International Journal of Multiphase Flow 116 (2019) 250–269

Contents lists available at ScienceDirect

International Journal of Multiphase Flow

journal homepage: www.elsevier.com/locate/ijmulflow

Single vapour bubble growth under flash boiling conditions using a

modified HLLC Riemann solver
D. Dietzel a, T. Hitz b, C.-D. Munz b, A. Kronenburg a,∗
a
Institut für Technische Verbrennung, Universität Stuttgart, Herdweg 51, 70174 Stuttgart
b
Institut für Aerodynamik und Gasdynamik, Universität Stuttgart, Pfaffenwaldring 21, 70569 Stuttgart

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

Article history: Common fuel-oxidizer combinations in orbital manoeuvring systems consist of toxic substances but will
Received 18 June 2018 be replaced in the future. Liquid oxygen (LOX) is one potential oxidizer but it rapidly superheats under
Revised 28 February 2019
the initial low-pressure conditions in the combustion chamber. Bubble nucleation and growth dominate
Accepted 12 April 2019
the efficient disintegration of the liquid jet, which is called flash boiling. An investigation of the small
Available online 17 April 2019
scale bubble dynamics will help to improve existing models for the break-up of the LOX jet and the mix-
Keywords: ing with the fuel. Direct numerical simulations (DNS) can be used to analyse the underlying processes if
Single vapour bubble the solver handles phase transition effects at extreme ambient conditions. In this paper we use a fully
Riemann solver compressible discontinuous Galerkin solver combined to a level-set equation and to a modified HLLC
Level-set Riemann solver. The main goal of our investigation is to assess closure conditions for the mass transfer
Multiphase flow models which accurately represent the physics of vapour bubble growth. We couple three models for the
Phase change
estimation of vaporisation mass fluxes to the interface Riemann solver. The Hertz-Knudsen relation and a
kinetic relation predict the volumetric expansion well but cannot represent the instantaneous mass flux.
A sub-grid scale heat flux model predicts the mass flux qualitatively better but the volumetric bubble ex-
pansion matches only at late time intervals. The dependency of the calibrated coefficients on the physical
conditions is similar for the different vaporisation models.
© 2019 Elsevier Ltd. All rights reserved.

1. Introduction
Space probes, rockets and satellites require manoeuvring sys-
tems for e.g. attitude control, re-boost and re-entrance into the
atmosphere. Small thrusters provide the necessary propulsion
using the recoil energy of the exhaust gases from an upstream
combustion process (Tanaka et al., 2011). Such orbital maneuvering
systems typically use hypergolic propellants such as monomethyl
hydrazine, but current interest focuses on their replacement by
non-toxic fuel-oxidizer combinations. One such combination for
the new generation thrusters is gaseous methane and liquid
oxygen (LOX) (Manfletti, 2014). The disintegration of the LOX jet
controls the subsequent mixing with the fuel and determines the
prospects of successful ignition events. If the chamber pressure is
below saturation conditions, the LOX becomes superheated and
small vapour bubbles nucleate and grow. Consequently, the jet
quickly expands and breaks up into small droplets. This process
is called flash boiling and can enhance the mixing process and
efficiency of the subsequent combustion (Lamanna et al., 2014;
∗
Corresponding author.
E-mail address: kronenburg@itv.uni-stuttgart.de (A. Kronenburg).
2015). However, any current modelling strategy does not resolve
bubble nucleation and growth of individual bubbles and will
require closures at the sub-grid scale (SGS). A quantification of
these SGS processes, i.e. bubble nucleation and bubble growth,
is necessary to improve the existing models. A possible strategy
would be the use of direct numerical simulations (DNS) to build a
database that allows for a detailed investigation of the underlying
physics and possible closures. DNS of flash vaporisation resolving
all the necessary processes are, however, extremely difficult to
realize and two of the key aspects are listed here:
• A compressible flow field solver is a precondition for the ac-
curate modelling of the early stages of flash vaporisation. The
early bubble growth is initiated by the local pressure field
which changes as the bubbles grow. The vapour densities in-
side the bubbles are directly linked to their pressures and the
fluid cannot be considered incompressible.
• It is important to realize that the term DNS – as it is com-
monly used –refers to a macroscopic approach where contin-
uum assumptions remain valid, all (turbulent) scales including
interface boundaries are well resolved and do not require mod-
elling. For flash vaporisation, however, the phase change ap-
pears in unclosed form and modelling at molecular scale would

