You are on page 1of 23

Large Eddy Simulation of Transitional and Turbulent

Hypersonic Flow

Natan Hoffmann∗ , Amareshwara Sainadh Chamarthi† , Hemanth Chandra Vamsi K‡ , and Steven Frankel§
Technion-Israel Institute of Technology, Haifa, Israel

Large eddy simulation (LES) of transitional and turbulent hypersonic flow are carried out.
We focus on two canonical flow systems: oblique shock impingement on a hypersonic boundary
layer and hypersonic compression ramp flow. We employ the gradient based reconstruction
method, monotonicity-preserving (MP) limiting, a novel discontinuity sensor, and the Harten-
Lax-van Leer-Contact (HLLC) approximate Riemann solver to discretize the inviscid terms of
the governing equations with excellent dissipation and dispersion properties. For the viscous
terms, we employ a fourth-order 𝛼-damping scheme to avoid odd-even decoupling. The
importance of low dissipation smooth flow numerical methods, sensitive discontinuity detecting,
and wall stress modelling is emphasized. The first case undertaken is that of oblique shock
impingement caused by a 4◦ deflection on a disturbed Mach 6 laminar boundary layer. The
compressible equilibrium ODE wall model is used to reduce computational cost. The present
wall modelled LES (WMLES) quantitatively matches previous experiments, direct numerical
simulation (DNS), and previous WMLES. The second case carried out is Mach 7.7 transitional
flow over a 15◦ compression ramp. The Reynolds number based on the flat plate length of
0.1 m is 4.2 × 105 . The wall model is not used for this case as the flow does not fully transition
to turbulence. The present implicit LES (ILES) is shown to quantitatively match previous
experiment and DNS, albeit with some discrepancy due to coarse wall-normal grid spacing. The
final case examined is a longer 15◦ compression ramp at a higher Reynolds number of 8.6 × 105 .
The third case becomes fully turbulent and as such, WMLES is used for this case. While some
regions quantitatively match experiment and DNS, the separation bubble is not well captured
by the equilibrium assuming wall model, the reattachment location heating is over-predicted,
and the coarse grid under-predicts wall heating in the downstream portion of the domain where
the flow is fully turbulent.

I. Introduction
ypersonic boundary layer transition has a drastic effect on heat transfer, skin friction, boundary layer separation,
H and the design of a hypersonic vehicle’s thermal protection system [1]. The assumption of the location of boundary
layer transition can affect the vehicle gross take off weight by a factor of two or more [2]. As such, knowledge of when
and where transition to turbulence occurs is critical to the design of hypersonic vehicles. The presence of unsteady shock
structures and shock boundary layer interactions complicate the prediction of boundary layer transition considerably.
There are various model problems that feature both hypersonic boundary layer transition, shocks, and their interaction.
Two of the most basic are transitional hypersonic shock boundary layer interactions over flat plates [3] and transitional
hypersonic compression ramp flow [4]. In these flow systems, a hypersonic freestream flows over a flat plate and is
subject to an adverse pressure gradient by virtue of either an impinging shock or ramp. If this adverse pressure gradient
is sufficiently large, the boundary layer may separate from the wall. Depending on various factors including the velocity
of the flow, deflection angle, or wall temperature, a separation bubble can form, adding new complexities to the flow
system. From the separation bubble comes a separation shock, expansion fan, and reattachment shock, along with
shock-shock interactions and shear layers. Together with external perturbations arising from the freestream, surface
roughness, or nose bluntness, disturbances may enter the boundary layer, amplify, destabilize, and eventually transition
to turbulence.
∗ PhD Student.
† Post-Doctoral Fellow.
‡ PhD Student.
§ Professor.

1
Numerical simulations of hypersonic turbulent flows involving shocks represent a significant computational challenge.
While effort must be put forth to minimize high-frequency oscillations incumbent to shocks, excessive dissipation may
dampen flow structures required for accurate simulations of turbulence. This dilemma is especially acute for hypersonic
boundary layer transition simulations, in which small disturbances, perturbations, or instabilities may be critical to the
process.
In recent years, the theoretical, experimental, and numerical studies of hypersonic shock boundary layer interaction
and hypersonic compression ramp flows have received considerable attention. Sandham et al. [5] experimentally and
numerically studied oblique shock impingement from a 4◦ wedge on a hypersonic disturbed laminar boundary layer.
Schülein [6] and Willems et al. [7] studied the same case experimentally. In all experiments, the boundary layer
transitioned shortly downstream of the shock impingement region with the help of wind tunnel freestream disturbances.
In an attempt to model the freestream disturbances, Sandham et al. [5] used density disturbances in their subsequent
DNS. Without the disturbances, the shock impingement was insufficient to cause boundary layer transition alone.
Roghelia et al. [8, 9] studied hypersonic transitional flow over a 15◦ compression ramp. These experiments served
as the basis for subsequent DNS studies by Cao et al. [10–13] as well as global stability analyses of Hao et al. [14].
Cao et al. [11] explored the unsteady and three-dimensional behavior of the nominally two-dimensional flow system
at Mach 7.7 with a unit Reynolds number of Re/𝑚 = 4.2 × 106 m−1 . Later, Cao et al. [13] slightly increased the unit
Reynolds number to 𝑅𝑒 /𝑚 = 8.6 × 106 m−1 and the length of the flow domain to investigate the intrinsic instability of
the flow system. The DNS studies of this case culminated in the finding that the flow system was intrinsically unstable
and would transition to turbulence through numerical error given a long enough flow domain.
While DNS studies of such flow systems are integral to the understanding of their underlying physics, it is currently
infeasible to use DNS for complex geometries or moderate to high Reynolds number flows on a wide scale. Yang
and Griffin [15] revisited the grid point and time step requirements of Choi and Moin [16] and concluded that the
computational cost for a DNS scales as Re2.91 𝐿 . This nearly-cubic dependence on the Reynolds number renders DNS
extremely computationally expensive. Even for wall resolved LES (WRLES), Yang and Griffin [15] estimate the
computational cost to scale as Re2.72𝐿 , which while slightly alleviated, is still extremely computationally expensive.
Indeed, WRLES has been termed quasi-DNS [17]. Fortunately, WMLES, in which all inner layer dynamics are modelled,
have received a lot of attention and development recently – and for good reason: the computational cost estimate is a
drastically lower, nearly-linear Re1.14
𝐿 .
Yang et al. [18], Mettu and Subbareddy [19], and Ganju et al. [20] all performed WMLES of the shock disturbed
laminar boundary layer interaction case of Sandham et al. [5] using the compressible equilibrium ODE wall model of
Kawai and Larsson [21]. However, the disturbance amplitude had to be increased to match the DNS and experimental
results. Furthermore, the choice of numerical scheme was surely an extremely important factor in the ability to simulate
transition on the coarse grids typical of WMLES. This was recently reported in De Vanna et al. [22] in their study of
the effect of convective schemes in WRLES and WMLES of compressible wall turbulence. WMLES of this shock
boundary layer interaction case serves as an effective benchmark of the numerical method and wall model employed
since transition will not be predicted without extremely sensitive numerics and a wall stress model to correctly compute
the wall stresses. WMLES of compression ramp cases using the compressible equilibrium ODE wall model of Kawai
and Larsson [21] are particularly difficult, since the strong adverse pressure gradient caused by flow deflection directly
contradicts the equilibrium assumption central to the formulation of their wall model. However, the equilibrium wall
model has been used to some success for compression ramp flows, even with separation bubbles. Mettu and Subbareddy
used the equilibrium wall model to simulate Mach 5 flow over a 36◦ compression ramp. Later, Mettu and Subbareddy
[23] tested the equilibrium wall model and a non-equilibrium wall model using the full RANS equations for multiple
compression ramp flows. Van Noordt et al. [24] used the equilibrium wall model in the context of the immersed
boundary method to simulate Mach 7.2 flow over an 8◦ compression ramp. It is evident from these studies that while
in some regions the equilibrium wall model can successfully predict correct wall stresses, regions including strong
non-equilibrium effects show discrepancy and improvements must be made.
On the numerical methods front, it is clear that little-to-no dissipation schemes in regions of smooth flow are a
requirement for transitional WMLES. On the other hand, effective shock capturing methods are required to apply
sufficient dissipation in regions of discontinuities to stabilize high frequency oscillations. To synthesize these two
requirements, a shock sensor is necessary to detect when to and when not to apply the two candidate schemes. In
this work, we employ the gradient-based-reconstruction method of Chamarthi [25] to spatially discretize the inviscid
flux terms of the governing equations. This method uses high-order Legendre-based interpolation to cell interfaces
in a conservative finite difference framework to achieve excellent spectral properties. Moreover, the scheme requires
derivatives to be computed for interpolation, allowing for their re-use in the viscous flux discretization, as well as for

2
shock sensor methods or post process quantities. Furthermore, for this work, we employ an MP limiter in conjunction
with the Ducros shock sensor to detect and limit oscillations due to shock waves. We also make use of a novel
characteristic based discontinuity sensor [26]. For spatial discretization of the viscous fluxes, we employ the 𝛼-damping
approach of Chamarthi [25], which was introduced by Nishikawa [27] to remedy the well-known odd-even decoupling
problem. This viscous flux discretization was shown to be very important for turbulent flow simulations [28]. In
summary, for transitional hypersonic flow simulations on coarse grids, it is important to a) carefully consider the
numerical methods used, and b) use a model to represent the unresolved inner layer dynamics.
The rest of the paper goes as follows: in Section II, the governing equations are presented. In Section III, the
numerical methods are presented, including a delineation of the gradient-based-reconstruction approach, the implemented
shock and discontinuity sensor, 𝛼-damping viscous flux spatial discretization, and the compressible equilibrium ODE
wall model. After, results for the present WMLES of oblique shock impingement on a hypersonic disturbed laminar
boundary layer are discussed. Then, results for the present ILES of transitional hypersonic flow over a 15◦ compression
ramp are laid out. Lastly for the results, the present WMLES of the intrinsically unstable hypersonic 15◦ compression
ramp are examined. Finally, concluding remarks are made, and future work is set forth.

