discontinuities. A classical second-
order approximation is adopted for the viscous terms. This scheme is referred to as DNC3 (third-order Directional Non
Compact) in the following of this paper.
The approximation of the inviscid terms is briefly recalled hereafter in the simple case of a 1D system of conservation
laws:
∂w ∂ f (w)
+ =0 (7)
∂t ∂x
where w is the vector of conservative variables, f is the flux function and t and x are the time and space variable. Given
a uniform grid with space step h, so that x j = j h the semi-discrete scheme in space writes

(δF ) j
(wt ) j + =0 (8)
h
where δ is the classical difference operator over one cell, δ(•) j+ 1 := (•) j+1 − (•) j , and Fj+ 1 is the numerical flux at
2 2
cell interface j + 21 :
" ! #
1 2
Fj+ 1 = I − δ µf − D (9)
2 6 j+ 1 2
 
where f is the physical flux, µ is the cell average operator, µ(•) j+ 1 := 1
2 (•) j+1 + (•) j , and D is the artificial
2
dissipation term. The latter writes: f g
D j+ 1 = ε 2 δw + ε 4 δ3 w 1 (10)
2 j+ 2

with ε 4 j+ 1 = max(0, k4 − ε 2 ,j+ 1 ) and ε 2 ,j+ 1 = k 2 λ j+ 1 max(ν j , ν j+1 ) where k2 ∈ [0, 1], k4 ≈ 12
1
are the dissipation
2 2 2 2
∂f
coefficients, λ is the spectral radius of the flux Jacobian , and ν j is the well-known pressure-based shock sensor of
∂w
Jameson et al. [5].

D. Time integration scheme

Mesh clustering near solid walls leads to strong restrictions of the maximum allowable time step to ensure the
stability when an explicit time integration scheme is used. This restriction is often smaller than the smallest resolved
time scale in the simulation. To overcome this restriction, an implicit scheme is required. However, fully implicit
methods are prohibitively expensive in computation time, due to the inversion of a large matrix at each time step.
Additionally, they usually require sub-iterations to relax numerical errors introduced by the approximations used in the
inversion procedure. Furthermore, they are generally limited to second-order accuracy at best.
In this work, we use an implicit method recently proposed in [6], allowing a moderate increase of the CFL number
(by a factor 2 to 5) with respect to standard explicit schemes, while implying a small overcost. This scheme belongs to
the family of Implicit Residual Smoothing (IRS) methods, consisting in filtering high frequency modes in the residual
when running an explicit scheme with a time step larger than the one imposed by the stability limit. They have been
extensively used in the literature at first or second order [5, 7, 8] in order to accelerate the convergence of steady state
calculations. The present version is a fourth-order extension, referred to as IRS4, which is well-suited for DNS and LES
simulations, since it introduces a low additional error and requires a small computational overcost for matrix inversion.
It was succesfully appplied to perfect gas flows in turbomachinery [9].
Specifically, the IRS4 method is applied to an optimized six-step Runge–Kutta scheme (RK6), originally proposed
by Bogey and Bailly [10]. The latter is second-order accurate in the sense of the truncation error, but its coefficients are
optimized in the Fourier space to minimize numerical errors.
Let us consider the following conservation law, discretized in space by means of some approximation technique:

wt + R (w) = 0 (11)

3
where R is the space approximation operator (in our case, the DNC3 scheme described in the preceding Section) and
wt is the time derivative of the solution. An r-stage Runge–Kutta time stepping scheme is written as:



 w (0) = wn
 ∆w (k ) = ∆tR (w (k−1) ), k = 1, ...r

 −ak (12)
 w n+1 = w (s)



where w n is the numerical solution at time n∆t, ∆w (k ) = w (k ) − w (0) is the solution increment at the k th RK stage, r
is the number of stages, and ak are the scheme coefficients. A larger stability domain is obtained by smoothing the
residual at each stage with the IRS operator J , introduced in the left-hand side of (12):



 w (0) = wn
 J ∆w (k ) = −ak ∆tR (w (k−1) ), k = 1, ...r

 (13)
 w n+1 = w (s)