https://doi.org/10.1016/j.ijmultiphaseflow.2019.04.010
0301-9322/© 2019 Elsevier Ltd. All rights reserved.
 D. Dietzel, T. Hitz and C.-D. Munz et al. / International Journal of Multiphase Flow 116 (2019) 250–269
be needed. The coupling of molecular dynamic simulations and
DNS for fully 3D bubble dynamics is beyond current capabilities
and some modelling of phase change is required.
The current paper will primarily focus on the implementation
of different closure strategies for the phase transfer rates in a com-
pressible DNS code, their calibration and their influence on bubble
growth. A good number of compressible multiphase solvers have
been developed within the last decade with recent examples in-
troduced in (Saurel et al., 2008; Zein et al., 2010; Lauer et al.,
2012; Chang et al., 2013; Houim and Kuo, 2013; Saurel et al., 2016;
Lee and Son, 2017; Fechter et al., 2017; 2018). Saurel and his co-
workers (Saurel et al., 2008; 2016) as well as (Zein et al., 2010)
solved separate equation systems for the two phases including one
volume fraction equation for each phase. The phase indicator dif-
fuses across several cells which are treated as homogeneous mix-
tures of liquid and vapour. This allows to treat two-phase flows
for larger length scales because interfaces of individual bubbles
do not need to be resolved. However, a sharp separation of the
two phase, as it is needed for a DNS of bubble dynamics at a
sub-grid scale, is difficult to achieve. Chang et al. (2013) solved
the fully compressible Navier-Stokes equations with a level-set
approach (Sussman and Smereka, 1994) that keeps the inter-
face between liquid and vapour sharp. However, Chang et al.
did not account for phase change. Lauer et al. (2012), Lee and
Son (2017) and Fechter et al. (2017, 2018) combined the level-set
algorithm with compressible solvers that can account for phase
transition. Lauer et al. (2012) used interface Riemann solvers with-
out phase change but source terms accounted for the mass trans-
fer from liquid to vapour. They estimated the mass fluxes with
the Hertz–Knudsen relation and calibrated the model coefficients
for oscillating cavitation bubbles. However, the physics of oscil-
lating bubbles differs from the physics of vapour bubble growth.
Moreover, Lauer et al. (2012) considered water and the coefficients
may not be valid for other fluids. Lee and Son (2017) used a ghost
fluid method and incorporate phase transfer into the jump condi-
tions which determine the ghost states. They estimated the mass
flux from the heat balance at the interface which requires a suit-
able resolution of the temperature boundary layer. The heat flux
determines the mass flux at the interface and no model coeffi-
cient requires calibration. Lee and Son (2017) used this solver to
investigate the growth and shrinkage of single compressible bub-
bles. However, the current paper focuses on LOX jets under flash
boiling conditions where thermal boundary layers are thin and
their resolution is not feasible if numerous bubbles are to be re-
solved in the computational domain. Fechter et al. developed an
exact (Fechter et al., 2017) and an approximative (Fechter et al.,
2018) Riemann solver for general two-phase Riemann problems.
The equation system consisted of Rankine–Hugoniot and interface
jump conditions and they used a kinetic relation to estimate va-
porisation rates. In Fechter et al. (2018) they showed that the ap-
proximate Riemann solver, called HLLP Riemann solver, is faster
than the exact solver and it gives a reasonable approximation of
the phase transition in shock tubes and for single droplets. This
Riemann solver is a modification of HLLC-type Riemann solvers
(Hu et al., 2009) and will be the basis of our work. However, gen-
eral calibration laws for the model parameter in the kinetic re-
lation do not exist but are required to accurately predict the ex-
pansion of vapour bubbles under flash boiling conditions. None of
these approaches can directly be used for the simulation of multi-
ple bubble growth in superheated LOX jets. The model coefficients
in Lauer et al. (2012) and Fechter et al. (2018) are not necessarily
valid for the physical conditions of our investigation and general
calibration laws do not exist. The resolution of the thermal bound-
ary layer as in Lee and Son (2017) is not feasible in cases with mul-
tiple bubbles for our process conditions. Thus, the main goal of our
251
investigation is to find a suitable calibration for different types of
vaporisation models such that the expansion of vapour bubbles in
superheated jets is accurately predicted. For this purpose we use
the growth of single vapour bubbles as a reference to validate the
model calibration. Using this simplified setup we determine the re-
quired coefficients, we demonstrate challenges in the compressible
DNS of vapour bubble growth and show how these challenges can
be overcome. In particular, we cover the following aspects:
• We quantify the growth of single vapour bubbles for cryogenic
oxygen for different temperature, pressure and time ranges.
• We justify simplifications in the initial conditions for vapour
bubble growth that considerably reduce the computational
time. This aspect is particularly important for future simula-
tions of multiple bubble growth.
• We use the Hertz–Knudsen relation (HKR), the kinetic relation
(KR) from Fechter et al. (2018) and an SGS heat flux model
(SHFM) to estimate the vaporisation mass flux and analyse their
calibration coefficients for the different temperature, pressure
and time ranges.
The paper is organized as follows: In Section 2 the fundamental
equations are presented. The numerical approach and the setup of
the present test cases are described in Sections 3 and 4. We show
our results in Section 5 and summarize our findings in Section 6.
2. Fundamentals
The fully compressible multiphase solver presented in
Fechter et al. (2018) serves as a basis for the work presented
here. Viscous forces are neglected since pressure force and phase
transition will dominate the expansion of the vapour bubbles. The
thermal boundary layers at the bubble walls do not exceed 15%
of the bubble diameter for the relevant process conditions (e.g.
Fig. 1(b)) and are typically even smaller for the majority of our
test cases (cf. Section 4 for details on the numerical configuration).
An adequate resolution would require cell sizes of much less than
1 μm, but such high resolution requirements are not feasible for
simulations of multiple bubbles and their interactions. Thus, we
do not resolve the heat diffusion but account for its influence by
the calibration of the vaporisation models. Under these conditions
the Navier–Stokes equations can be simplified to the Euler equa-
tions as introduced in Section 2.1. In Section 2.2, we present the
different models for the vaporisation mass flux and the interface
capturing approach is outlined in Section 2.3.
2.1. Conservation laws and jump conditions at the interface
The compressible Euler equations that govern the conservation
of mass, momentum and energy read
Ut + ∇ · F (U ) = 0 (1)
where
U = ( ρ , ρ u1 , ρ u2 , ρ u3 , E )T (2)
is the vector of the conserved variables. Here, ρ , u1 , u2 , u3 and E
are the density, the three Cartesian velocity components and the
total energy, respectively. The latter is the sum of the internal en-
u2
ergy e and the kinetic energy, E = ρ (e + 2i ). Here, i indicates the
three Cartesian velocity components and we use Einsteins notation
to sum up. The subscript t in Eq. (1) indicates the partial derivative
in time. The flux vectors F(U) for the three Cartesian coordinate
252 D. Dietzel, T. Hitz and C.-D. Munz et al. / International Journal of Multiphase Flow 116 (2019) 250–269

Fig. 1. Aspects of the sub-grid scale heat flux model.

directions read recent review on its application and calibration. The model equa-
⎛ ρ u1 ⎞ ⎛ ρ u2 ⎞ ⎛ ρu ⎞ tion reads
3 