II. Governing Equations


In this study, the three-dimensional compressible Navier-Stokes equations in conservative form were solved in
curvilinear coordinates:
𝜕U 𝜕F𝑐 𝜕G𝑐 𝜕H𝑐 𝜕F𝑣 𝜕G𝑣 𝜕H𝑣
+ + + + + + = 0, (1)
𝜕𝑡 𝜕𝜉 𝜕𝜂 𝜕𝜁 𝜕𝜉 𝜕𝜂 𝜕𝜁
where 𝑡 is time and (𝜉, 𝜂, 𝜁) are the computational coordinates. U is the conserved variable vector, and F𝑐 , G𝑐 , and H𝑐
are the convective flux vectors defined as:

© 𝜌ª © 𝜌𝑈 ª © 𝜌𝑉 ª © 𝜌𝑊 ª
­ 𝜌𝑢 ® ­ 𝜌𝑈𝑢 + 𝜉e𝑥 𝑝 ® ­ 𝜌𝑉𝑢 + 𝜂e𝑥 𝑝 ® ­ 𝜌𝑊𝑢 + 𝜁e𝑥 𝑝 ®
1 ­­ ®® ­ ® ­ ® ­ ®
U = ­ 𝜌𝑣 ® , F𝑐 = ­ 𝜌𝑈𝑣 + 𝜉e𝑦 𝑝 ® , G𝑐 = ­ 𝜌𝑉 𝑣 + 𝜂e𝑦 𝑝 ® , H𝑐 = ­ 𝜌𝑊𝑣 + 𝜁e𝑦 𝑝 ® ,
­ ® ­ ® ­ ®
(2a-2d)
𝐽­ ® ­ ® ­ ® ­ ®
­ 𝜌𝑤 ® ­ 𝜌𝑈𝑤 + 𝜉e𝑧 𝑝 ® ­ 𝜌𝑉 𝑤 + 𝜂e𝑧 𝑝 ® ­ 𝜌𝑊 𝑤 + 𝜁e𝑧 𝑝 ®
­ ® ­ ® ­ ® ­ ®
« 𝜌𝐸 ¬ « 𝜌𝑈𝐻 ¬ « 𝜌𝑉 𝐻 ¬ « 𝜌𝑊 𝐻 ¬

where 𝐽 is the Jacobian of the transformation from physical to computational space, the wide tilde, f (·), denotes Jacobian
normalized grid metrics (e.g. 𝜉e𝑥 = 𝜉 𝑥 /𝐽), 𝜌 is the density
 and 𝑢, 𝑣, and 𝑤 are the velocities in the 𝑥, 𝑦, and 𝑧 directions,
respectively. 𝑝 is the pressure, 𝐸 = 𝑒 + 𝑢 2 + 𝑣 2 + 𝑤 2 /2 is the specific total energy, and 𝐻 = 𝐸 + 𝑝/𝜌 is the specific
total enthalpy. The equation of state is for a calorically perfect gas so that 𝑒 = 𝑝/[𝜌(𝛾 − 1)] −1 is the internal energy,
where 𝛾 = cp /cv is the ratio of specific heats with cp as the isobaric specific heat and cv as the isochoric specific heat. 𝑈,
𝑉, and 𝑊 are the contravariant velocities defined as:

𝑈 = 𝜉e𝑥 𝑢 + 𝜉e𝑦 𝑣 + 𝜉e𝑧 𝑤, 𝑉 = 𝜂e𝑥 𝑢 + 𝜂e𝑦 𝑣 + 𝜂e𝑧 𝑤, 𝑊 = 𝜁e𝑥 𝑢 + 𝜁e𝑦 𝑣 + 𝜁e𝑧 𝑤, (3a-3c)

where the subscripted computational coordinates are the grid metrics computed in conservative form as in Nonomura
et al. [29]. These metrics were computed using the same numerical method applied for spatial discretization, which
ensures freestream preservation [30]. F𝑣 , G𝑣 , and H𝑣 are the viscous flux vectors defined as:

© 0 ª © 0 ª © 0 ª
­𝜉e𝑥 𝜏𝑥 𝑥 + 𝜉e𝑦 𝜏𝑥 𝑦 + 𝜉e𝑧 𝜏𝑥𝑧 ® ­𝜂e𝑥 𝜏𝑥 𝑥 + 𝜂e𝑦 𝜏𝑥 𝑦 + 𝜂e𝑧 𝜏𝑥𝑧 ® ­𝜁e𝑥 𝜏𝑥 𝑥 + 𝜁e𝑦 𝜏𝑥 𝑦 + 𝜁e𝑧 𝜏𝑥𝑧 ®
­ ® ­ ® ­ ®
F𝑣 = − ­𝜉e𝑥 𝜏𝑦 𝑥 + 𝜉e𝑦 𝜏𝑦 𝑦 + 𝜉e𝑧 𝜏𝑦𝑧 ® , G𝑣 = − ­𝜂e𝑥 𝜏𝑦 𝑥 + 𝜂e𝑦 𝜏𝑦𝑦 + 𝜂e𝑧 𝜏𝑦𝑧 ® , H𝑣 = − ­ 𝜁e𝑥 𝜏𝑦 𝑥 + 𝜁e𝑦 𝜏𝑦𝑦 + 𝜁e𝑧 𝜏𝑦𝑧 ® , (4a-4c)
­ ® ­ ® ­ ®
­ ® ­ ® ­ ®
­ 𝜉e𝑥 𝜏𝑧 𝑥 + 𝜉e𝑦 𝜏𝑧 𝑦 + 𝜉e𝑧 𝜏𝑧𝑧 ® ­ 𝜂e𝑥 𝜏𝑧 𝑥 + 𝜂e𝑦 𝜏𝑧𝑦 + 𝜂e𝑧 𝜏𝑧𝑧 ® ­ 𝜁e𝑥 𝜏𝑧 𝑥 + 𝜁e𝑦 𝜏𝑧𝑦 + 𝜁e𝑧 𝜏𝑧𝑧 ®
­ ® ­ ® ­ ®
« 𝜉 𝑥 𝛽 𝑥 + 𝜉 𝑦 𝛽𝑦 + 𝜉𝑧 𝛽𝑧 ¬ « 𝜂e𝑥 𝛽 𝑥 + 𝜂e𝑦 𝛽 𝑦 + 𝜂e𝑧 𝛽 𝑧 ¬ « 𝜁 𝑥 𝛽 𝑥 + 𝜁 𝑦 𝛽𝑦 + 𝜁𝑧 𝛽𝑧 ¬
e e e e e e

where the normal stresses are defined as:

3
   
𝜕𝑢 ˆ 𝜕𝑢 𝜕𝑣 𝜕𝑤 𝜕𝑣 ˆ 𝜕𝑢 𝜕𝑣 𝜕𝑤
𝜏𝑥 𝑥 = 2 𝜇ˆ +𝜆 + + , 𝜏𝑦𝑦 = 2 𝜇ˆ +𝜆 + + , (5a-5b)
𝜕𝑥 𝜕𝑥 𝜕𝑦 𝜕𝑧 𝜕𝑦 𝜕𝑥 𝜕𝑦 𝜕𝑧
 
𝜕𝑤 ˆ 𝜕𝑢 𝜕𝑣 𝜕𝑤
𝜏𝑧𝑧 = 2 𝜇ˆ +𝜆 + + , (5c)
𝜕𝑧 𝜕𝑥 𝜕𝑦 𝜕𝑧

where 𝜇ˆ = 𝜇Ma/Re is the scaled dynamic viscosity as a result of non-dimensionalization and Stokes’ hypothesis is
 −1/2
assumed so that 𝜆ˆ = −2 𝜇/3.
ˆ Ma = 𝑢 ∞ 𝛾𝑅𝑔𝑎𝑠 𝑇 and Re = 𝜌∞ 𝑢 ∞ 𝐿 𝑟 𝑒 𝑓 /𝜇∞ are the Mach and Reynolds numbers,
respectively, where 𝑅𝑔𝑎𝑠 is the universal gas constant and the ∞ subscript denotes a freestream value. The shear stresses
are defined as:

     
𝜕𝑢 𝜕𝑣 𝜕𝑣 𝜕𝑤 𝜕𝑢 𝜕𝑤
𝜏𝑥 𝑦 = 𝜏𝑦 𝑥 = 𝜇ˆ + , 𝜏𝑦𝑧 = 𝜏𝑧𝑦 = 𝜇ˆ + , 𝜏𝑥𝑧 = 𝜏𝑧 𝑥 = 𝜇ˆ + , (6a-6c)
𝜕𝑦 𝜕𝑥 𝜕𝑧 𝜕𝑦 𝜕𝑧 𝜕𝑥

and the components of 𝛽 are:

𝜕𝑇 𝜕𝑇
𝛽 𝑥 = 𝑢𝜏𝑥 𝑥 + 𝑣𝜏𝑥 𝑦 + 𝑤𝜏𝑥𝑧 + 𝜅ˆ , 𝛽 𝑦 = 𝑢𝜏𝑦 𝑥 + 𝑣𝜏𝑦𝑦 + 𝑤𝜏𝑦𝑧 + 𝜅ˆ , (7a-7b)
𝜕𝑥 𝜕𝑦
𝜕𝑇
𝛽 𝑧 = 𝑢𝜏𝑧 𝑥 + 𝑣𝜏𝑧 𝑦 + 𝑤𝜏𝑧𝑧 + 𝜅ˆ , (7c)
𝜕𝑧

where 𝜅ˆ = 𝜇ˆ [(𝛾 − 1)Pr] −1 is the scaled thermal conductivity, Pr is the Prandtl number, and 𝑇 is the temperature. The
dynamic viscosity was taken as a function of temperature by Sutherland’s law:

1 + 𝑆/𝑇𝑟 𝑒 𝑓
𝜇(𝑇) = 𝑇 3/2 , (8)
𝑇 + 𝑆/𝑇𝑟 𝑒 𝑓
where 𝜇𝑟 𝑒 𝑓 is the reference dynamic viscosity, 𝑇𝑟 𝑒 𝑓 is the reference temperature, and 𝑆 = 110.4 K is Sutherland’s
constant. The equations were non-dimensionalized using the freestream density 𝜌∞ , the freestream speed of sound
𝑐 ∞ , reference length 𝐿 𝑟 𝑒 𝑓 , the freestream temperature 𝑇∞ , and the freestream dynamic viscosity 𝜇∞ such that the
temperature is related to pressure and density via 𝑝 = 𝜌𝑇/𝛾.

III. Numerical Methods


In this work, we employed the gradient-based-reconstruction method of Chamarthi [25] with the extension of a novel
discontinuity sensor to spatially discretize the numerical convective fluxes. In what follows, we delineate the details of
this method. After, we briefly present the viscous flux discretization, wall stress model, and the time integration method.

A. Convective Flux Spatial Discretization Scheme


Using
h a conservative
i h numerical i method,
h the igoverning equations cast in semi-discrete form for a curvilinear cell
𝐼𝑖, 𝑗,𝑘 = 𝜉𝑖− 1 , 𝜉𝑖+ 1 × 𝜂𝑖− 1 , 𝜂𝑖+ 1 × 𝜁𝑖− 1 , 𝜁𝑖+ 1 can be expressed via the following ordinary differential equation:
2 2 2 2 2 2


d dF̌𝑐 dǦ𝑐 dȞ𝑐
Ǔ𝑖, 𝑗,𝑘 = Res𝑖, 𝑗,𝑘 = − − −
d𝑡 d𝜉 𝑖, 𝑗,𝑘 d𝜂 𝑖, 𝑗,𝑘 d𝜁 𝑖, 𝑗,𝑘
(9)
dF̌𝑣 dǦ𝑣 dȞ𝑣
+ + + ,
d𝜉 𝑖, 𝑗,𝑘 d𝜂 𝑖, 𝑗,𝑘 d𝜁 𝑖, 𝑗,𝑘

ˇ indicates a numerical approximation of a physical quantity, Res𝑖, 𝑗,𝑘 is the residual function,
where the check accent, (·),
and the remaining terms are cell center numerical flux derivatives of the physical fluxes in Eqn. 1. For brevity, we
continue with only the 𝜉-direction, however, the following may be extended to all three dimensions straightforwardly.