For a one-dimensional problem, the implicit operator reads

J = 1 + θ( δ∆tx ) 4 δ(λ e 4 δ3 ) (14)

where θ is the smoothing coefficient, δ the differential operator, λ e the spectral radius of inviscid fluxes and ∆t the time
step. This corresponds to a damping term base on fourth derivatives applied to the solution residual. The additional
error introduced by J at each RK stage is of the order of ∆t 4 :

1 ∂f e
− θ∆t 4 λ 4 d5 + o(∆t 4 ) (15)
12 ∂x d

As a consequence, the leading truncation error term of the time-stepping scheme is not altered by the IRS procedure.
IRS4 has been shown to be unconditionally stable for θ ' 0.005. For stability constraint in the following LES
calculations, the smoothing parameter is fixed at a higher value, ie. θ = 0.01 unless otherwise stated. Additionally, the
accuracy of the IRS4-RK6 scheme remains close to that of the explicit scheme for CF L numbers up to about 10 [9].
At each RK stage, the IRS4 requires the inversion of a scalar pentadiagonal system per direction, which corresponds
to the solution of two embedded Thomas algorithms. Thanks to the availability of a highly efficient pentadiational solver,
the overcost associated with the application of IRS4 with respect to the explicit RK6 is rather small, and computational
gains of a factor 3 to 5 have be obtained thanks to the increased time step for realistic 3D calculations.

III. Flow configurations

A. Geometry and numerical setup

Simulations were performed for the well known VKI LS-89 planar cascade instrumented by Arts et al. in 1990 [11],
which originally uses air as the working fluid. This case was chosen because it has been largely documented in the
literature, using both experiments and numerical simulations. This configuration is not an optimised dense gas shape as
it can be found in the previous dense gas studies [12, 13]. The LS-89 geometry was chosen to assess our numerical
strategy and then to put results for a dense gas in perspective with those obtained for a perfect gas.
More precisely, the blade chord C is 67.647 mm long with a pitch-to-chord ratio of 0.85 and a stagger angle of
χ = 55◦ . The flow angle at turbine inlet is equal to 0◦ . The computational grid is depicted in Figure 1, where plane
sections consist in a H-type mesh with 180 points in the pitchwise direction and 850 points in the streamwise direction
(Figure 1b). For 3D simulations, the mesh is simply extruded in the spanwise direction over 200 points, yielding 30.6
×106 grid points. The blade is discretized by 550 points on the upper and lower surfaces. To make a fair comparison
between perfect and dense gas, the same pressure ratio and the same outlet Reynolds number have been kept for dense
gas cases. The pressure ratio is indeed the main parameter of a turbine cascade and using the same Reynolds number
ensures a similar resolution. To maintain the same exit Reynolds number, the geometry has been rescaled (by a factor
20 in this case) compared to the original VKI-LS89 geometry, resulting in a chord of 3.38 mm.
In the perfect gas simulations, the working fluid is air and the physical parameters corresponding to the case referred
to as MUR129 in experiments are selected (see Table 1). An isothermal wall condition is used with Tw =297.75 K.

4
For dense gas simulations, the working fluid is the perfluoro-perhydrophenanthrene (chemical formula C14 F24 , called
hereafter with its commercial name PP11), a heavy fluorocarbon gas. The PP11 is a BZT fluid that has been previously
studied in the dense gas literature [14, 15] and which exhibits a relatively wide inversion zone and can thus potentially
induce BZT effects. The Martin-Hou [2] equation of state is used supplemented by the transport properties of Chung et
al. Since there is no guiding experimental configuration for dense gases, the inlet working point and the wall temperature
are additional parameters to determine. To this aim, a parametric study based on 2D simulations has been realized.

Re2 Mi s,2 p1◦ (Pa) p1◦ /p2 Tu

MUR129 1.13 × 106 0.84 1.87 × 105 1.58 0%

Table 1 Free-stream conditions of MUR129 LS-89 configuration. Subscript 1 refers to the inlet section and
subscript 2 to the outlet section. Re denotes the Reynolds number based on the chord, Mi s is the isentropic
Mach number, p◦ the total pressure, p the static pressure and Tu the incoming turbulence rate.

0.10 C

0.85 C

0.7 C

1.5 C

(a) (b)

Fig. 1 Computational domain of LS-89 cascade (a) and close-up view of the grid (b), where one point out of 4
is represented.