⎜ ρ u1 + p ⎟ ⎜ ρ u1 u2 ⎟ ⎜ ρ u1 u3 ⎟
2
1 psat pvap
⎜ ⎟ F2 = ⎜ ρ u22 + p ⎟, F3 = ⎜ ρ u2 u3 ⎟, (3) m˙  = √ λvap − λcond
⎝ ρ u2 u1 ⎠,
F1 = (6)
⎝ ⎠ ⎝ 2 ⎠ 2π R Tliq Tvap
ρ u3 u1 ρ u3 u2 ρ u3 + p
u1 ( E + p ) u2 ( E + p ) u3 ( E + p ) where λvap and λcond are the vaporisation and condensation coeffi-
cients, respectively. They have to be scaled for a certain application
and p denotes pressure. An equation of state (EOS) gives the rela-
and we outline the calibration for our cases in Section 5.
tion between density, pressure and energy, i.e. p = f (ρ , e ), closing
the equation system (1) for a single phase. The open-source EOS The kinetic relation (KR)
library CoolProp (Bell et al., 2014) is coupled to the flow solver Fechter et al. (2018) used the KR in their original formulation
and any correlation is obtained from the functions provided in of the HLLP Riemann solver. They provided a detailed derivation
Bell et al. (2014). for the final model equation (Fechter et al., 2017),
We solve Eq. (1) separately for the individual bulk phases, i.e.
1
the liquid and the vapour phases. To couple these two bulk phases m˙  = T [[s − ssat ]]. (7)
we impose the jump conditions (Fechter et al., 2018) cent ref
This model ensures consistency with the second law of thermody-
[[ρ (un − Sint )]] = 0, namics, but entropy production at the interface is unknown and
[[ρ un (un − Sint ) + p]] = σ κ , (4) requires scaling by the coefficient cent (Fechter et al., 2017). For
[[ρ e(un − Sint ) + pun ]] + m˙  hlv = Sint σ κ . constant interface states, m˙  is inversely proportional to cent such
that the estimated mass flux decreases with an increasing model
Here, Sint , σ , κ and hlv are the interface velocity, the surface ten- coefficient.
sion coefficient, the curvature and the latent heat of vaporisation,
respectively. The mass flux due to vaporisation is m˙  and the oper- A sub-grid scale heat flux model (SHFM)
ator [[φ ]] denotes the difference between the vapour and the liq- The SHFM is based on the energy balance at the interface. Ne-
uid state, i.e [[φ ]] = φvap − φliq . The Euler equations do not resolve glecting the heat diffusion in the vapour phase we obtain the mass
the heat flux but the term m˙  hlv supplies the enthalpy jump across flux as the ratio of an approximated heat flux in the liquid phase
the interface. We model m˙  as will be presented in Section 2.2 and and the latent heat, viz.
evaluate hlv at Tref = 0.5(Tvap + Tliq ) from the EOS (Fechter et al., qliq − qvap qliq
2017; 2018). m˙  = ≈ . (8)
hlv hlv
The assumption of negligible diffusive heat fluxes in the vapour
2.2. Vaporisation model phase is justified as small temperature gradients in the bubble at
the interface can be expected. The heat diffusivity is higher in the
The interface velocity Sint and the mass flux m˙  in Eq. (4) are liquid and the heat flux in the vapour becomes important only for
coupled by very large temperature gradients. As the HLLP Riemann solver does
m˙  = ρliq (uliq − Sint ) = ρvap (uvap − Sint ) (5) not allow for diffusive effects, the thermal boundary layer will not
be resolved and we approximate the heat flux, qliq , by
and we use the following models to estimate m˙  and to close the 
dT  T∞ − Tsat
system of equations as given by the jump conditions (4). qliq = −kliq 
dx x=xint
≈ −kliq 2
δ
(9)

The Hertz–Knudsen relation (HKR)
The HKR originates from Hertz (1882) and Knudsen (1950) and
is based on a mass balance at the liquid-vapour interface. This
balance law is derived from the kinetic gas theory comparing the
molecule fluxes from and to the interface. Schrage (1953) compre-
hensively discussed the HKR and (Persad and Ward, 2016) gave a
assuming a parabolic temperature profile in the boundary layer
as illustrated in Fig. 1(a). Similar approaches to estimate the heat
flux in an unresolved thermal boundary layer exist for droplets
(Sazhin, 2014) but have not intensively been discussed in the con-
text of vapour bubble growth. The concept of the model within our
compressible solver is the following:
 D. Dietzel, T. Hitz and C.-D. Munz et al. / International Journal of Multiphase Flow 116 (2019) 250–269 253

• The saturation temperature is determined according to the multiplying the resulting system of equations by a test function φ ,
vapour pressure inside of the bubble. Starting from isothermal the governing equations in reference space read
conditions (see Section 4) the mass flux is initially set to zero.  
• The bubble growth is triggered numerically (see Section 4) and JUt φ dξ + ∇ξ · Fˆ (U )φ dξ = 0, (13)
E E
with increasing size the pressure in the bubble will decrease.
• Consequently, the saturation temperature decreases and the with the Jacobi determinant of the transformation J. An integration
temperature gradient in Eq. (9) becomes non-zero. The mass by parts of the second term in Eq. (13) gives the weak formulation
flux increases proportional to the temperature gradient. of the conservation laws in the discontinuous Galerkin framework,
• If the mass flux increases continuously, the vapour pressure will