4
Moreover, we drop the 𝑗 and 𝑘 indices in the interest of clarity. The cell center numerical convective flux derivative is
expressed as:

dF̌𝑐 1  𝑐 𝑐

= F̌ − F̌ , (10)
d𝜉 𝑖 Δ𝜉 𝑖+ 21 𝑖− 12

where 𝑖 ± 12 indicates right and left cell interface values, respectively. F̌𝑐 are computed using an approximate Riemann
𝑖± 12
solver, since a Riemann problem exists at each cell interface. The interface numerical convective fluxes are computed
from:
1h 𝑐 𝐿   i 1  
𝑐
F̌ Ǔ𝑖± 1 + F̌𝑐 Ǔ𝑖±𝑅 𝑅 𝐿
F̌𝑖± 1 = − A𝑖± 1 Ǔ𝑖± 1 − Ǔ

1 1 ,
𝑖± 2
(11)
2 2 2 2 2 2 2


where the 𝐿 and 𝑅 superscripts denote the left- and right-biased states, respectively, and A𝑖± 1 denotes the convective
2
flux Jacobian. In this work, the HLLC approximate Riemann solver was used, the details for which may be found
in Section III.A.5. The objective is to obtain the left- and right-biased states. These were computed with the
gradient-based-reconstruction method, which will be explained in the following subsection.

1. Gradient-Based-Reconstruction Method: Linear Scheme


Gradient-based reconstruction methods employ the first two moments of the Legendre polynomial evaluated on
𝜉𝑖− 1 ≤ 𝜉 ≤ 𝜉𝑖+ 1 for interpolation. This may be written for a general variable, 𝜙, as:
2 2
" #
𝜙𝑖′ 3𝜙𝑖′′ 2
Δ𝜉𝑖2
𝜙(𝜉) = 𝜙𝑖 + (𝜉 − 𝜉𝑖 ) + 𝒦 (𝜉 − 𝜉𝑖 ) − , (12)
Δ𝜉 2Δ𝜉𝑖2 12
where 𝜙𝑖′ and 𝜙𝑖′′ respectively represent the first and second derivatives of 𝜙𝑖 . If 𝜉 = 𝜉𝑖 + Δ𝜉/2 and 𝒦 = 1/3, the
following equations for the left- and right-biased states are obtained:

𝐿 1 ′ 1 ′′ 𝑅 1 ′ 1 ′′
𝜙𝑖+ 1 = 𝜙𝑖 + 𝜙 + 𝜙 , 𝜙𝑖+ 1 = 𝜙 𝑖+1 − 𝜙𝑖+1 + 𝜙𝑖+1 . (13a–13b)
2 2 𝑖 12 𝑖 2 2 12

In this work, 𝜙𝑖′ was computed using eighth order explicit central differences:
 
1 1 4 1 4 4 1 4 1
𝜙𝑖′ = 𝜙𝑖−4 − 𝜙𝑖−3 + 𝜙𝑖−2 − 𝜙𝑖−1 + 𝜙𝑖+1 − 𝜙𝑖+2 + 𝜙𝑖+3 − 𝜙𝑖+4 . (14)
Δ𝜉 280 105 5 5 5 5 105 280
𝜙𝑖′′ was computed from:
2 1
𝜙𝑖′′ = (𝜙𝑖+1 − 2𝜙𝑖 + 𝜙𝑖−1 ) − 𝜙′ − 𝜙𝑖−1
′ 
. (15)
Δ𝜉 2 2Δ𝜉 𝑖+1
Since 𝜙 is an arbitrary variable, either primitive or conservative variables can be used. In this work, the conservative
variables were used.

2. Gradient-Based-Reconstruction Method: Non-Linear Scheme


Eqns. 13 are linear gradient-based interpolations. Therefore, they may be susceptible to oscillations in the presence
of discontinuities. So, MP limiting was employed. The following delineates the MP limiting procedure for the left-biased
state, however, the procedure is the same for the right-biased state. The MP limiting criterion is:
  
 𝐿,𝐿𝑖𝑛𝑒𝑎𝑟 𝐿,𝐿𝑖𝑛𝑒𝑎𝑟
− 𝜙𝑖 𝜙 1 𝐿,𝐿𝑖𝑛𝑒𝑎𝑟 𝐿,𝑀 𝑃
−𝜙 1 ≤ 10−10 ,

𝐿
𝜙 1

 if 𝜙 1
1 =
𝜙𝑖+ 𝑖+ 2 𝑖+ 2 𝑖+ 2 𝑖+ 2 (16)
2  𝐿, 𝑁 𝑜𝑛−𝐿𝑖𝑛𝑒𝑎𝑟
 𝜙𝑖+ 1
 otherwise,
 2
where 𝜙 𝐿,𝐿𝑖𝑛𝑒𝑎𝑟
1 corresponds to Eqn. 13a, and the remaining terms are:
𝑖+ 2

5
 
𝜙 𝐿, 1𝑁 𝑜𝑛−𝐿𝑖𝑛𝑒𝑎𝑟 = 𝜙 𝐿,𝐿𝑖𝑛𝑒𝑎𝑟
1 + minmod 𝜙 𝐿,𝑀 𝐼 𝑁
1 − 𝜙 𝐿,𝐿𝑖𝑛𝑒𝑎𝑟
1 , 𝜙 𝐿,𝑀 𝐴𝑋
1 − 𝜙 𝐿,𝐿𝑖𝑛𝑒𝑎𝑟
1 , (17a)
𝑖+ 2 𝑖+ 2 𝑖+ 2 𝑖+ 2 𝑖+ 2 𝑖+ 2

𝜙 𝐿, 1𝑀 𝑃 = 𝜙 𝐿,𝐿𝑖𝑛𝑒𝑎𝑟
1 + minmod [𝜙𝑖+1 − 𝜙𝑖 , 𝒜 (𝜙𝑖 − 𝜙𝑖−1 )] , (17b)
𝑖+ 2 𝑖+ 2
    
𝜙 𝐿, 1𝑀 𝐼 𝑁 = max min 𝜙𝑖 , 𝜙𝑖+1 , 𝜙 𝐿,𝑀
1
𝐷
, min 𝜙 𝑖 , 𝜙 𝐿,𝑈𝐿
1 , 𝜙 𝐿,𝐿𝐶
1 , (17c)
𝑖+ 2 𝑖+ 2 𝑖+ 2 𝑖+ 2
    
𝜙 𝐿, 1𝑀 𝐴𝑋 = min max 𝜙𝑖 , 𝜙𝑖+1 , 𝜙 𝐿,𝑀
1
𝐷
, max 𝜙 𝑖 , 𝜙 𝐿,𝑈𝐿
1 , 𝜙 𝐿,𝐿𝐶
1 , (17d)
𝑖+ 2 𝑖+ 2 𝑖+ 2 𝑖+ 2

1 1
𝜙 𝐿, 1𝑀 𝐷 = (𝜙𝑖 + 𝜙𝑖+1 ) − 𝑑 𝐿,𝑀 , (17e)
𝑖+ 2 2 2 𝑖+ 21
𝜙 𝐿,𝑈𝐿
1 = 𝜙𝑖 + 4 (𝜙𝑖 − 𝜙𝑖−1 ) , (17f)
𝑖+ 2

1 4
𝜙 𝐿,𝐿𝐶
1 = (3𝜙𝑖 − 𝜙𝑖−1 ) + 𝑑 𝐿,𝑀 , (17g)
𝑖+ 2 2 3 𝑖− 21
𝑑 𝐿,1𝑀 = minmod 𝜙𝑖′′ , 𝜙𝑖+1
′′ 
, (17h)
𝑖+ 2

1
where 𝒜 = 4 and minmod (𝑎, 𝑏) = 2 [sgn(𝑎) + sgn(𝑏)] min (|𝑎| , |𝑏|).

3. Ducros Shock Sensor


While MP limiting effectively mitigates oscillations arising from discontinuities, the limiting criterion in Eqn. 16
can become too sensitive and cause excessive dissipation. To remedy this issue, the Ducros shock sensor, which is
designed specifically to sense shocks, was used:

(∇ · u) 2
Ω𝑖 = 𝜃 𝑖 , (18)
(∇ · u) + |∇ × u| 2 + 𝜖
2

where,

| 𝑝 𝑖+1 − 2𝑝 𝑖 + 𝑝 𝑖−1 |
𝜃𝑖 = , (19)
| 𝑝 𝑖+1 + 2𝑝 𝑖 + 𝑝 𝑖−1 |
u is the velocity vector and 𝜖 = 10−30 to avoid division by zero. Note that the derivatives already computed from Eqn.
14 were re-used for Eqn. 18. We modified Ω𝑖 by using it’s maximum value in a three cell neighborhood:

Ω𝑖 = max (Ω𝑖+𝑚 ) , for 𝑚 = −1, 0, 1. (20)


Using Ω𝑖 , Eqn. 16 was modified to:
𝐿,𝐿𝑖𝑛𝑒𝑎𝑟
 𝜙𝑖+ 1


 if Ω𝑖 ≤ Ω,
𝐿
𝜙𝑖+ 1 = 2
𝐿, 𝑁 𝑜𝑛−𝐿𝑖𝑛𝑒𝑎𝑟
(21)
2  𝜙𝑖+ 1
 otherwise,
 2
where Ω is a cutoff value that is case-dependent. Using this method, shocks are detected well and the non-linear scheme
effectively limits oscillations.

4. Characteristic Space Transformation


To obtain results with even fewer oscillations, interpolation and limiting can be done in characteristic space.
Accordingly, we carried out the following procedure:
1) Compute Roe-averaged variables following Blazek [31] (Equation 4.89) to construct the left, L𝑛 , and right, R𝑛 ,
eigenvectors of the normal convective flux Jacobian.

6
2) Since in this work we used the conservative variables, transform Ǔ𝑖 , Ǔ𝑖′ , and Ǔ𝑖′′ to characteristic space by
multiplying them by L𝑛 :

′ ′
Č𝑖+𝑚,𝑏 = L𝑛,𝑖+ 1 Ǔ𝑖+𝑚 , Č𝑖+𝑚,𝑏 = L𝑛,𝑖+ 1 Ǔ𝑖+𝑚 , (22a–22b)
2 2

′′ ′′
Č𝑖+𝑚,𝑏 = L𝑛,𝑖+ 1 Ǔ𝑖+𝑚 , (22c)
2

for 𝑚 = −2, −1, 0, 1, 2, 3 and 𝑏 = 1, 2, 3, 4, 5, representing the vector of characteristic variables.