B. Choice of inlet condition and wall temperature for the dense gas study
The following section describes preliminary 2D simulations used to select the inlet working point, that is to say
the total inlet pressure and temperature, and the temperature for the isothermal wall condition. As explained in the
introduction, significant deviations from the ideal gas behavior appear when the fundamental derivative of gas dynamics
Γ is smaller than one or even negative. The region where Γ is smaller than unity is called the dense gas region, whereas
the region where Γ < 0 (present only in the so-called Bethe–Zel’dovich–Thompson, BZT, fluids) is known as the
inversion zone. This regions are located in the vapour phase near the saturation curve in thermodynamic conditions
close to the critical point. In the inversion region non-classical shock waves may appear. By adjusting the inlet working
conditions, it is possible to impose an inversion zone in the region where the shock system is expected.

To make a preliminary choice of the thermodynamic conditions, the flow path is assumed, to a first approximation,
to remain nearly isentropic throughout the cascade. A first inlet point, referred to as IC1 (see Figure 2 and Table 2),
has been considered. These conditions were used in Sciacovelli et al. [14] and correspond roughly to starting from
the middle of the Γ < 0 region. The range of thermodynamic states along a streamline through the blade passage are
depicted in a Clapeyron’s diagram in Figure 2 with the white line starting from IC1. The expansion process begins
in the BZT region but Γ values rapidly increase well above zero, leading to the formation of classical shock waves at

5
the blade trailing edge, as depicted in Figure 3a. The inlet condition can be shifted to the left on the same isentropic
line to impose BZT conditions near the trailing edge of the turbine (where the shock system is expected to appear).
This new inlet point is denoted IC2 (see Figure 2 and Table 2). The corresponding map of Γ obtained with these IC2
conditions is shown in Figure 3b. It corresponds to a very high value of Γ near the inlet, which rapidly diminishes due
to the expansion across the vane, leading to values close or less than zero in the vicinity of the trailing edge. Note that
this point corresponds to a very high pressure at the entrance, which is difficult to realize in practice.

p1◦ /pc ρ◦1 /ρc T1◦ /Tc Γ1 p1◦ /p2 Re2

IC1 0.98 0.62 1.001 -0.093 1.58 1.13 × 106
IC2 1.35 1.47 1.019 6.706 1.58 1.13 × 106

Table 2 Inlet thermodynamic conditions for dense gas simulations.p1◦ , ρ◦1 and T1◦ are the total inlet pressure,
density and temperature respectively, pc , ρc and Tc the critical pressure, density and temperature, Γ1 the inlet
fundamental derivative of gas dynamic, p2 the outlet static pressure and Re2 the outlet Reynolds number.

2 T/Tc
Γ=3 2
1.8 Γ=1 1.9

1.6
s=s0 1.8
1.7
1.6
1.4 IC2
p/pc

1.5
1.2 1.4
1.3
IC1 Γ = 0 1.2
1
1.1
0.8 1
0.9
0.6 0.8
1 1.5 2 2.5 3 3.5 4
v/vc
Fig. 2 Clapeyron’s diagram of PP11 computed with the Martin-Hou equation of state. Extraction of the
thermodynamic states along a streamline starting at the middle of the inlet for IC1 (white line) and IC2 (red
line) inlet condition. Some iso-contours of Γ are also represented.

The choice of the wall temperature is an important parameter because the temperature imposed to materials in
turbine is usually a critical issue in the design of turbines. Heating of materials in perfect gas turbines can often
be a limiting factor to avoid the blade to be damaged. For dense gases, this temperature is less critical. Indeed, as
discussed in the introduction, dense gas turbines operate at temperatures around or less than 300◦ C, so that there are
few constraints for the blades compared to perfect gas turbines with temperatures over 1000◦ C. However, as seen in
Figure 2, the flow path starting from IC2 passes very close of the saturation curve. The main issue to fix the value of
wall temperature consists in avoiding the condensation of the flow, which is not compatible with the present flow solver
and could lead to corrosion of the material in a real turbine blade. After some tries, a value of the wall temperature of
656 K is retained, which the lowest value to avoid condensation. Maps of the mean temperature in the 2D simulations
of perfect and dense gases are shown in Figure 4. The expansion induces a temperature decrease. For the perfect gas,
the temperature variations in Figure 4a are shown in the range 300 K-420 K, meaning that the flow throughout the blade
passage is hotter than the wall. On the contrary, for the dense gas, the blade can be warmer than the flow. However low
temperature variations are experienced (in the range 640 K-660 K) due to the high thermal conductivity of dense gases,
so that the exact value of wall temperature is not very important. Note that the result would not have changed a lot by

6
using an adiabatic wall condition, but we have retained an isothermal condition at 656 K to keep the same numerical
setup as the perfect gas case.