increase and the decreasing temperature gradient relaxes the
mass flux.
The accuracy of the SHFM depends on the estimation of the
boundary layer thickness in Eq. (9) which we determine as
δ = Cδ R.
Fig. 1(b) shows an example for the ratio δ /R = Cδ as a function of
the bubble radius. In a first approach we keep Cδ constant within
one simulation but calibrate the coefficient for each individual test
case.
2.3. Level-set equation
We track the liquid-vapour interface using the level-set ap-
proach introduced by Sussman and Smereka (1994) where a signed
distance function (x, t) is transported as
∂
+ uLS ∇ · ( ) = 0.
∂t
The level-set determines the distance to the interface and the zero-
isocontour of defines the interface location. We obtain the ad-
vection velocity uLS of the interface from Eq. (5) and extend this
velocity to several neighbouring cells. We solve a Hamilton–Jacobi
equation to reinitialize such that it remains a signed distance
function (Sussman and Smereka, 1994). The normal vector of the
interface and its curvature are determined using the space deriva-
tives of according to
∇ ∇
 int =
n and κ = ∇ · . (12)
|∇ | |∇ |
3. Numerical methods
3.1. Discontinuous Galerkin spectral element method
The Euler equations are solved numerically by a discontinuous
Galerkin (DG) spectral element method. Here, we give a brief intro-
duction but the reader is referred to Hindenlang et al. (2012) and
Kopriva (2009) for a more detailed description. Within the dis-
continuous Galerkin approach the simulation domain is discretised
with a set of control volumes . Each cell is handled separately
and in a first step, a control volume is mapped from physical space,
x, to a reference space, ξ . The computational cells are spanned to a
unit reference element E of the size [-1;1] in each direction. After
Fig. 2. Wave fan of a two-phase Riemann problem with phase change. Left: Exact solution of the Riemann problem. Right: Approximate solution of the Riemann problem
in the HLLP context. The contact wave is omitted reducing the complexity of the problem.
∂
JU φ dξ + (Fˆ · N )φ dS − Fˆ (U ) · (∇ξ φ )dξ = 0. (14)
∂t E ∂E E
The first and the second terms on the left hand side define the
time derivative of the conserved quantity and the fluxes across the
surface of the DG cell, respectively. Each cell provides an inner cell
(10)
resolution which is accounted for by a continuous in-cell ansatz
function. As basis of the ansatz function Lagrange polynomials are
used, where the polynomial degree N determines the accuracy of
integration given in the third term on the left side of Eq. (14). As
polynomials are local for every cell, discontinuities may arise at
the cell face and Riemann solvers are used to determine consistent
fluxes across the cell boundaries. We apply HLLC Riemann solvers
at single phase cell boundaries (Toro, 2009) and couple the DG ap-
proach to a third order Runge–Kutta scheme for time integration.
3.2. Resolution at the interface
(11)
At the phase interface of small bubbles a strong contact dis-
continuity appears for many physical quantities, most notably for
density and pressure. Fluxes across the interface cannot be deter-
mined in the same manner as for single phase cell interfaces since
the discontinuity would lead to strong instabilities in the solution
procedure. Moreover, phase change complicates the Riemann prob-
lem at the interface where additional terms in the jump conditions
appear. We determine the fluxes across the interface using the
approximate Riemann solver from Fechter et al. (2018), which is
based on the ghost fluid method of Fedkiw et al. (1999). Fig. 2 de-
picts a typical two-phase Riemann problem including rarefaction,
shock and contact waves as well as the phase interface as an ad-
ditional discontinuity. The different fluid states describe the initial
liquid state, rarefied liquid, vaporised fluid, compressed vapour and
the initial vapour state from the left to the right, respectively. An
exact solution of the Riemann problem requires an iterative solu-
tion of the Rankine–Hugoniot and jump conditions across the dif-
ferent waves. The HLLP Riemann solver simplifies its solution using
the following assumptions:
• The outer wave speeds SL and SR are approximated by a lin-
earization of the Lax curves according to Fechter et al. (2013).
• The contact wave is omitted as proposed by
Harten et al. (1983) for single phase flow.
• The energy coupling across the Rankine–Hugoniot condition on
the vapour side is dropped.
254 D. Dietzel, T. Hitz and C.-D. Munz et al. / International Journal of Multiphase Flow 116 (2019) 250–269
Fig. 3. 1D example for the DG scheme with FV sub-cells: Each DG cell provides an inner cell resolution. In the vicinity of the interface the nodes of DG polynomials are
treated as FV cells. The numerical fluxes in the bulk phase are determined using a single-phase HLLC Riemann solver. The numerical fluxes at the phase interface are
determined using the HLLP Riemann solver.
The mathematical tool Maplesoft® solves the remaining system
of equations and provides explicit equations for the internal states
2σ 4μ
U∗L and U∗R . Fechter et al. (2018) discussed these simplifications in
detail and provided the explicit equations for the internal states.
The numerical fluxes are determined as functions of the outer and
the inner states according to Hu et al. (2009).
As continuous polynomials are used within one DG cell, it is
not possible to capture the interface discontinuity within the cell
at polynomial nodes. To enhance the interface resolution, the poly-
nomial nodes are treated as finite volume (FV) cells in the vicinity
of the interface by subdividing each DG cell into N + 1 FV sub-cells
per spatial direction. These cells are decoupled from their polyno-
mial ansatz function and discontinuities may arise between them.
For the present investigation a sharp interface approach is used.
Every computational cell is considered as either liquid or gaseous
and there are no mixed cells. Thus, interface fluxes can be deter-
mined directly across FV cell faces in the refined region and the
interface is shifted from one cell face to the next whenever the
sign of changes. This approach is needed to avoid a smearing
of the interface over several computational cells. Fig. 3 depicts this
concept that has been adapted from Sonntag and Munz (2014) and
Fechter and Munz (2015).
4. Physical conditions and numerical setup
The reader is reminded here of the objective of the work: the
development and validation of a solver that approximates vapour
bubble growth under flash boiling conditions. The problem as such
involves three-dimensional multiple bubble dynamics and their in-
teractions. The macroscopic simulation requires, however, mod-
elling the interface mass transfer. Hence, a suitable strategy is de-
veloped and validated here with the aid of the relatively simple
growth dynamics of a single, spherical bubble in a superheated
liquid. For such a case, analytical models and semi-analytical so-
lutions exist (Plesset and Properetti, 1977; Prosperetti and Plesset,
1978; Brennen, 1995; Lee and Herman Merte, 1996; Robinson and
Judd, 2004; Prosperetti, 2017) and serve as reference for the mod-
els used in the discontinuous Galerkin solver. In Section 4.1 we
introduce the most important aspects of single vapour bubble
growth and the reference solution. We summarize the relevant
physical conditions in Section 4.2, and outline the numerical setup
in Sections 4.3 and 4.4.
4.1. Single vapour bubble growth
Plesset (1949) and Plesset and Properetti (1977) extended the
early work of Rayleigh (1917) to the Rayleigh-Plesset equation
(RPE)
3 1

RR̈ + (R˙ )2 = p − p∞ − − R˙ (15)
2 ρliq sat R R
for single bubble growth. Here, R, R˙ and R̈ are the bubble radius
and its first and second time derivative, respectively. The solution
of Eq. (15) provides the temporal evolution of the bubble radius R
and its derived quantities (growth velocity, acceleration). Lee and
Herman Merte (1996) coupled the RPE to the temperature equation
 2 
∂T ∂T ∂ T 2 ∂T
+ ur =α + (16)
∂t ∂r ∂ r2 r ∂ r
and derived suitable boundary conditions at the bubble wall from
Eq. (8). We illustrate the assumptions commonly made for single
vapour bubble growth in Fig. 4. The surrounding liquid is super-
heated, i.e. it is below its saturation pressure at a metastable state
A, and vapour bubbles nucleate isothermally under saturation pres-
sure (state B). The surface tension force balances the pressure dif-
ference between vapour and liquid for a bubble of critical radius
(Brennen, 1995)
2σ
Rcrit = . (17)
psat (T∞ ) − pliq
Tiny fluctuations overcome the mechanical equilibrium and the
bubble grows initially at an extremely low rate. After the initial
delay the bubble grows at high rates driven by the inertia of the
liquid. The heat required for the vaporisation of the liquid cools
down the interface, and the vapour pressure as well as the growth
rate decrease. This process is indicated in Fig. 4 by the path from
state B to state C on the vapour side. A small boundary layer devel-
ops from the bubble wall towards the liquid, the vapour pressure
approaches the ambient pressure and heat conduction drives the
bubble growth. Eqs. (15) and (16) provide a relatively simple, one-
dimensional integrable solution (Lee and Herman Merte, 1996),
they account for the inertia and the heat diffusion driven stages of
bubble growth and can thus serve as a reference for the DG code
development.
The Rayleigh-Plesset equation (Rayleigh, 1917; Plesset and Prop-
eretti, 1977) and its coupling to the temperature equation by
Lee and Herman Merte (1996) describe the growth of isolated bub-
bles using several assumptions that warrant some further discus-
sion:
1. 1st assumption: The system consists of a spherical bubble in an
infinite liquid reservoir. This assumption allows to employ ra-
dial symmetry for the solution of bubble expansion. In a truly
 D. Dietzel, T. Hitz and C.-D. Munz et al. / International Journal of Multiphase Flow 116 (2019) 250–269 255