Using Eqns. 13, we then obtained the unlimited interpolation to cell interfaces in characteristic space via:

𝐿 1 ′ 1 ′′ 𝑅 1 ′ 1 ′′
Č𝑖+ 1 = Č𝑖,𝑏 + Č𝑖,𝑏 + Č𝑖,𝑏 , Č𝑖+ 1 = Č𝑖+1,𝑏 − Č𝑖+1,𝑏 + Č𝑖+1,𝑏 . (23a–23b)
2 ,𝑏 2 12 2 ,𝑏 2 12

The interpolations were then treated by an algorithm that selectively treats the characteristic waves. This is the subject
of a novel discontinuity sensor that has been submitted for peer-review [26]. After obtaining Č 𝐿,𝑅 1 , the interpolated
𝑖+ 2
states were then transformed back to physical space by multiplying the characteristic variable interpolation by R𝑛 :

Ǔ 𝐿,𝑅 𝐿,𝑅
1 = R𝑛,𝑖+ 1 Č 1 . (24)
𝑖+ 2 2 𝑖+ 2

Once the interface interpolations were obtained, we computed the flux through each cell interface via the HLLC
approximate Riemann solver.

5. Approximate Riemann Solver


The HLLC approximate Riemann solver in generalized coordinates was used for the Riemann problem at each cell
interface. Only the 𝜉 direction is presented here. This flux is defined as:



 F𝐿 , if 0 ≤ 𝑆 𝐿 ,

𝐻 𝐿𝐿𝐶
F𝐿 ,

 ∗ if 𝑆 𝐿 ≤ 0 ≤ 𝑆∗ ,
F𝑖±1/2 = (25)


 F∗𝑅 , if 𝑆∗ ≤ 0 ≤ 𝑆 𝑅 ,

 F𝑅 ,
 if 0 ≥ 𝑆 𝑅 ,
where,
 
F∗𝐾 = F𝐾 + 𝑆 𝐾 U∗𝐾 − U𝐾 , (26)

where 𝐾 is the left, 𝐿 or right, 𝑅 state. Extending from Pathak and Shukla [32], the star state is defined as:



 1 



 


 


   

𝜉b𝑥 𝑆∗ + 𝜉b𝑦2 + 𝜉b𝑧2 𝑢 𝐾 − 𝜉b𝑦 𝜉b𝑥 𝑣 𝐾 − 𝜉b𝑧 𝜉b𝑥 𝑤 𝐾

 


 


 

 



 𝜉b𝑥2 + 𝜉b𝑦2 + 𝜉b𝑧2 




   

 
𝜉b𝑦 𝑆∗ − 𝜉b𝑥 𝜉b𝑦 𝑢 𝐾 + 𝜉b𝑥2 + 𝜉b𝑧2 𝑣 𝐾 − 𝜉b𝑧 𝜉b𝑦 𝑤 𝐾
 𝐾  
 

𝐾 𝑆 −𝑈
𝐾 

 


𝐾
U∗ = 𝜌 , (27)
𝑆 − 𝑆∗ 
𝐾
 𝜉b𝑥2 + 𝜉b𝑦2 + 𝜉b𝑧2 


 


   

𝜉 𝑧 𝑆∗ − 𝜉 𝑥 𝜉 𝑧 𝑢 − 𝜉 𝑦 𝜉 𝑧 𝑣 + 𝜉 𝑥 + 𝜉 𝑦 𝑤
𝐾 𝐾 2 2 𝐾

 b b b b b b b 


 


 

 



 𝜉𝑥 + 𝜉𝑦 + 𝜉𝑧
b 2 b 2 b 2 




 " #


 𝐾


  𝑆 ∗ 𝑝 
𝐾 + 𝑆∗ − 𝑈 𝐾 + 𝐾 𝐾
 
𝐸

  

𝜉𝑥 + 𝜉𝑦 + 𝜉𝑧 𝜌 𝑆 − 𝑈
2 2 2 𝐾
b b b 
 

7
where the wide hat symbol, c(·) denotes grid metrics interpolated from cell centers to cell interfaces. Note that the
temporal metrics have been omitted. The left, right, and star wave speeds are respectively:
√︃
𝑆 𝐿 = min(𝑈 𝐿 − 𝑐 𝐿 𝜉b𝑥2 + 𝜉b𝑦2 + 𝜉b𝑧2 , 𝑈˘ − 𝑐),
˘ (28)

√︃
𝑆 𝑅 = max(𝑈 𝑅 + 𝑐 𝑅 𝜉b𝑥2 + 𝜉b𝑦2 + 𝜉b𝑧2 , 𝑈˘ + 𝑐),
˘ (29)

   
𝜌 𝑅 𝑈 𝑅 𝑆 𝑅 − 𝑈 𝑅 − 𝜌 𝐿 𝑈 𝐿 𝑆 𝐿 − 𝑈 𝐿 + 𝑝 𝐿 − 𝑝 𝑅 𝜉b𝑥2 + 𝜉b𝑦2 + 𝜉b𝑧2
𝑆∗ =   , (30)
𝜌𝑅 𝑆𝑅 − 𝑈 𝑅 − 𝜌𝐿 𝑆𝐿 − 𝑈 𝐿
˘ denotes Roe-averaged variables.
where the breve accent, (·),

B. Viscous Flux Spatial Discretization Scheme


In this subsection, we describe the spatial discretization of the numerical viscous fluxes. We used the fourth-order
𝛼-damping scheme of Chamarthi [25] and Chamarthi et al. [33], which is based on the 𝛼-damping approach of
Nishikawa [27]. The importance of using such a viscous flux discretization in the context of turbulent flows was recently
highlighted in Chamarthi et al. [28]. For simplicity and without loss of generality, we present a one-dimensional
scenario. The cell center numerical viscous flux derivative is:

dF̌𝑣 1  𝑣 𝑣

= F̌𝑖+ 1 − F̌𝑖− 1 , (31)
d𝜉 𝑖 Δ𝜉
2 2

The cell interface numerical viscous flux is:

© 0 ª
𝑣
F̌𝑖+ 1 =­
­ −𝜏𝑖+ 1 ®
®, (32)
2 ­ 2 ®
−𝜏 1 𝑢 1 + 𝑞 𝑖+ 1
« 𝑖+ 2 𝑖+ 2 2¬

where,


4 𝜕𝑢 𝜕𝑇
𝜏𝑖+ 1 = 𝜇ˆ 𝑖+ 1 , 𝑞 𝑖+ 1 = −𝜅ˆ𝑖+ 1 . (33a–33b)
2 3 2 𝜕𝑥 1
𝑖+
2 2 𝜕𝑥 𝑖+ 1
2 2

As mentioned by Eqn. 8, the dynamic viscosity was computed using Sutherland’s Law. Since the dynamic viscosity is
required at the cell interface, the cell interface temperature is used. The cell interface temperature was computed via
an arithmetic average, i.e. 𝑇𝑖+ 1 = (𝑇𝑖 + 𝑇𝑖+1 ) /2. For an arbitrary variable, 𝜙, the 𝛼-damping approach computes cell
2
interface gradients as:
 
𝜕𝜙 1 𝜕𝜙 𝜕𝜙 𝛼
= + + (𝜙 𝑅 − 𝜙 𝐿 ) , (34)
𝜕𝑥 𝑖+ 1 2 𝜕𝑥 𝑖 𝜕𝑥 𝑖+1 2Δ𝑥
2

where,


𝜕𝜙 Δ𝑥
𝜙 𝐿 = 𝜙𝑖 + + 𝛽 (𝜙𝑖+1 − 2𝜙𝑖 + 𝜙𝑖−1 ) , (35a)
𝜕𝑥 𝑖 2

𝜕𝜙 Δ𝑥
𝜙 𝑅 = 𝜙𝑖+1 − + 𝛽 (𝜙𝑖+2 − 2𝜙𝑖+1 + 𝜙𝑖 ) , (35b)
𝜕𝑥 𝑖+1 2

where, in this work, 𝛼 = 4 and 𝛽 = 0. The gradients at cell centers were the same ones computed in Eqn. 14.

8
C. Wall Model
In this work, the compressible equilibrium ODE wall model of Kawai and Larsson [21] was used to model the
unresolved, inner layer dynamics of the boundary layer. In this model, the flow is assumed to be steady, parallel, and in
equilibrium so that the convective and pressure gradient terms of the compressible RANS equation perfectly balance
each other. These assumptions result in the simplified momentum and total energy equations:
 
d  d |𝑢 𝑤𝑚 |
𝜇ˆ 𝑤𝑚 + 𝜇𝑡 ,𝑤𝑚 = 0, (36)
d𝑦 𝑤𝑚 d𝑦 𝑤𝑚
    
d 1 𝜇ˆ 𝑤𝑚 𝜇𝑡 ,𝑤𝑚 d  d |𝑢 𝑤𝑚 |
+ =− 𝜇ˆ 𝑤𝑚 + 𝜇𝑡 ,𝑤𝑚 |𝑢 𝑤𝑚 | , (37)
d𝑦 𝑤𝑚 𝛾 − 1 Pr Prt,wm d𝑦 𝑤𝑚 d𝑦 𝑤𝑚
√︁
where the “wm” subscript denotes that the quantity is related solely to the wall model and |𝑢 𝑤𝑚 | = 𝑢 2𝑤𝑚 + 𝑤 2𝑤𝑚 is the
magnitude of the wall parallel velocities. An eddy viscosity assumption is invoked and has the formulation:
√︄
|𝜏𝑤𝑚 | 𝑤
𝑦 𝑤𝑚 1 − exp −𝑦 ∗ /𝐴+ ,
  2
𝜇𝑡 ,𝑤𝑚 = K 𝜌 𝑤𝑚 (38)
𝜌 𝑤𝑚
where,
√︁
∗ 𝑦 𝑤𝑚 𝜌 𝑤𝑚 |𝜏𝑤𝑚 | 𝑤
𝑦 = . (39)
𝜇ˆ 𝑤𝑚
Note that the scaled wall distance in the van Driest damping function is in semi-local scaling, which has been shown
to be more favorable for non-adiabatic flows. The parameters were taken as K = 0.41, 𝐴+ = 17, and Pr𝑡 ,𝑤𝑚 = 0.9.
Note that the dynamic viscosity was computed via Sutherland’s law, and was scaled in accordance with the outer LES
non-dimensionalization method. Therefore, just as in the outer LES, 𝜇ˆ 𝑤𝑚 = 𝜇 𝑤𝑚 Ma/Re. Moreover, constant pressure
is assumed and the same ideal gas law is used as in the outer LES, i.e. 𝜌 𝑤𝑚 = 𝛾 𝑝/𝑇𝑤𝑚 , where 𝑝 is taken to be constant.
The equations are solved on one-dimensional finite-volume grids normal to each wall cell center. Each one-
dimensional grid begins at 𝑦 𝑤𝑚 = 0 and extend up to 𝑦 𝑤𝑚 = ℎ 𝑤𝑚 , where ℎ 𝑤𝑚 is the desired height of the wall model
grids. ℎ 𝑤𝑚 should be in the resolved logarithmic layer of the outer LES solution. The interface locations were obtained
from:
 
tanh (𝛼𝑠 𝑗/𝑁 𝑤𝑚 )
𝑦 𝑤𝑚 | 𝑗+ 1 = ℎ 𝑤𝑚 1 − , for 𝑗 ∈ [𝑁 𝑤𝑚 ...0], (40)
2 tanh (𝛼𝑠 )
where 𝛼𝑠 is a stretching parameter and 𝑁 𝑤𝑚 is the number of cells used for the wall model grids. This stretched grid
formulation has the advantage of placing the top cell interface at a height of 1. Therefore, to ensure that the wall model
grids extend up to ℎ 𝑤𝑚 , the cell interfaces are simply scaled accordingly. The cell centers were then computed from:
 