(a) (b)

Fig. 3 Time-averaged distributions of the fundamental derivative of gas dynamic Γ obtained with 2D dense
gas simulations using IC1 (a) and IC2 (b) inlet conditions.

(a) (b)

Fig. 4 Time-averaged distributions of temperature obtained with 2D simulations using perfect gas (a) and
dense gas with IC2 inlet conditions (b).

7
 IV. LES Results
In this section, LES results are presented, first for air considered as a perfect gas, which serves as validation of the
numerical strategy and constitutes a reference point to highlight dense gas effects. Then the flow dynamics for the dense
gas is analysed.

A. Baseline perfect gas large eddy simulation

The flow conditions of MUR 129 are selected (see section III). First, the wall distributions of cell sizes ∆x + , ∆y +
and ∆z + expressed in wall units are given in Figure 5 for the mesh described in section III. The average first cell size
is 2.5 µm, corresponding to ∆y + ≈ 2. The resolutions in the streamwise and spanwise directions are ∆x + ≈ 100 and
∆z + ≈ 25. These values correspond to a coarse LES, but are similar to those used in [16]. The LES is initialized with a
preliminary 2D laminar calculation. Then, a sinusoidal perturbation of the conservative variables with an amplitude of
10 % is added in the spanwise direction to trigger transition toward a fully 3D field. The simulation is first run over
about ten flow-through times to evacuate the initial transient. A flow-through time is calculated as the time required for a
particle dropped at the blade leading edge to reach the trailing edge, when travelling at a constant velocity approximated
as the arithmetic average of the velocity at the passage inlet (x = 0 in our reference frame with origin at the blade
leading edge) and the velocity at the passage outlet (x = C cos χ). Afterwards, the statistics are collected over the five
subsequent flow-through times. The artificial dissipation coefficients of the spatial scheme were set to k 2 = 0. and
k 4 = 0.064 for this calculation. The dimensional time-step is 3 × 10−8 seconds, which corresponds to a maximum CF L
number of approximately 7, and a typical value of the smoothing parameter of IRS4 scheme is used, namely θ = 0.01.

160 4 40

140 3.5 35

120 3 30

100 2.5 25
+

+
∆x

∆y

80 2 ∆z 20

60 1.5 15

40 1 10

20 0.5 5

0 0 0
-0.06 -0.04 -0.02 0 0.02 0.04 0.06 0.08 -0.06 -0.04 -0.02 0 0.02 0.04 0.06 0.08 -0.06 -0.04 -0.02 0 0.02 0.04 0.06 0.08
Curvilinear Abscissa Curvilinear Abscissa Curvilinear Abscissa

(a) (b) (c)

Fig. 5 Resolution in wall units, ∆x + (a), ∆y + (b) and ∆z + (c), along the blade sides for the perfect gas simulation.

The comparison of the perfect gas LES with experimental and numerical literature data for the wall distribution of
the isentropic Mach number and heat transfer as a function of the curvilinear abscissa is presented in Figure 6. The
pressure side is represented with negative values of the curvilinear abscissa and the suction side with positive ones.
The leading edge corresponds to the abscissa 0 and the trailing edge to both extremes of the graph. The experimental
data of Arts et al. [11] and two different numerical results provided by Collado et al. [16] and Segui et al. [17] are
superimposed for comparison. The isentropic Mach number distribution is in very good agreement with the numerical
references, which is not unheard since this quantity arises essentially from the Eulerian dynamics. The comparison of
the heat transfer distribution with both experimental and numerical data is fairly good. In particular, the location of
the laminar-to-turbulent transition is well captured by the current LES and appears to be located around 70 mm on the
suction side. However, the simulation slightly underpredicts the heat transfer between 20 mm and 70 mm on the suction
and between -64 mm and -30 mm on the pressure side. The behavior near the leading edge is also fairly reproduced by
the current LES.