Fig. 4. Initial conditions for the present setup. State A indicates a metastable superheated liquid. State B indicates a vapour bubble which nucleates within the liquid. We
assume an isothermal nucleation with the vapour bubble being saturated. The resulting pressure gap is balanced by the surface tension.

flashing liquid multiple bubbles occur and the bubble density Table 1
Initial conditions for bubble growth parametric study. Densities are given in
is typically high. The bubbles may therefore interact with their
[kg/m3 ], pressures are given in [bar], critical radii in [μm] and temperatures in
neighbour and deviate from sphericity. The present study, how- [K]. The Jakob number Ja is dimensionless.
ever, considers isolated bubbles in a sufficiently extended liquid
Rp = 5 Rp = 10 Rp = 50 Rp = 100
environment only, and the fluid properties need to be selected
with respect to the ambient conditions (i.e. pressure and tem- T∞ = 90.0 pliq 0.199 0.099 0.02 0.01
perature) typical for flash boiling of cryogenic oxygen. Multiple psat = 0.994, ρ liq 1141.927 1141.905 1141.887 1141.885
ρvap = 4.39 Rcrit 0.332 0.295 0.271 0.268
bubble arrangements will be discussed in forthcoming work.
Ja 117.545 298.334 1965.017 4185.898
2. 2nd assumption: The vapour in the bubble is saturated and in
T∞ = 100.0 pliq 0.508 0.254 0.051 0.025
thermodynamic equilibrium with the liquid at the interface. If the psat = 2.54, ρ liq 1090.303 1090.230 1090.171 1090.164
growth rate of the bubble exceeds the rate at which vapour ρvap = 10.425 Rcrit 0.106 0.094 0.086 0.085
can be generated, the bubble pressure drops below the sat- Ja 62.883 158.303 1022.836 2162.338
uration pressure (Plesset and Zwick, 1954; Plesset and Prop- T∞ = 110.0 pliq 1.087 0.543 0.109 0.054
eretti, 1977). An estimation of the maximum rate of vapour psat = 5.434, ρ liq 1033.779 1033.566 1033.395 1033.374
generation is difficult because the accommodation coefficients ρvap = 21.281 Rcrit 0.039 0.034 0.031 0.031
Ja 39.698 99.442 632.984 1329.518
of the underlying kinetic model are in general not known
T∞ = 120.0 pliq 2.045 1.022 0.204 0.102
(Prosperetti, 2017). Shusser et al. (20 0 0) gave an example for
psat = 10.0223, ρ liq 969.208 968.612 968.133 968.073
an extension of the Rayleigh-Plesset equation incorporating a ρvap = 39.308 Rcrit 0.015 0.013 0.012 0.012
kinetic model for phase transfer rates, and nonequilibrium ef- Ja 28.780 72.016 453.795 948.112
fects are briefly addressed in Prosperetti (2017). Up to today,
uncertainties in the estimation of these nonequilibrium effects
at the interface persist, but it is reasonable to assume here that
processes at molecular scale are much faster than macroscopic context of flash boiling (Sher et al., 2008; Bar-Kohany and Levy,
processes (e.g. bubble expansion) as interface velocities are not 2016) and it is deemed reasonable to use the implementation by
too high (cf. Fig. 24). We therefore use the equilibrium assump- Lee and Herman Merte (1996) for the present investigation.
tion as it is common in similar studies.
3. 3rd assumption: the vapour pressure inside the bubble is con-
stant and the liquid is incompressible. The growth velocities of 4.2. Physical conditions
the isolated bubbles in this work do not exceed 20 m/s and
are therefore considerably smaller than the speed of sound in Experiments on flash boiling LOX jets are currently conducted
the vapour. It is valid to assume that the bubble pressure in- at the German Aerospace Centre (DLR) in Lampoldshausen and
staneously follows the pressure at the bubble wall (Plesset and first results of the jet break-up under such conditions have been
Zwick, 1954). The speed of sound of the liquid oxygen is larger presented in Lamanna et al. (2015). In these experiments, the
than 640 m/s for the present operating conditions confirming temperature of the cryogenic oxygen that is injected into a test
the assumption of an incompressible liquid phase. chamber ranges from T = 90 K to T = 120 K and the same range
is covered here. We define the superheat ratio between saturated
It is emphasized here that the approximation of Lee and Her-
vapour pressure and ambient pressure as (Lamanna et al., 2014)
man Merte (1996) is in general valid for both, the inertia and the
heat diffusion controlled regimes of bubble growth. Plesset and psat
Zwick (1954) derived an analytical model for the heat diffusion Rp = . (18)
p∞
controlled stage taking into account the evaporative cooling. This
model is, however, valid for thin boundary layers only. The proce- Table 1 lists the initial conditions of the different test cases that
dure of Lee and Merte resolves the temperature field and assump- span the entire temperature range and provide conditions from
tions regarding the boundary layer thickness are not needed. Up to very moderate to high superheat. We label the cases CX − Y with
today, the Rayleigh-Plesset equation is a common reference for the X and Y being the superheat ratio Rp and the initial temperature,
understanding and the quantification of the growth process in the respectively (e.g. C5-120 for R p = 5 and T0 = 120 K).
256 D. Dietzel, T. Hitz and C.-D. Munz et al. / International Journal of Multiphase Flow 116 (2019) 250–269

Fig. 5. Dimensionless bubble radius over time for different combinations of T0 and
Rp , respectively. The dashed line defines the analytical solution in the inertia con-
trolled range. The dash-dotted line indicates the analytical solution in the heat dif-
fusion controlled range.
Fig. 6. Reference solution for case C100-90 with (solid lines) and without (dashed-
lines) coupling of the temperature Eq. (16). Black lines show the dimensionless ra-
dius R+ and blue lines show the instantaneous mass flux. (For interpretation of the
references to colour in this figure legend, the reader is referred to the web version
of this article.)