𝑦 𝑤𝑚 | 𝑗 = 𝑦 𝑤𝑚 | 𝑗+ 1 + 𝑦 𝑤𝑚 | 𝑗 − 1 /2, for 𝑗 ∈ [1...𝑁 𝑤𝑚 ]. (41)
2 2

Eqns. 36 and 37 were discretized using a second order central difference scheme in the interior, whereas at the
wall and matching location, first order forward and backward differences were used, respectively. This resulted in two
tridiagonal systems of equations for which the Thomas algorithm was employed. At the wall, an isothermal viscous wall
boundary condition was applied (i.e., |𝑢 𝑤𝑚 | 𝑤 = 0, 𝑇𝑤𝑚 | 𝑤 = 𝑇𝑤 ), whereas at ℎ 𝑤𝑚 , the boundary simply matched the
values of the LES at the matching location (i.e., |𝑢 𝑤𝑚 | ℎ𝑤𝑚 = |𝑢 𝐿𝐸𝑆 | ℎ𝑤𝑚 , 𝑇𝑤𝑚 | ℎ𝑤𝑚 = 𝑇𝐿𝐸𝑆 | ℎ𝑤𝑚 ).
Eqns. 36 and 37 were solved iteratively for the wall modelled velocity and temperature profiles at each one-
dimensional finite-volume grid until the differences of the successive values of the wall modelled stresses were less than
10−4 . The wall modelled wall shear stress and wall heat flux were computed from:

|𝑢 𝑤𝑚 | 1 − |𝑢 𝑤𝑚 | 𝑤 𝑇𝑤𝑚 | 1 − 𝑇𝑤𝑚 | 𝑤
𝜏𝑤𝑚 | 𝑤 = 𝜇ˆ 𝑤𝑚 | 𝑤 , 𝑞 𝑤𝑚 | 𝑤 = 𝜅ˆ𝑤𝑚 | 𝑤 . (42a-42b)
𝑦 𝑤𝑚 | 1 − 𝑦 𝑤𝑚 | 𝑤 𝑦 𝑤𝑚 | 1 − 𝑦 𝑤𝑚 | 𝑤

Once converged wall modelled stresses were obtained, the wall modelled shear stress was decomposed:

9
𝑢 𝑤𝑚 | ℎ𝑤𝑚 𝑤 𝑤𝑚 | ℎ𝑤𝑚
𝜏𝑤𝑚, 𝑥 𝑤 = 𝜏𝑤𝑚 | 𝑤 , 𝜏𝑤𝑚,𝑧 𝑤 = 𝜏𝑤𝑚 | 𝑤 , (43a-43b)
|𝑢 𝑤𝑚 | ℎ𝑤𝑚 |𝑢 𝑤𝑚 | ℎ𝑤𝑚

where 𝑢 𝑤𝑚 | ℎ𝑤𝑚 and 𝑤 𝑤𝑚 | ℎ𝑤𝑚 are the streamwise and spanwise velocities at the matching location, respectively. The
wall-normal viscous flux vector at the wall was then modified to:

0
© ª
­ 𝜂e𝑥 𝜏𝑥 𝑥 + 𝜂e𝑦 𝜏𝑤𝑚,𝑥 + 𝜂e𝑧 𝜏𝑥𝑧 ®
­ 𝑤 ®
­ ®
𝑣 ­𝜂e𝑥 𝜏𝑤𝑚,𝑥 + 𝜂e𝑦 𝜏𝑦𝑦 + 𝜂e𝑧 𝜏𝑤𝑚,𝑧 ®
G |𝑤 = − ­ 𝑤 𝑤® , (44)
­ ®
­ 𝜂e 𝜏 + 𝜂e 𝜏
𝑦 𝑤𝑚,𝑧 𝑤 + 𝜂 e𝑧 𝜏𝑧𝑧 ®®
­ 𝑥 𝑧𝑥

­ ®
« 𝜂
e 𝛽
𝑥 𝑥 + 𝜂
e 𝛽
𝑦 𝑦 + 𝜂
e 𝛽
𝑧 𝑧 ¬
where,

𝜕𝑇
𝛽 𝑥 = 𝑢𝜏𝑥 𝑥 + 𝑣 𝜏𝑤𝑚, 𝑥 𝑤 + 𝑤𝜏𝑥𝑧 + 𝜅ˆ , 𝛽 𝑦 = 𝑢 𝜏𝑤𝑚,𝑥 𝑤 + 𝑣𝜏𝑦𝑦 + 𝑤 𝜏𝑤𝑚,𝑧 𝑤 + 𝑞 𝑤𝑚 | 𝑤 , (45a-45b)
𝜕𝑥
𝜕𝑇
𝛽 𝑧 = 𝑢𝜏𝑧 𝑥 + 𝑣 𝜏𝑤𝑚,𝑧 𝑤 + 𝑤𝜏𝑧𝑧 + 𝜅ˆ . (45c)
𝜕𝑧

1. A Note on Modifications for the Wall Model


• Note that it is common to either use one-sided differences or to neglect altogether the non-wall modelled stresses
in the viscous flux vector. In this work, we neither modified nor neglected these stresses, so they can be considered
to be unresolved.
• For case 3 (Section IV.C), the wall model became highly unstable in the separation bubble region, which is to be
expected since this region is very much not in equilibrium. Thus, the viscous fluxes were only modified if the
local streamwise velocity at ℎ 𝑤𝑚 was greater than zero. Effectively, the wall model was not used in any regions
with streamwise local flow reversal.

D. Time Integration
The explicit third-order total-variation-diminishing Runge-Kutta (RK3TVD) [34] method was used for time
integration. The timestep, Δ𝑡, was computed from the CFL condition. We used both a convective and viscous analogue
of the CFL condition. For all simulations, CFL = 0.2. The convective Δ𝑡 was computed from:
"  #
√︃   √︃−1   √︃ −1 −1
Δ𝑡 𝑐 = min 𝑈 + 𝑐 𝜉 𝑥2 + 𝜉 𝑦2 + 𝜉 𝑧2 , 𝑉 + 𝑐 𝜂2𝑥 + 𝜂2𝑦 + 𝜂2𝑧 , 𝑊 + 𝑐 𝜁 𝑥2 + 𝜁 𝑦2 + 𝜁 𝑧2 , (46)

√︁
where 𝑐 = 𝛾 𝑝/𝜌 is the local speed of sound. The viscous Δ𝑡 was computed from:
  −1   −1   −1 
𝜇ˆ
Δ𝑡 𝑣 = min 𝜉 𝑥2 + 𝜉 𝑦2 + 𝜉 𝑧2 , 𝜂2𝑥 + 𝜂2𝑦 + 𝜂2𝑧 , 𝜁 𝑥2 + 𝜁 𝑦2 + 𝜁 𝑧2 , (47)
𝛼
where 𝛼 = 4 corresponds to that employed in the viscous spatial discretization method. Finally, the timestep was
computed from:

Δ𝑡 = CFL × min (Δ𝑡 𝑐 , Δ𝑡 𝑣 ) . (48)

E. Boundary Conditions
In the current solver, a type 2 grid consistent with Laney [35] (pages 430-431) is used. As such, boundary conditions
are implemented using ghost cells. Dirichlet boundary conditions were set according to:

10
𝜙𝐺𝐶 = 2𝜙 𝑤 − 𝜙 𝐼𝐶 , (49)
where 𝜙 is an arbitary variable and 𝐺𝐶, 𝑤, and 𝐼𝐶 are ghost cell, wall value, and interior cell, respectively. Zero
gradient Neumann boundary conditions were set according to:

𝜙𝐺𝐶 = 𝜙 𝐼𝐶 . (50)

IV. Results and Discussion

A. Case 1: Oblique Shock Impingement on Mach 6 Disturbed Laminar Boundary Layer

Fig. 1 Grid and relevant dimensions of case 1. Every third point is shown for clarity. The flat plate leading edge
is at 𝑥 = −46𝛿0∗ . Note the wall-normal grid clustering and streamwise glid clustering towards the leading edge.

Table 1 Non-dimensional parameters of case 1.

Ma∞ Re∞, 𝛿0∗ Re∞,𝑥0 Pr 𝛾


6.0 6830 314252 0.72 1.4

Table 2 Reference values of case 1.

𝑇∞ , K 𝑇𝑤 , K 𝜌∞ , kg m−3 𝜇∞ , Pa s 𝑢 ∞ , m s−1 𝑅𝑔𝑎𝑠 , J kg−1 K −1


65 292.5 0.0267 4.16 × 10−6 969.69 287.05

Table 3 Boundary conditions of case 1. RH refers to Rankine-Hugoniot jump conditions.

𝑖 𝑚𝑖𝑛 𝑖 𝑚𝑎𝑥 𝑗 𝑚𝑖𝑛 , 𝑥 ≤ −46𝛿0∗ 𝑗 𝑚𝑖𝑛 , 𝑥 > −46𝛿0∗ 𝑗 𝑚𝑎𝑥 𝑘 𝑚𝑖𝑛 , 𝑘 𝑚𝑎𝑥
Freestream Extrapolation Inviscid Wall Isothermal Viscous Wall RH Conditions Periodic

Table 4 Grid details of case 1.