8
 1000

PFG LES
Experiment
0.8 Collado et al.
800 Segui et al.
Isentropic Mach number

0.6

Heat Transfer
600
PFG LES
Experiment
Collado et al.
0.4
400

0.2
200

0
0
-0.06 -0.04 -0.02 0 0.02 0.04 0.06 0.08 -0.06 -0.04 -0.02 0 0.02 0.04 0.06 0.08
Curvilinear Abscissa Curvilinear Abscissa

(a) (b)

Fig. 6 Validation of the perfect gas calculation. Wall distributions of the isentropic Mach number (a) and the
heat transfer (b).

B. Analysis of dense gas results

The inlet conditions correspond to point IC2, whose parameters are summarized in Table 2 and the wall temperature
is fixed at 656 K. The cell sizes in wall units, ∆x + , ∆y + and ∆z + , are shown in Figure 7 along the blade surface. The
average first cell size is 0.125 µm, corresponding to ∆y + ≈ 1. The average resolutions in the streamwise and spanwise
directions are ∆x + ≈ 60 and ∆z + ≈ 15. The resolution obtained for the dense gas is finer than the one in the perfect gas
case due to the different behavior of the friction on the blade, yielding a resolution acceptable for a LES calculation.
As in the perfect gas calculation, the initial transient is evacuated after ten flow-through times and the statistics are
collected over the five subsequent flow-through times. The artificial dissipation coefficients of the spatial scheme were
taken equal to k2 = 2. and k4 = 0.083 for this calculation. The simulation involving several shocks, the dissipation
parameters were increased with respect to the perfect gas case for sake of numerical stability. The dimensional time-step
is 6.5 × 10−9 seconds, which corresponds to a maximum CF L number of approximately 7, and the smoothing parameter
of IRS4 is θ = 0.02.

250 4 50

45
3.5

200 40
3
35

2.5
150 30
+

+
∆x

∆y

∆z

2 25

100 20
1.5

15
1
50 10

0.5
5

0 0 0
-0.003 -0.002 -0.001 0 0.001 0.002 0.003 0.004 -0.003 -0.002 -0.001 0 0.001 0.002 0.003 0.004 -0.003 -0.002 -0.001 0 0.001 0.002 0.003 0.004
Curvilinear abscissa Curvilinear abscissa Curvilinear abscissa

(a) (b) (c)

Fig. 7 Resolution in wall units, ∆x + (a), ∆y + (b) and ∆z + (c), along the blade sides for the dense gas simulation.

A global view of the flow through the turbine cascade is provided in Figure 8, where the Q-criterion (Q=1000)
colored by the velocity magnitude is displayed. Flows in perfect and dense gas are very different. The first one is fully
subsonic with no shocks in the throat whereas the second is supersonic after the throat and exhibits a complex shock
system. That is why quantitative comparisons between perfect and dense gases are very complicated. Qualitatively,

9
an other difference between perfect and dense gas is the deviation of the wake. For the perfect gas, the flow remains
aligned with the end of the blade while, for the dense gas, a deviation of the wake of roughly 20◦ is observed.

(a) (b)

Fig. 8 Instantaneous snapshots for the perfect (a) and dense gas (b) simulations. The isocontour represents
the Q-criterion, Q=1000, colored by the velocity magnitude, and the background shows the normalized density
gradient.