Note that single bubble growth is typically characterized by the

Jakob number Ja
ρliq cp,liq (T∞ − Tsat ( p∞ ))
Ja = . (19)
ρsat,vap ( p∞ )hlv
It measures the degree of superheat as the ratio between heat
stored in the liquid and the latent heat needed for vaporisation
(Prosperetti, 2017). We base our parameter study on the pressure
ratio Rp instead because it is one important key parameter in flash
boiling (e.g. Lamanna et al. (2014)). Table 1 lists the corresponding
Ja which we also consider in the evaluation of our results.
We have discussed in Section 4.1 that bubble growth passes
different stages, where different physical mechanisms determine
the growth rates. If time and bubble size are normalized by
Prosperetti (2017)
A2 t AR
t+ = , R+ = 2 ,
B2 B tive cooling on the reference solution. The black lines depict the
 
2 hlv ρvap (T∞ −Tsat )
with A = 3 ρliq Tsat and B = Ja 12
π αliq , then the asymp-
totes of the inertia and heat conduction driven stages can be
approximated by linear and square root dependencies of R+ on
t + , respectively (Prosperetti, 2017). The normalization given by
Eq. (20) was originally introduced by Mikic et al. (1970) and covers
the inertia and heat diffusion driven stages. It uses several assump-
tions and empirical correlations which may introduce appreciable
uncertainties. However, Eq. (20) compares suprisingly well with
multiple experimental and numerical studies (Prosperetti, 2017;
Avdeev and Zudin, 2005). In the context of the current study, these
correlations provide first estimates of the relevant bubble growth
stages and the asymptotic limits for flash boiling of cryogenic oxy-
gen. Other analytical solutions coupling the different stages of bub-
ble growth are presented, among others, by Theofanous and Pa-
tel (1976), Prosperetti and Plesset (1978) as well as Avdeev and
Zudin (2005), but are not further discussed here.
The The t + − R+ relationship according to Mikic et al. (1970) is
demonstrated in Fig. 5. Here, the reference solutions for sin-
gle vapour bubble growth have been normalized as introduced
by Eq. (20) and the dashed and dash-dotted lines indicate the
asymptotic limits. It is apparent that inertia driven growth holds
for times smaller than t + = 10−2 only, while diffusion controlled
growth requires growth for time scales longer than t + = 50. The
intermediate range is defined from 10−2 < t + < 50. In Fig. 5 we
show the cases C5-90, C5-120, C100-90 and C100-120, which cover
a wide spectrum of Ja including the highest and the lowest degree
of superheat within our parameter range. We restrict the maxi-
mum physical simulation time of our investigation to t = 100 μs
because flash vaporisation is not limited to single bubble growth,
nucleation rates are high and bubbles will rapidly interact and
merge. Only case C5-120 becomes heat diffusion controlled within
this time interval and all other cases are either inertia controlled
or lie within the intermediate range. Increasing Rp or Ja shifts the
dimensionless solution towards the inertia controlled range for the
same final time while an increase in T0 has the opposite effect.
Lee and Herman Merte (1996) observed the same trend comparing
different experimental investigations. This analysis shows that the
majority of our test cases is inertia controlled justifying the use
of the Euler equations. At the same time, we cannot calibrate the
vaporisation models towards a mass flux which is solely derived
from the solution of Eq. (15) without coupling of the temperature
equation as we explain with the aid of Fig. 6.
Fig. 6 demonstrates the effect of heat diffusion and evapora-
(20)
bubble radius over time for case C100-90 with (solid lines) and
without (dashed lines) diffusive terms, where the initial bubble
growth is nearly identical. The blue lines show the correspond-
ing vaporisation mass flux and we observe considerable differences
between the coupled and the uncoupled solutions even at early
stages, where the bubble radii still feature very similar values. The
mass flux is determined as m˙  = ρvap R˙ . Without heat diffusion the
mass flux increases and remains at a high level. At the same time,
the pressure in the bubble and the vapour density remain con-
stant. If we consider evaporative cooling (and thereby heat diffu-
sion), the mass flux decreases after an early peak, the pressure in
the bubble decreases and the vapour density decreases as well. In
the inertia controlled range, the growth rate of the bubble (R˙ ), is
not affected by the different mass fluxes because the vapour den-
sity changes for the different approximations. Therefore, the volu-
metric flow rate V˙ = A · m˙  /ρvap (with A being the bubble surface)
is similar for the two approaches in the inertia controlled regime.
We conclude that evaporative cooling affects the bubble pressure
and the mass flux throughout the entire growth process, but the
growth rate at later stages only. In the DG solver, the bubble pres-
sure will also drop despite the omission of heat conduction and
subsequent temperature decrease at the interface since the bub-
ble is fully resolved and surface effects become small as the radius
increases. The mass flux of the DG solver should therefore not be
calibrated towards the constant mass flux derived from the RPE
solution without heat diffusion, as this strongly overpredicted the
growth rate. We conclude that evaporative cooling and heat diffu-
sion are required to predict all physical quantities accurately, even
for the inertia driven stage. In Section 2 we have discussed the
 D. Dietzel, T. Hitz and C.-D. Munz et al. / International Journal of Multiphase Flow 116 (2019) 250–269
Fig. 7. Results obtained from the RPE coupled to the temperature equation for different temperatures (colours) at R p = 5.
challenges in resolving the temperature boundary layers, in par-
ticular regarding our forthcoming work with multiple bubbles. To
keep the computational effort reasonable, we continue using the
Euler equations and we calibrate our vaporisation models such that
the growth rates of the bubbles are matched for particular time in-
tervals. We assess the validity of this approach in Section 5.1.
Fig. 6also indicates that the relevant length scales of bubble
growth easily cover three orders of magnitude. For different phys-
ical conditions these ranges may even be larger for the same time
interval. This is shown in Fig. 7(a) where we depict the physical
bubble radii at R p = 5 for the different temperatures. The bubble
grows by four orders of magnitude for T0 = 120 K which is impos-
sible to resolve in full 3D computations. As the volumetric jet ex-
pansion and jet break-up are key quantities of interest, it is fair
to stipulate that the early stages of growth are not relevant. In
Fig. 7(b) we depict the total vaporised mass (solid lines) and the
mass flux (dashed lines) as functions of the bubble radius R. The
mass flux peaks at low R but the total vaporised mass accumulates
during the final stages, due to its R2 -dependence. Thus,
• it is valid to assume that early bubble growth can be neglected
and that
• the initial bubble radius for the DG simulation can be increased
to R0 = 0.1Rend . The computations would then cover 99.9% of
the volume expansion of a single bubble growing from Rcrit to
Rend .
We introduce a normalized radius
R f, NDG and rref are an adjustable stretching coefficient, the num-
R∗ =
Rcrit
and use the initial radii
R0
R∗0 = = {1, 5, 10, 25, 50}
Rcrit
for each combination of T∞ and Rp . The reference solution defines
our final simulation time for each combination of R∗0 , T∞ and Rp
with tend = f (Rend = 10R0 ). When we initialize the DG simulation
with R∗0 = 1 we increase the start radius by one FV sub-cell (i.e.
one discretisation point) to trigger the growth of the bubble. Note
that we always start from the critical radius in our reference solu-
tion. In summary, our test matrix is defined by three independent
variables. Two of these parameters, T0 and Rp , define the physical
conditions and the third parameter, R∗0 , defines the range of the
length and time scales scales that are covered during the simula-
tion time.
257
4.3. Numerical setup
The spherically symmetric reference solution allows for the use
of identical symmetry conditions for the DG solver, and the source
term (Toro, 2009)