Total Grid Size Grid Δ𝑦| 𝑤


1.3M 298 × 64 × 72 0.255𝛿0∗
5.5M 596 × 64 × 144 0.255𝛿0∗

The first case considered was oblique shock impingement on a Mach 6 disturbed laminar boundary layer. This case
was studied experimentally [6], [7], experimentally and with DNS [5], as well as with WMLES [18], [19], [20]. In
this flow system, an oblique shock caused by a 4◦ wedge atop the domain is made to impinge on a disturbed laminar
boundary layer. The non-dimensional parameters for this case may be found in Table 1. Note that the Mach number is

11
based on the freestream velocity and temperature. The Reynolds number is based on the displacement thickness at a
distance from the flat plate leading edge, 𝑅𝑒 𝑥 = 314252 or 𝑥 = 46𝛿0∗ . The reference values and boundary conditions
for this case can be found in Tables 2 and 3, respectively. Note that it is conventional to simulate this case using the
compressible similarity solution as the inflow condition, avoiding the need to simulate the flat plate leading edge.
However, in this work, we include the flat plate leading edge. To avoid any potential instabilities from the leading edge
shock, the region immediately preceding the leading edge (−50𝛿0∗ ≤ 𝑥 ≤ −46𝛿0∗ ) is treated as an inviscid wall to avoid
artificial Mach waves at the domain inlet. Moreover, the MEG scheme of Chamarthi [25], is employed from the domain
inlet until 𝑥 = 0𝛿0∗ (refer to Fig. 1) to better dissipate any instabilities caused by the leading edge.
The oblique shock impingement on the boundary layer is insufficient to cause boundary layer transition alone. This
was noted in the original DNS study of Sandham et al. [5]. However, in experiments of the same case, boundary layer
transition was observed. This is believed to be a result of freestream disturbances in addition to instabilities caused by
the shock boundary layer interaction. As such, in the DNS study, freestream disturbances were added to the density at
𝑥 = 0𝛿0∗ in an attempt to characterize the freestream disturbances present in the experiments. These disturbances are of
the form:
𝐽
∑︁ 𝐾
 ∑︁
𝜌 ′ = 𝐴𝑊 (𝑦) cos 2𝜋 𝑗 𝑧/𝐿 𝑧 + 𝜙 𝑗 sin (2𝜋 𝑓 𝑘 𝑡 + 𝜓 𝑘 ) , (51)
𝑗=1 𝑘=1

where 𝑊 (𝑦) = 1 − exp −𝑦 3 is a window function to dampen the disturbances in the boundary layer, 𝐽 = 16 and 𝐾 = 20
are cutoff wavenumbers, 𝑓 𝑘 = 0.02𝑘 is the frequency, and 𝜙 𝑗 and 𝜓 𝑘 are random phases in [0, 2𝜋]. The disturbance
amplitude, 𝐴, is tuned based on the grid resolution considered. For example, in the DNS of Sandham et al. [5], using
approximately 200 million cells, 𝐴 = 0.0005. Whereas in the subsequent WMLES studies of Yang et al. [18], Mettu
and Subbareddy [19], and Ganju et al. [20], using either one million cells or four millions cells, 𝐴 = 0.001. Thus for
this study, we took 𝐴 = 0.001.
The domain size for this case is 350𝛿0∗ × 25𝛿0∗ × 45𝛿0∗ . The grid sizes considered may be found in Table 4. For both
grid sizes, Δ𝑦| 𝑤 = 0.255𝛿0∗ . The wall model exchange location was placed at the third grid cell center off the wall,
which was at ℎ 𝑤𝑚 = 0.64𝛿0∗ . This placed ℎ 𝑤𝑚 at an instantaneous 𝑦 +𝑤𝑚 ℎ𝑤𝑚 ∼ 100. Moreover, we used 𝑁 𝑤𝑚 = 30,

with 𝛼𝑠 = 3.0. This placed the instantaneous 𝑦 +𝑤𝑚 𝑤 < 1.
Since transitional and turbulent cases are extremely sensitive to excessive numerical dissipation, the Ducros cutoff
value, Ω, was an important factor in the accuracy of this case. While this is obviously not ideal, it is currently a
necessary evil for similar configurations [22], [24]. This is because a very low dissipation or kinetic energy/entropy
preserving scheme is critical to resolve instabilities and turbulent structures while effective shock capturing schemes
are necessary to mitigate any stability-jeopardizing oscillations. For this work, Ω = 0.1 was found to be sufficient to
effectively track shocks while diminishing any excess dissipation from the application of the limiter. Moreover, to reduce
numerical dissipation further, we averaged the left- and right-biased states to obtain a central scheme and effectively rid
the interpolation of numerical dissipation:

𝐿 𝑅 1 𝐿 𝑅

Č𝑖+ 1
,𝑏
= Č𝑖+ 1
,𝑏
=Č𝑖+ 1 ,𝑏 + Č𝑖+ 1 . (52)
2 2 2 2 2 ,𝑏

Fig. 2 displays various instantaneous flow figures taken from the 5.5 million cell grid simulation. In Fig. 2a,
the density gradient magnitude shows the shock structure, separation bubble, and laminar and turbulent boundary
layer. As aforementioned, the leading edge shock is present near the inlet of the domain. Downstream, the leading
edge shock passes through the impinging shock emanating from the top boundary. Slightly before this shock-shock
interaction, a faint footprint of the density disturbances is evident. Near the shock boundary layer interaction, a
separation bubble is apparent from the heightening of the incoming, disturbed, laminar boundary layer, which further
downstream reattaches to the wall. From this separation bubble, a separation shock occurs, as well as an expansion fan
at the maximum height of the bubble, and finally a strong reattachment shock. This separation bubble and interaction
region introduce added instabilities to the flow system, which cause the boundary layer to transition to turbulence just
downstream of the interaction. This transition to turbulence is evident in the flow structures also plotted in Fig. 2a.
These structures are instantaneous iso-surfaces of the second invariant of the velocity gradient tensor, otherwise known
as the Q-criterion. The structures are colored by wall-normal distance to highlight their growth with streamwise distance
from the interaction region. The complete transition to turbulence is preceded by the emergence of Görtler vortices
which arise from local streamline concavity caused by the separation bubble [36].

12
(a) 𝑥-𝑦 slice of density gradient magnitude contour and iso-surfaces of Q-criterion (𝑄 = 0.1) colored by wall-normal distance.

(b) Stanton number contour.

(c) Near-wall (𝑦 = 0.1275𝛿0∗ ) temperature contour.

Fig. 2 Instantaneous flow figures from the 5.5 million cell grid of case 1.

Transition to turbulence is accompanied by intense near-wall heating. This is represented by the Stanton number in
Fig. 2b, wherein the transition process and turbulence cause high heating loads to the wall, owing to large aerodynamic
heating common to hypersonic flows. The Stanton number was computed from:

𝑞 𝑤𝑚 | 𝑤
St = , (53)
𝜌∞ 𝑢 ∞ cp (𝑇𝑟 − 𝑇𝑤 )
 
where 𝑇𝑟 = 1 + 𝑟 (𝛾 − 1) Ma2∞ /2 is the recovery temperature, 𝑟 = Pr1/2 = 0.83 is the recovery factor, and 𝜌∞ = 1,
𝑢 ∞ = Ma, and cp = (𝛾 − 1) −1 due to non-dimensionalization. Slightly above the wall, the temperature can be observed
to decrease post-transition, as is shown in Fig. 2c. This is because the height of the slice, 𝑦 = 0.1275𝛿0∗ , is above the
large near-wall temperature gradient attributed to aerodynamic heating. Despite this being the first cell center off the
wall, the one-dimensional wall model grids, which extend from the wall to the third cell center off the wall, capture
these large temperature gradients. Therefore, the correct wall heat transfer is computed.
Fig. 3 displays the time and spanwise averaged Stanton number as a function of Re 𝑥 , which was computed from:

Re 𝑥 = 𝑥Re∞, 𝛿0∗ + Re 𝑥0 . (54)


A quantitative comparison was made to available experimental, DNS, and WMLES data. All WMLES data fall between
the two experiments, which were conducted at separate facilities, while matching the DNS data satisfactorily. Since
WMLES for this case requires the tuning of the density disturbance amplitude, increased resolution may even cause
over-prediction of the Stanton number, relative to DNS. Moreover, most of these studies, including the present work,
make use of a hybrid scheme by employment of a shock sensor. Therefore, the choice of the shock sensor cutoff value
surely plays a large role in quantitative accuracy here.

13
Fig. 3 Time and spanwise averaged Stanton number compared with available experimental, DNS, and WMLES
data.

B. Case 2: Hypersonic Transitional Flow Over a 15◦ Compression Ramp

Fig. 4 Grid and relevant dimensions of case 2. Every tenth point is shown for clarity. The flat plate leading edge
is at 𝑥 = 0𝐿. The grid is lightly clustered in the wall-normal direction and in the streamwise direction near the
leading edge, corner, and outlet. Care was taken to ensure that near wall cells were orthogonal, however this is
not depicted here since every tenth point is shown.

Table 5 Non-dimensional parameters of case 2.

Ma∞ Re∞,𝐿 Pr 𝛾
7.7 4.2 × 105 0.71 1.4

The second case considered was hypersonic transitional flow over a 15◦ compression ramp. The geometry consists
of a flat plate with a sharp leading edge followed by a ramp of angle 15◦ and stems from the experimental work of
Roghelia et al. [8, 9], which was later investigated extensively using DNS and global stability analysis (GSA) by Cao et

14
Table 6 Reference values of case 2.

𝑇∞ , K 𝑇𝑤 , K 𝜌∞ , kg m−3 𝜇∞ , Pa s 𝑢 ∞ , m s−1 𝑅𝑔𝑎𝑠 , J kg−1 K −1


125 293 0.0212 8.7 × 10−6 1726 287.05

Table 7 Boundary conditions of case 2.

𝑖 𝑚𝑖𝑛 𝑖 𝑚𝑎𝑥 𝑗 𝑚𝑖𝑛 , 𝑥 ≤ 0𝐿 𝑗 𝑚𝑖𝑛 , 𝑥 > 0𝐿 𝑗 𝑚𝑎𝑥 𝑘 𝑚𝑖𝑛 , 𝑘 𝑚𝑎𝑥


Freestream Extrapolation Inviscid Wall Isothermal Viscous Wall Freestream Periodic

Table 8 Grid details of case 2.

Total Grid Size Grid Δ𝑦| 𝑤


16.3M 817 × 200 × 100 0.001𝐿

al. [10–12] and Hao et al. [14]. The non-dimensional parameters, reference values, boundary conditions, and grid
details for this case may be found in Tables 5, 6, 7, and 8, respectively. Note that the Mach number is based on the
freestream velocity and temperature. The Reynolds number is based on the flat plate length of 𝐿 = 0.1 m.

(a) Wall pressure coefficient comparison. (b) Skin friction comparison.

Fig. 5 Comparison of wall pressure coefficient and skin friction with the fine grid two-dimensional results of
Hao et al. [14].

In this case, a Mach 7.7 freestream flows over the geometry causing a leading edge shock and an adverse pressure
gradient by the corner. This adverse pressure gradient is large enough to cause a separation bubble, which then brings
about a separation shock, expansion fan, and reattachment shock. The various shock-shock interactions cause shear
layers to emerge as well.
A two-dimensional simulation was first run to establish a quasi-steady mean flow that was then used as an initial
condition for the subsequent three-dimensional simulation. A grid of 817 × 200 was used for the two-dimensional
simulation. This grid count was the coarsest simulation considered in Hao et al. [14]. The pressure coefficient and skin
friction coefficient are plotted in Figs. 5a and 5b, respectively. These quantities were respectively computed from:

15
2 𝑝| 𝑤 2 𝜏| 𝑤
Cp = , Cf = . (55a-55b)
𝜌∞ 𝑢 2∞ 𝜌∞ 𝑢 2∞

Eqns. 55 are only used for this two-dimensional case. All other instances of Cp were computed with Eqn. 57b. A
comparison is made to the finest grid simulation in Hao et al. [14]. The pressure coefficient matches excellently.
However, it is clear that on the downstream portion of the ramp where scales become smaller, the near-wall spacing was
insufficient to accurately capture the skin friction coefficient. Similarly, the negative skin friction coefficient in the
separation bubble is slightly under-predicted. The discrepancy is majorly related to the large difference in wall-normal
grid spacing between the present simulation (𝐿 × 10−3 ) and that of Hao et al. [14] (𝐿 × 10−6 ). Future work must address
this.
Despite the slight discrepancy in the two-dimensional mean flow simulation, the three-dimensional simulation was
still conducted with the mean flow initial condition. The spanwise length was 𝐿 𝑧 = 0.3𝐿. While the system is unstable,
to bring about three-dimensionality more rapidly, disturbances were added into the domain. These are in the form:

𝑤 ′𝑗 = 0.01R, for 𝑗 ∈ [1...20] and 𝑥/𝐿 ∈ (0.5, 1.5) , (56)


where 𝑤′ is a perturbation for the spanwise velocity and R is a random number in [−1, 1]. The intrinsic flow system
and this disturbance bring about a three-dimensional flow field that does not fully transition to turbulence but showcases
various transitional flow phenomena. This is because the Reynolds number is slightly inadequate and the domain length
is not long enough to allow for instabilities to grow sufficiently large for complete transition to turbulence. It is important
to note that the wall model was not used for this case since the flow does not become fully turbulent.