To better understand what is happening in the dense gas case, time-averaged distributions of several quantities are
displayed in Figure 9. Furthermore, the results along two streamlines, denoted "line 1" and "line 2", are extracted and
plotted in Figure 10. The shock system is composed of three different shocks labeled with letters in Figures 9 and
10. The first shock, denoted A, corresponds to the lower fish-tail shock from the preceding blade, which impinges on
the suction side. Then a weak shock, denoted B, is visible in the middle of the suction side. Finally, the last shock,
denoted C, is located after the trailing edge and corresponds to the upper part of the fish-tail shock attached to the
trailing edge. The density map in Figure 9a shows the expansion process with highest values at the inlet, followed by a
rapid decay induced by the converging walls. The fundamental derivative of gas dynamics Γ, represented in Figure 9d,
decreases rapidly from a value of nearly 7 at the inlet. As long as Γ is greater than one, the sound speed (Fig.9b and
10b) is diminishing and the flow becomes supersonic, as seen in the Mach number plots (Fig.9c and 10c). Near the
neck formed between the round section of the trailing edge and the suction side, Γ becomes lower than unity while the
density continues to decrease, leading to a change in the direction of variation of the sound speed which undergoes a
strong increase (Fig.10b). The drop and rise of the sound speed induces shock A. Since it occurs in a region where the
flow is expanding, this shock can be called an expansion shock. It induces a weak dissipation and does not produce an
important entropy variation as shown in Figures 9f and 10f. Shocks B and C correspond to an increase of density, thus
an increase of pressure, necessary to adapt the pressure to its imposed exit value. Shock B is seen to play a role in the
laminar-to-turbulent transition of the suction-side boundary layer. A significant entropy deviation is generated at this
location (Fig.10f). Both shocks B and C are compression shocks associated with an increase of the density. Since Γ is
still lower than unity, the sound speed is seen to diminish across these shocks (Fig.10b). The velocity norm (Fig.9e and
10e) also decreases, so that the flow remains supersonic up to the exit. Shock C is the second shock of the fish-tail shock
attached to the trailing edge. The different nature of the two shocks forming the fish-tail system can be responsible to
the important deviation of the wake by inducing a local pressure gradient.

10
 (a) (b)

(c) (d)

(e) (f)

Fig. 9 Time averaged distribution of dense gas LES spanwise averaged of density (a), Sound speed (b), Mach
number (c), Γ (d), velocity norm (e) and entropy increase Σ (f). Two streamline are also displayed: line 1 denoted
with points having a subscript 1 and line 2 with a subscript 2.

11
 60 line 1
line 1 line 2
800
line 2
55
B2
A1 B1
50

Sound Speed
600
45
ρ

C2
40
A2
400
C1 35
A2 C1
C2
A1 B2 30
B1
200
25
0.001 0.0015 0.002 0.0025 0.003 0.001 0.0015 0.002 0.0025 0.003
X X

(a) (b)

1.4 2

line 1
A2 B2 line 2
1.2 A1 C2
B1 1.5

1 C1

1
Mach

0.8
Γ

0.6
0.5

0.4 B1 B2
A2
line 1 C2
C1
line 2 A1
0.2 0

0.001 0.0015 0.002 0.0025 0.003 0.001 0.0015 0.002 0.0025 0.003
X X

(c) (d)

0.2

70 0.18 C1 line 1
line 2
0.16
60 A1 B1
B2 B1
0.14
50 C2
Norm Velocity

0.12
Σ

40 0.1
C1
0.08
30 A2
0.06

20 0.04
line 1
line 2 C2
0.02 B2
10
A1 A2
0
0.001 0.0015 0.002 0.0025 0.003 0.001 0.0015 0.002 0.0025 0.003
X X

(e) (f)

Fig. 10 Extraction of values on streamlines 1 and 2 displayed in Figure 9 of density (a), sound speed (b), Mach
number (c), Γ (d), velocity norm (e) and entropy increase Σ (f).

12
C. Transition mechanism
The laminar-to-turbulent transition of the suction-side boundary layer for the dense gas case deviates significantly
from the natural transition observed in the perfect-gas MUR129 configuration. Figure 6 locates the transition around
70 mm in the perfect gas case, whereas the transition appears earlier in the dense gas case. In Figure 11, wall friction
coefficient and heat transfer show a clear peak around 2.5 mm on the suction side, which seems related to the position
of the shock B.
The heat transfer takes negative values on the suction side after 1,7 mm due to heating of the flow by the blade.
Recall that the wall temperature was chosen to avoid condensation but here, the wall warms the flow instead of cooling
it as in usual turbine flows. Regardless of those negative values, the heat transfer is also 100 times higher than for the
perfect gas because the thermal conductivity is approximately 1000 times greater and the temperature gradient is 10
times smaller. However the shape of the wall heat transfer distribution is similar to the perfect gas one in the vicinity of
the leading edge. It also can be noticed that the friction coefficient exhibits two regions of negative values caused by
small separation bubbles on the suction side at positions 2.6 mm and 3.1 mm. These separation bubbles explain the
substantially different flow topology compared to the perfect gas case.
To better visualize the transition mechanism the distribution of turbulence intensity is displayed in Figure 12. The
transition starts at the foot of shock B. Important turbulent fluctuations are visible in the shock region, which could be
explained by an oscillating movement of the shock. Spectral analysis would be helpful to investigate this point. Shock
A is attached to the trailing edge of the preceding blade, so that it does not induced kinetic energy variation. The foot of
shock A only generates a very small separation bubble, visible in the close-up view of Figure 12b. Another large patch
of high turbulence intensity is visible after the blade passage, which can be related to a complex interaction between the
shock system and the wake.