2
ρu
S= ρ u2 (23)
r
u (E + p )
is added on the right hand side of Eq. (1) reverting the 3D sys-
tem to a 1D configuration with spherical symmetry. Fig. 8 depicts
the geometrical setup where we set the bubble in the left part of
the domain. The left boundary at r = 0 indicates the bubble centre
and we apply a symmetry condition here. The superheated liquid
fills the right side with an outflow boundary condition at r = L. We
keep pressure wave reflections small by using a weak outflow con-
dition and a large domain size L. In the weak outflow condition the
Riemann problem is solved for the initial state of the liquid and the
internal state at the boundary. The default domain size satisfies the
condition
tend cliq
L= . (24)
2
ensuring that a pressure wave can travel only one time forth and
back through the domain. A stretched computational grid is de-
fined by
1 − C rref
C = f · NDG , r ∗ = and rnode = L · r ∗ (25)
1 −C
determining the radial position rnode of a polynomial node. Here,
(21)
ber of discontinuous Galerkin elements and the node coordinate in
an equidistant reference domain ranging from 0 to 1, respectively.
The degree of the polynomial is kept constant for all simulations
with N = 3. The default stretching factor is f = 1.03 and the num-
(22) ber of DG elements NDG is set such that a bubble is initially re-
solved with at least ten discretisation points. The numerical test in
the present section are accomplished using the Hertz–Knudsen re-
lation. The model coefficients are calibrated against the reference
solution from Section 4.2 to λvap = 0.038 and λcond = 0.0376, re-
spectively.
The preceding section discussed the condition for the initial ra-
dius according to Eq. (22). The increase in initial radius reduces
the computational cost compared to R0 = Rcrit for two reasons:
firstly, for a fixed number of cells across the bubble radius the to-
tal number of computational cells reduces. Secondly, the required
time step increases with the cell width and therefore with R0 if,
again, the number of cells across the bubble radius is constant. For
258 D. Dietzel, T. Hitz and C.-D. Munz et al. / International Journal of Multiphase Flow 116 (2019) 250–269

Fig. 8. Geometrical setup for single bubble growth.

Table 2 Table 3
Relative computational time for different R∗0 in percent. Parameters for domain size sensitivity
study. The default length L is given by
R∗0 = 5 R∗0 = 10 R∗0 = 25 R∗0 = 50 Eq. (24).
tCPU /tCPU,Rcrit 15% 7% 2% 1%
L in [mm] f [–] DG cells

L/2 0.165 1.029 100

L 0.33 1.03 120
a given final simulation time, the total number of time steps re- 2L 0.66 1.031 140
duces with increasing time step size. The total computational time
is proportional to the number of time steps and the number of
cells. The computational time relative to the computational time Table 4
Parameters for cell size sensi-
with R0 = Rcrit reads
tivity study.
tCPU tRcrit NDG
= . (26) Nbub f [–] DG cells
tCPU,Rcrit t NDG,Rcrit 5 1.062 60
Table 2 lists the results for R∗0 = 5 . . . 50. They do not strongly de- 10 1.03 120
20 1.015 240
pend on T0 nor on Rp . The meshes for the present analysis were
constructed such that the number of cells across the bubble radius
and the cell width approaching the right end of the domain (cf.
Fig. 8) are constant for R0 = Rcrit and R0 = R∗0 · Rcrit , respectively. at the final time and an increase in p on the liquid side is observed
The computational time reduces to at least 15% compared to the for L/2 compared to L and 2L. The differences arise due to weak
cases with R0 = Rcrit . At R∗0 = 50 the computational time reduces shock wave reflections at the outflow boundary condition. The de-
to 1% emphasising the benefit of using Eq. (22). viation in the pressure field does not affect the bubble growth but
using the domain size L in Eq. (24) guarantees mesh independence
4.4. Mesh resolution and time step restriction throughout the investigation, and L is used for the remainder of
the paper. Fig. 10 depicts the same quantities for a varying reso-
A grid dependency study has been conducted to ensure inde- lution of the initial bubble with Nbub = [5, 10, 20]. Table 4 lists the
pendence of results on the major numerical parameters. These in- corresponding mesh parameters. The three cases are in very good
clude the domain size, L, the bubble resolution, i.e. the grid size, agreement indicating mesh independent results for the given crite-
and the time step that is here expressed with the aid of the ria. Fig. 10 also illustrates results when using a reduced time step
CFL number. The influence of these parameters is studied for C5- (here, the CFL number is set to 0.45). Again, no significant change
120 with R∗0 = 50. First, the domain size is varied to assess the in the growth rates and in the resulting pressure field is observed.
influence of wave reflections from the boundary: L is given by When comparing the different numerical setups, the bubble radii
Eq. (24) and increased and decreased by a factor of 2 for the do- deviate by a maximum of 2.2%, 1.4% and 0.9% for cases with a vari-
main size sensitivity study. The specific parameters are given in able cell width, time step size and domain length, respectively. It
Table 3. Fig. 9(a) depicts the bubble radius over time which is not is noted here (1) that - similar to R - the numerical parameters do
sensitive to the domain size. Diagram 9(b) shows the pressure field not unduly affect the evolution of m˙  and V and (2) that very

Fig. 9. Results of C5-120 with different domain sizes L for Nbub = 10.
 D. Dietzel, T. Hitz and C.-D. Munz et al. / International Journal of Multiphase Flow 116 (2019) 250–269 259

Fig. 10. Results of C5-120 with different bubble resolutions Nbub and with CF L = 0.45 for L = L2 .

similar mesh sensitivities have been observed when using the ki-
netic relation (KR) and the sub-grid scale heat flux model (SHFM)
as a vaporisation model. Mesh independence of all results reported
below can be assumed.