(a) 𝑥-𝑦 slice of density gradient magnitude contour, 𝑥-𝑧 slice of near-wall Mach number, and
iso-surfaces of Q-criterion (𝑄 = 0.1).

(b) Stanton number contour. White lines denote lines of zero skin friction coefficient.

Fig. 6 Instantaneous flow figures for case 2.

Fig. 6 displays instantaneous flow figures of case 2. The shock structure, separation bubble, and laminar and
transitional boundary layer can be seen in the density gradient magnitude slice. As above-mentioned, the ramp angle

16
coupled with the freestream and wall conditions causes a sufficiently large adverse pressure gradient to occur, resulting
in a separation bubble and the associated separation shock, expansion fan (very faint in the figure), and reattachment
shock. These additional shocks then interact with each other and with the leading edge shock and form shear layers,
expansion fans, and reflections. The three-dimensionality of the flow reattachment location can be seen in the near-wall
slice of the Mach number. Near-wall streaks are also present post-reattachment, owing to baroclinic effects [37] and the
Görtler instability caused by local flow concavity. Secondary instabilities are depicted with Q-criterion structures that
coincide with the location of the near-wall streaks. The largest streak wavelengths are approximately 4 mm, which are
consistent with the DNS and experimental streak wavelengths of 4 ∼ 6mm. These streaks are associated with increased
wall heating, as is shown in Fig. 6b, in which the Stanton number is plotted. White lines denoting regions of zero skin
friction are also plotted in Fig. 6b, which exhibit the length of the separation bubble along with the three-dimensionality
of the reattachment location along the span of the wall.

Fig. 7 Time and spanwise averaged Stanton number and wall pressure coefficient compared with the experiments
of Roghelia et al. [8, 9] and the DNS of Cao et al. [11].

A quantitative comparison is also made with the experiments of Roghelia et al. [8, 9] and the DNS of Cao et al.
[11]. Fig. 7 displays the time and spanwise averaged Stanton number and wall pressure coefficient. Note that these
quantities are computed slightly differently than previously shown:


𝑞| 𝑤 2 𝑝| 𝑤 − 𝑝 ∞
St = , Cp = , (57a-57b)
𝜌∞ 𝑢 ∞ cp (𝑇𝑟 − 𝑇𝑤 ) 𝜌∞ 𝑢 2∞

where for the Stanton number, the ILES value of 𝑞| 𝑤 is used in lieu of the wall modelled value, and for the pressure
coefficient, the freestream pressure is subtracted from the wall pressure value.
Regarding the pressure coefficient, there is excellent agreement with the DNS and satisfactory agreement with the
experiment. Cao et al. [10] noted that the discrepancy between experiment and DNS in the region 𝑥/𝐿 < 0.5 can be
attributed to a lack of resolution in the pressure transducers in the experiments. The present Stanton number profile
matches the DNS prediction quite well with the exception of a small region of the separation bubble and near the outlet
of the domain. Moreover, in the present simulations, the separation bubble distance is 0.53 ≤ 𝑥/𝐿 ≤ 1.27, which is
close to the DNS and experiment separation bubble length of 0.59 ≤ 𝑥/𝐿 ≤ 1.26 and adequate for the present grid
resolution. These discrepancies can also be attributed to the coarse near-wall spacing, which will be addressed in future
work.

17
C. Case 3: Hypersonic Transition to Turbulence Over a 15◦ Compression Ramp

Fig. 8 Grid and relevant dimensions of case 3. Every tenth point is shown for clarity. The flat plate leading edge
is at 𝑥 = 0𝐿. The grid is lightly clustered in the wall-normal direction and in the streamwise direction near the
leading edge, corner, and outlet. Care was taken to ensure that near wall cells were orthogonal, however this is
not depicted here since every tenth point is shown.

Table 9 Non-dimensional parameters of case 3.

Ma∞ Re∞,𝐿 Pr 𝛾
7.7 8.6 × 105 0.71 1.4

Table 10 Reference values of case 3.

𝑇∞ , K 𝑇𝑤 , K 𝜌∞ , kg m−3 𝜇∞ , Pa s 𝑢 ∞ , m s−1 𝑅𝑔𝑎𝑠 , J kg−1 K −1


125 293 0.0432 8.7 × 10−6 1726 287.05

Table 11 Boundary conditions of case 3.

𝑖 𝑚𝑖𝑛 𝑖 𝑚𝑎𝑥 𝑗 𝑚𝑖𝑛 , 𝑥 ≤ 0𝐿 𝑗 𝑚𝑖𝑛 , 𝑥 > 0𝐿 𝑗 𝑚𝑎𝑥 𝑘 𝑚𝑖𝑛 , 𝑘 𝑚𝑎𝑥


Freestream Extrapolation Inviscid Wall Isothermal Viscous Wall Freestream Periodic

Table 12 Grid details of case 3.

Total Grid Size Grid Δ𝑦| 𝑤


25.5M 1297 × 200 × 100 0.001𝐿

The third and final case considered was hypersonic flow over a 15◦ compression ramp featuring full transition to
turbulence. In this case, the ramp length was increased to 0.22 m and the Reynolds number based on the flat plate
length of 𝐿 = 0.1 m was increased to 8.6 × 105 . This configuration was investigated using DNS in Cao et al. [13]. The
non-dimensional parameters, reference values, boundary conditions, and grid details for this case may be found in
Tables 9, 10, 11, and 12, respectively. The flow phenomena are similar to case 2. However, shortly downstream of flow
reattachment, the boundary layer fully transitions to turbulence.
Like case 2, a two-dimensional simulation was run to bring about a quasi-steady mean flow that was used as an initial
condition for the three-dimensional simulation. A grid of 1297 × 200 was used for the two-dimensional simulation. The

18
Fig. 9 Comparison of Stanton number with the two-dimensional results of Cao et al. [13].

Stanton number compared with the two-dimensional computation of Cao et al. [13] is shown in Fig. 9. Good agreement
is shown with the exception of a slight over-prediction of the peak Stanton number by the corner as well as a departure
from the reference towards the downstream portion of the ramp. It was thought that this discrepancy would not be
present in the three-dimensional simulation as a wall model would be used. As such, the two-dimensional quasi-steady
mean flow was still used for the subsequent three-dimensional simulation. For this case, disturbances were not added
since the system is intrinsically unstable and transition to turbulence occurs post-reattachment solely by numerical error.
The spanwise length was 𝐿 𝑧 = 0.3𝐿.
Fig. 10 displays instantaneous flow figures of case 3. In Fig. 10a, a similar shock structure to case 2 is apparent.
Contrary to case 2 however, the boundary layer transitions to turbulence shortly downstream of reattachment. In fact,
the instantaneous temperature-colored Q-criterion iso-surfaces seem to indicate the existence of near wall streaks even
within the end of the separation bubble [37]. The turbulent structures begin to grow in size downstream of reattachment,
however they dissipate farther downstream. This is consistent with the near wall temperature farther downstream of
the ramp: the structures are both smaller and of higher temperature, which indicates the decay of the structures rather
than their growth. This is consistent with the Stanton number, shown in Fig. 10b: the heat transfer largely increased in
1.5 ≲ 𝑥/𝐿 ≲ 2, but the wall heating quickly diminishes after this region. This decay of turbulence does not match the
phenomena seen in the reference DNS and is likely due to an inadequate mesh rather than a shortcoming of the wall
model. A finer grid case must be run in order to assess this and is left for future work.
Fig. 11 shows the time and spanwise averaged Stanton number and pressure coefficient compared with the data
from Cao et al. [13]. Regarding the pressure coefficient, the separation bubble region is under-predicted compared
to both the DNS and experiment. Moreover, the separation location is brought slightly upstream. At reattachment,
the pressure coefficient is then over-predicted and remains that way for the remainder of the ramp. Observing the
Stanton number, the separation bubble is badly represented by the present WMLES. An apparent reversed trend is
observed near the reattachment location, which is most probably attributed to the strong non-equilibrium effects in
that region. Besides for this, this reversed trend region is near a secondary separation bubble within the primary
separation bubble, which the wall model misrepresents. This may explain the apparent trend reversal, as well. Upon
full reattachment, the present WMLES matches the experimental Stanton number quite well, but then dwindles with
increasing streamwise distance from reattachment. This is consistent with Fig. 10b, which shows the largest values of
the Stanton number post-reattachment, and then a sudden tapering-off downstream. The tapering-off of the Stanton
number on the downstream portion of the ramp and the over-prediction of the pressure coefficient in the same region are
most likely due to an inadequate mesh, which must be addressed in future work. However, the remainder of the errors
such as, early separation, misrepresentation of the separation bubble, and an over-prediction of the pressure coefficient
at reattachment are surely due to the violation of the equilibrium assumption in and near the separation bubble. Dawson
et al. [38] showed that the inclusion of the pressure gradient term and it’s balancing convection term in a wall model are

19
(a) 𝑥-𝑦 slice of density gradient magnitude contour and iso-surfaces of Q-criterion (𝑄 = 0.1) colored by
temperature.

(b) Stanton number contour. White lines denote lines of zero skin friction coefficient.

Fig. 10 Instantaneous flow figures for case 2.

key to the accurate representation of the adverse pressure gradient region in compression ramp flows. As such, this is
likely the explanation of these errors in the present WMLES.