0.35 4

0.3
2

0.25
Heat transfer (x 10-4)

0
Friction coefficient

0.2

0.15 -2

0.1
-4
0.05

-6
0

-0.05
-8
-0.003 -0.002 -0.001 0 0.001 0.002 0.003 0.004 -0.003 -0.002 -0.001 0 0.001 0.002 0.003 0.004
Curvilinear abscissa Curvilinear abscissa

(a) (b)

Fig. 11 Wall distribution of friction coefficient (a) and heat transfer (b) for the dense gas flow.

V. Conclusion and perspectives

The LS-89 turbine cascade has been simulated using dense gas equation of state. The simulation strategy has first
been validated for perfect gas on the well-documented MUR129 configuration of Arts et al. [11], showing a good
agreement with the experimental and numerical literature results. A preliminary study has then been conducted to
choose the inlet thermodynamic conditions. This operating point was chosen in order to locate the region of negative
values of the fundamental derivative of gas dynamics Γ in the vicinity of the trailing edge, where the shock system
develops.
A peculiar shock system is observed in the LES of the dense gas turbine cascade. A first shock, corresponding to
the lower fish-tail shock attached to the trailing edge, appears at the location where the sound speed variation changes
of sign. This shock has the characteristics of an expansion shock and induces almost no dissipation. After this shock,
two compression shocks take place to adapt the pressure to the imposed exit value. The first weak compression shock is

13
 (a) (b)

Fig. 12 Time-averaged distribution of turbulence intensity generated by the flow (a) and close-up view around
the foot of the expansion shock (b).

located on the suction side and the second one corresponds to the upper part of the fish-tail shock. The different nature
of the two shocks forming the fish tail induces a pressure gradient responsible for the deviation of the wake by roughly
20◦ . The flow remains supersonic from the neck to the oulet. The role of negative values of fundamental derivative is
still not clear, and further analyses are necessary to understand its role in the shock dynamics.
The transition mechanism is also described. In the dense gas case, it is triggered by the interaction of the compression
shock above the suction side with the boundary layer. The shock induces an adverse pressure gradient, leading to a
small separation bubble. The strong levels of turbulent fluctuations, coupled with shock oscillations, trigger the eruption
of turbulence in the boundary layer. The picture is very different for the perfect gas since no shock is present and a
natural by-pass transition occurs.
This simulation is one of the first LES database using a dense gas equation of state. It could be used to assess
RANS modeling for dense gases. Future work will study the accuracy of RANS simulations based on the k − ω model
compared to the present LES results. A parametric study will be performed to examine the influence of the pressure
ratio or the inlet thermodynamic operating conditions on the dense gas dynamics in order to analyse more precisely
the effects of negative values of the fundamental derivative of gas dynamics. Furthermore, LES simulations could be
performed for finest grids, higher Reynolds numbers or more realistic rotor or stator blade geometries. Due to the high
computational burden, turbulent wall models would certainly be required.

Acknowledgments
This work was granted access to the HPC resources of IDRIS National Scientific Computing Centre, under the
allocation number A0052A07332. This study was also funded by DGA (Direction Générale de l’Armement) under PhD
contract 2016 60 0043.

References
[1] Cinnella, P., and Congedo, P. M., “Inviscid and viscous aerodynamics of dense gases,” J. Fluid Mech., Vol. 580, 2007, pp.
179–217. doi:10.1017/S0022112007005290.

[2] Martin, J. J., and Hou, Y.-C., “Development of an equation of state for gases,” AIChE J., Vol. 1, No. 2, 1955, pp. 142–151.
doi:10.1002/aic.690010203.

[3] Chung, T., Ajlan, M., Lee, L., and Starling, K., “Generalized multiparameter correlation for nonpolar and polar fluid transport
properties,” Ind. & Eng. Chemistry, Vol. 27, No. 4, 1988, pp. 671–679. doi:10.1021/ie00076a024.

14
 [4] Cinnella, P., and Congedo, P. M., “Aerodynamic performance of transonic Bethe-Zel’dovich-Thompson flows past an airfoil,”
AIAA Journal, Vol. 43, No. 2, 2005, pp. 370–378. doi:10.2514/1.8627.

[5] Jameson, A., and Baker, T., “Solution of the Euler equations for complex configurations,” Fluid Dynamics and Co-located
Conferences, American Institute of Aeronautics and Astronautics, 1983, pp. –. doi:10.2514/6.1983-1929.

[6] Cinnella, P., and Content, C., “High-order implicit residual smoothing time scheme for direct and large eddy simulations of
compressible flows,” J. Comput. Phys., Vol. 326, 2016, pp. 1–29. doi:10.1016/j.jcp.2016.08.023.

[7] Lerat, A., Sidès, J., and Daru, V., “An implicit finite-volume method for solving the Euler equations,” Eighth International
Conference on Numerical Methods in Fluid Dynamics, edited by E. Krause, Springer Berlin Heidelberg, Berlin, Heidelberg,
1982, pp. 343–349. doi:10.1007/3-540-11948-5_41.

[8] Blazek, J., Rossow, C.-C., Kroll, N., and Swanson, R., “A comparison of several implicit residual smoothing methods in
combination with multigrid,” Thirteenth International Conference on Numerical Methods in Fluid Dynamics, edited by
M. Napolitano and F. Sabetta, Springer Berlin Heidelberg, Berlin, Heidelberg, 1993, pp. 386–390. doi:10.1007/3-540-56394-
6_253.

[9] Hoarau, J.-C., Cinnella, P., and Gloerfelt, X., “Large Eddy Simulation of turbomachinery flows using a high-order Implicit
Residual Smoothing scheme,” ICCFD Conference, Vol. ICCFD10-311, 2018.

[10] Bogey, C., and Bailly, C., “A family of low dispersive and low dissipative explicit schemes for flow and noise computations,” J.
Comput. Phys., Vol. 194, No. 1, 2004, pp. 194–214. doi:10.1016/j.jcp.2003.09.003.

[11] Arts, T., Lambert de Rouvroit, M., and Rutherford, A. W., “Aero-thermal investigation of a highly loaded transonic linear
turbine guide vane cascade. A test case for inviscid and viscous flow computations,” VKI Exp. Tech. Note 174, Vol. 91, 1990.

[12] Bufi, E., “Robust optimization of ORC turbine expanders,” Ph.D. thesis, Ecole Nationale Supérieure d’Arts et Métiers, 2016.

[13] Galiana, F., Wheeler, A., and Ong, J., “A study of trailing-edge losses in Organic Rankine Cycle turbines,” Journal of
Turbomachinery, Vol. 138, No. 12, 2016, pp. 121003–121003–9. doi:10.1115/1.4033473.

[14] Sciacovelli, L., Cinnella, P., and Grasso, F., “Small-scale dynamics of dense gas compressible homogeneous isotropic
turbulence,” J. Fluid Mech., Vol. 825, 2017, pp. 515–549. doi:10.1017/jfm.2017.415.

[15] Cramer, M. S., and Tarkenton, G. M., “Transonic flows of Bethe-Zel’dovich-Thompson fluids,” J. Fluid Mech., Vol. 240, 1992,
pp. 197–228. doi:10.1017/S0022112092000077.

[16] Collado Morata, E., Gourdain, N., Duchaine, F., and Gicquel, L., “Effects of free-stream turbulence on high pressure turbine
blade heat transfer predicted by structured and unstructured LES,” Int. J. Heat Mass Transfer, Vol. 55, No. 21, 2012, pp.
5754–5768. doi:10.1016/j.ijheatmasstransfer.2012.05.072.

[17] Segui, L., Gicquel, L., Duchaine, F., and Laborderie, J., “LES of the LS89 cascade: influence of inflow turbulence on the flow
predictions,” ETC 2017, , No. ETC2017-159, 2017. doi:10.29008/ETC2017-159.

15

View publication stats