V. Conclusion
In this work, ILES and WMLES were employed to explore three cases of transitional hypersonic flows with shocks:
1) oblique shock impingement on a Mach 6 disturbed laminar boundary layer, 2) Mach 7.7 transitional flow over a 15◦
compression ramp, and 3) Mach 7.7 transition to turbulence over a 15◦ compression ramp. The importance of the
applied numerical method was emphasized and thus in this work, we used the gradient-based-reconstruction method of
Chamarthi [25] with MP limiting and a novel discontinuity sensor [26] to ensure very low numerical dissipation and
effectively the ability to simulate transition to turbulence on coarse grids. Moreover, the 𝛼-damping scheme [27, 28, 33]
was used for the viscous terms. The compressible equilibrium ODE wall model of Kawai and Larsson [21] was employed
to model the inner layer dynamics of the turbulent boundary layer and reduce computational cost. For the first case
of oblique shock impingement on a Mach 6 transitional boundary layer, good qualitative and quantitative agreement
was shown for the present WMLES against the experimental, DNS, and past WMLES of the case. Near-wall velocity,
temperature, and heat transfer streaks were noted to precede the fully turbulent region. The system shock structure was
well captured with very few high frequency oscillations. For the second case of Mach 7.7 transitional flow over a 15◦
compression ramp, satisfactory qualitative and quantitative agreement with the reference experiment and DNS was
noted. ILES was used for this case as the boundary layer does not fully transition to turbulence. Some quantitative
measures such as the the separation bubble length and the Stanton number in the separation bubble region showed a
slight discrepancy with the reference DNS. This is likely due to an under-resolved grid and will be investigated further in
future work. However, streak wavelengths matched the reference DNS quite well along with the wall pressure coefficient
and the Stanton number in the majority of the domain. For the third case of Mach 7.7 transition to turbulence over a 15◦
compression ramp, WMLES was employed. The grid appeared to be under-resolved even for WMLES standards and
thus wall heating was under-predicted in the downstream portion of the ramp, whereas the pressure coefficient was

20
Fig. 11 Time and spanwise averaged Stanton number and wall pressure coefficient compared with data from
Cao et al. [13].

over-predicted. The strong non-equilibrium effects at the point of separation, the separation bubble, and at reattachment
were also noted to be misrepresented by the equilibrium assuming wall model. As noted in previous WMLES studies
employing equilibrium wall models for non-equilibrium flows, the neglected convective and pressure gradient terms are
essential to represent the non-equilibrium effects in and near a separation bubble. These are surely the dominant terms
in regions of non-equilibrium effects. While the ODE nature of the compressible equilibrium wall model is desirable
from a computational standpoint, it simply lacks the necessary physical terms to appropriately represent the nature of
the flow in and near the separation bubble. Future work will further address grid sensitivity studies to satisfy WMLES
grid standards, non-equilibrium wall models or improvements to the present equilibrium wall model, and a parametric
study on the effect of wall temperature on cases 2 and 3 to observe the effect of cold-wall to hot-wall conditions on the
overall flow.

Acknowledgements
The authors gratefully acknowledge the financial support from the Technion-Israel Institute of Technology in Haifa,
Israel.

References
[1] Leyva, I. A., “The relentless pursuit of hypersonic flight,” PhT, Vol. 70, No. 11, 2017, pp. 30–36.

[2] on the National Aerospace Plane, U. S. D. S. B. T. F., Report of the Defense Science Board Task Force on the National Aerospace
Plane (NASP)., Office of the Under Secretary of Defense for Acquisition, 1988.

[3] Babinsky, H., and Harvey, J. K., Shock wave-boundary-layer interactions, Vol. 32, Cambridge University Press, 2011.

[4] Simeonides, G., and Haase, W., “Experimental and computational investigations of hypersonic flow about compression ramps,”
Journal of Fluid Mechanics, Vol. 283, 1995, pp. 17–42.

[5] Sandham, N., Schuelein, E., Wagner, A., Willems, S., and Steelant, J., “Transitional shock-wave/boundary-layer interactions in
hypersonic flow,” Journal of Fluid Mechanics, Vol. 752, 2014, pp. 1–33.

21
[6] Schülein, E., “Effects of laminar-turbulent transition on the shock-wave/boundary-layer interaction,” 44th AIAA Fluid Dynamics
Conference, 2014, p. 3332.

[7] Willems, S., Gülhan, A., and Steelant, J., “Experiments on the effect of laminar–turbulent transition on the SWBLI in H2K at
Mach 6,” Experiments in Fluids, Vol. 56, No. 3, 2015, pp. 1–19.

[8] Roghelia, A., Olivier, H., Egorov, I., and Chuvakhov, P., “Experimental investigation of Görtler vortices in hypersonic ramp
flows,” Experiments in Fluids, Vol. 58, No. 10, 2017, pp. 1–15.

[9] Roghelia, A., Chuvakhov, P. V., Olivier, H., and Egorov, I., “Experimental investigation of Görtler vortices in hypersonic ramp
flows behind sharp and blunt leading edges,” 47th AIAA fluid dynamics conference, 2017, p. 3463.

[10] Cao, S., Klioutchnikov, I., and Olivier, H., “Görtler vortices in hypersonic flow on compression ramps,” AIAA Journal, Vol. 57,
No. 9, 2019, pp. 3874–3884.

[11] Cao, S., Hao, J., Klioutchnikov, I., Olivier, H., and Wen, C.-Y., “Unsteady effects in a hypersonic compression ramp flow with
laminar separation,” Journal of Fluid Mechanics, Vol. 912, 2021.

[12] Cao, S., Hao, J., Klioutchnikov, I., Olivier, H., Heufer, K. A., and Wen, C.-Y., “Leading-edge bluntness effects on hypersonic
three-dimensional flows over a compression ramp,” Journal of Fluid Mechanics, Vol. 923, 2021.

[13] Cao, S., Hao, J., Klioutchnikov, I., Wen, C.-Y., Olivier, H., and Heufer, K. A., “Transition to turbulence in hypersonic flow over
a compression ramp due to intrinsic instability,” Journal of Fluid Mechanics, Vol. 941, 2022.

[14] Hao, J., Cao, S., Wen, C.-Y., and Olivier, H., “Occurrence of global instability in hypersonic compression corner flow,” Journal
of Fluid Mechanics, Vol. 919, 2021.

[15] Yang, X. I., and Griffin, K. P., “Grid-point and time-step requirements for direct numerical simulation and large-eddy simulation,”
Physics of Fluids, Vol. 33, No. 1, 2021, p. 015108.

[16] Choi, H., and Moin, P., “Grid-point requirements for large eddy simulation: Chapman’s estimates revisited,” Physics of fluids,
Vol. 24, No. 1, 2012, p. 011702.

[17] Larsson, J., Kawai, S., Bodart, J., and Bermejo-Moreno, I., “Large eddy simulation with modeled wall-stress: recent progress
and future directions,” Mechanical Engineering Reviews, Vol. 3, No. 1, 2016, pp. 15–00418.

[18] Yang, X., Urzay, J., Bose, S., and Moin, P., “Aerodynamic heating in wall-modeled large-eddy simulation of high-speed flows,”
AIAA journal, Vol. 56, No. 2, 2018, pp. 731–742.

[19] Mettu, B. R., and Subbareddy, P. K., “Wall modeled LES of compressible flows at non-equilibrium conditions,” 2018 Fluid
Dynamics Conference, 2018, p. 3405.

[20] Ganju, S., van Noordt, W., and Brehm, C., “Progress in the Development of an Immersed Boundary Viscous-Wall Model for 3D
and High-SpeedFlows,” AIAA Scitech 2021 Forum, 2021, p. 0160.

[21] Kawai, S., and Larsson, J., “Wall-modeling in large eddy simulation: Length scales, grid resolution, and accuracy,” Physics of
Fluids, Vol. 24, No. 1, 2012, p. 015105.

[22] De Vanna, F., Baldan, G., Picano, F., and Benini, E., “Effect of convective schemes in wall-resolved and wall-modeled LES of
compressible wall turbulence,” Computers & Fluids, 2022, p. 105710.

[23] Mettu, B. R., and Subbareddy, P. K., “Wall-Modeled Large Eddy Simulation of High Speed Flows,” AIAA Journal, 2022, pp.
1–23.

[24] van Noordt, W., Ganju, S., and Brehm, C., “An immersed boundary method for wall-modeled large-eddy simulation of turbulent
high-Mach-number flows,” Journal of Computational Physics, Vol. 470, 2022, p. 111583.

[25] Chamarthi, A. S., “Gradient based reconstruction: Inviscid and viscous flux discretizations, shock capturing, and its
application to single and multicomponent flows,” Computers & Fluids, Vol. 250, 2023, p. 105706. https://doi.org/https:
//doi.org/10.1016/j.compfluid.2022.105706, URL https://www.sciencedirect.com/science/article/pii/S0045793022002997.

[26] Chamarthi, A. S., Hoffmann, N., and Frankel, S. H., “A Characteristic Based Discontinuity Detector for Gradient Based
Reconstruction Methods in Compressible Flows,” Submitted, 2022.

22
[27] Nishikawa, H., “Two ways to extend diffusion schemes to navier-stokes schemes: Gradient formula or upwind flux,” 20th AIAA
Computational Fluid Dynamics Conference 2011, 2011, pp. 27–30.

[28] Chamarthi, A. S., K., H. C., Hoffmann, N., Bokor, S., and Frankel, S. H., “On the role of spectral properties of viscous flux
discretization for flow simulations on marginally resolved grids,” Computers & Fluids, 2022, p. 105742. https://doi.org/https:
//doi.org/10.1016/j.compfluid.2022.105742, URL https://www.sciencedirect.com/science/article/pii/S0045793022003346.

[29] Nonomura, T., Iizuka, N., and Fujii, K., “Freestream and vortex preservation properties of high-order WENO and WCNS on
curvilinear grids,” Computers & Fluids, Vol. 39, No. 2, 2010, pp. 197–214.

[30] Visbal, M. R., and Gaitonde, D. V., “On the use of higher-order finite-difference schemes on curvilinear and deforming meshes,”
Journal of Computational Physics, Vol. 181, No. 1, 2002, pp. 155–185.

[31] Blazek, J., Computational fluid dynamics: principles and applications, Butterworth-Heinemann, 2015.

[32] Pathak, H. S., and Shukla, R. K., “Adaptive finite-volume WENO schemes on dynamically redistributed grids for compressible
Euler equations,” Journal of Computational Physics, Vol. 319, 2016, pp. 200–230.

[33] Chamarthi, A. S., Bokor, S., and Frankel, S. H., “On the Importance of High-Frequency Damping in High-Order Conservative
Finite-Difference Schemes for Viscous Fluxes,” Journal of Computational Physics, 2022, p. 111195.

[34] Gottlieb, S., and Shu, C.-W., “Total variation diminishing Runge-Kutta schemes,” Mathematics of computation, Vol. 67, No.
221, 1998, pp. 73–85.

[35] Laney, C. B., Computational gasdynamics, Cambridge university press, 1998.

[36] Fu, L., Karp, M., Bose, S. T., Moin, P., and Urzay, J., “Shock-induced heating and transition to turbulence in a hypersonic
boundary layer,” Journal of Fluid Mechanics, Vol. 909, 2021.

[37] Dwivedi, A., Sidharth, G., Nichols, J. W., Candler, G. V., and Jovanović, M. R., “Reattachment streaks in hypersonic compression
ramp flow: an input–output analysis,” Journal of Fluid Mechanics, Vol. 880, 2019, pp. 113–135.

[38] Dawson, D. M., Lele, S. K., and Bodart, J., “Assessment of wall-modeled large eddy simulation for supersonic compression
ramp flows,” 49th AIAA/ASME/SAE/ASEE Joint PropulsionConference, 2013, p. 3638.

23

You might also like