5. Results and discussion

5.1. Hertz–Knudsen relation

A direct calibration of the HKR is difficult because two model

coefficients appear in Eq. (6). Setting λcond = λvap reduces the ac-
curacy of the model but simplifies the estimation of the remain-
ing model coefficient. Lauer et al. (2012) calibrated this single co-
efficient according to a reference solution for the simulation of
Fig. 11. Normalized bubble radius R∗ over time.
cavitation bubbles. We use a similar strategy but suggest λcond =
0.99λvap to stabilize the simulations. At the initial time steps,
when the solution deviates only slightly from mechanical and ther-
modynamic equilibrium, vaporisation is favoured and early bubble
growth is stabilized. Here, we are not interested in bubble collapse
nor bubble oscillations. At later time steps, the proposed reduction
in the condensation rate does not change the results. The condition
λcond = 0.99λvap holds throughout the paper but we refer only to
λvap in the subsequent discussion. In the following, we compare
the DG results to the reference solution and we start our discus-
sion for case C5-120 with R∗0 = 50.

C5-120, R∗0 = 50
The base case of our analysis is C5-120 with R∗0 = 50. The black
line in Fig. 11 depicts the normalized bubble radius R∗ over time
obtained from the DG solver while the red line shows the corre- Fig. 12. Vaporisation mass flux m˙  of case C5-120 with R∗0 = 50.
sponding reference solution. The DG solution agrees well with the
reference solution for the relevant time and length scales. The de-
viations are large only for t/tend < 0.2 because the initial radius for
the DG solution had been increased. The corresponding vaporisa-
tion coefficient λvap = 0.038 scales the mass flux towards m˙  ≈ 65
kg/(m2 s), see Fig. 12. The vaporisation mass flux converges to a
constant value because the liquid remains superheated in the HLLP
approximation. Initially, the vapour in the bubble is saturated and
Eq. (6) gives a zero mass flux. We are not in mechanical equi-
librium, the bubble grows and the pressure in the bubble drops.
With decreasing pressure the mass flux increases and the pres-
sure drop is slowed down. After a short time the bubble pres-
sure converges towards the ambient pressure and a steady state
is reached for the bubble pressure and the mass flux. Fig. 13 de-
picts the pressure field at different time steps. We observe that the
bubble pressure has approximately reached the ambient level af-
ter t/tend = 0.08. After that time, the bubble pressure and the mass Fig. 13. Pressure field of C5-120 and R∗0 = 50 at different time steps.
260 D. Dietzel, T. Hitz and C.-D. Munz et al. / International Journal of Multiphase Flow 116 (2019) 250–269
Fig. 14. Integrated quantities of C5-120 and R∗0 = 50.
Fig. 15. Normalized bubble radius R∗ over time for different length scales.
flux remain unchanged. The qualitative behaviour of m˙  in the ref-
erence solution cannot be matched but the integrated mass flux
depicted in Fig. 14(a) deviates by only 10%. We have calibrated the
vaporisation coefficients such that we match the final bubble ra-
dius. The total vaporised mass deviates from the reference solution
because the vapour densities are not exactly the same as in the
reference solution and the same volume contains a different mass.
Our main focus is the volume expansion of the bubbles. Fig. 14(b)
shows the total change in the bubble volume where the deviations
are smaller. Thus, the HLLP Riemann solver coupled to the HKR is
a suitable approach to predict the volumetric expansion of single
vapour bubbles. Note that we can improve the results for total va-
porised mass adopting the calibration to the target quantity mtot .
However, we have found that the calibration towards mtot is more
difficult to obtain, particularly at low R∗0 . Given the good match in
V we calibrate the vaporisation models towards R in the remain-
ing investigation. The calibration of λvap is specific for the chosen
time scale. In the next step, we analyse the validity of our ap-
proach for the same physical conditions, i.e. constant T0 and Rp ,
at different time intervals implicitly defined by R∗0 (cf. Section 4.2).
Variation of the initial radius R∗0
We decrease R∗0 towards the critical radius R∗0 = 1 keeping the
time interval of each simulation determined by R∗end = 10R∗0 . We
depict the bubble radii as a function of time in Fig. 15 where each
individual calibration provides a good match to the reference so-
lution for the final simulation time. The growth at the relevant
length and time scales matches the reference solution well as seen
for R∗0 = 50 in Fig. 11. Fig. 16 depicts the total vaporised mass and
the total volume change as a function of time for C5-120 with
R∗0 = 1 and R∗0 = 10. We observe a deviation in mtot of more than
25% if we start from the critical radius R∗0 = 1. The generated vol-
ume matches well the reference solution only close to tend . In con-
trast to that, we observe a very good agreement for both mtot and
V starting from R∗0 = 10. While we predict mtot differently well
for the combinations of R∗0 , Rp and T0 we obtain very good matches
in V for all cases with minor restrictions for R∗0 = 1.
Variation in the physical conditions
We conduct the next simulations with R∗0 = 50 and vary the su-
perheat ratio Rp and the temperature T0 . Fig. 17 shows the volu-
metric expansion for the cases C100-120 and C5-100, which agree
well with the reference solution. We obtain similar results for the
various combinations of Rp , T0 and R∗0 . Regarding the simulation of
single vapour bubble growth we conclude that
• the simplifying assumptions discussed in Section 4.3 are valid
in most of our cases,
• we can predict the bubble growth in a wide parameter range
using the HLLP Riemann solver combined to the HKR,
• the instantaneous mass flux cannot be predicted with the HKR
within the HLLP Riemann solver.
5.2. Kinetic relation and sub-grid scale heat flux
In this section, we compare the two alternative vaporisation
models to the HKR. It is first noted that all findings discussed in
Figs. 15–17 equally hold for the KR and SHFM closures but are
not repeated here for brevity of presentation. Fig. 18 compares
the vaporisation mass fluxes, the bubble radii and the volumetric
expansion. The red and black dashed lines show results for the
KR and the SHFM, respectively. The peak mass flux (Fig 18(a)) is
slightly higher for the KR when compared with the HKR (black
solid line), but – as the evaporative cooling in the HLLP approach
is not resolved – it rapidly asymptotes towards a constant value.
The asymptotes of the HKR and KR deviate by approximately 5%.
Particularly at early times m˙  fluctuates but the volumetric ex-
pansion is nearly identical to the HKR approximation (Fig. 18(a)).
Similarly, trends observed for different initial radii R∗0 and for
different physical conditions follow trends described for the HKR
closure. The black dashed lines in Fig. 18 represent the SHFM.
Here, the mass flux agrees qualitatively better with the reference
solution than the HKR and the KR solutions. We observe a peak
during the initial stages and a continuous decrease for larger time
intervals. This results from the dependency of the vaporisation
model on the continuously growing boundary layer thickness as
it is defined by Eq. (10). However, we underpredict the reference
mass flux for the entire time interval. In contrast, the volumetric
expansion is overpredicted because the increased radius R∗0 = 50
D. Dietzel, T. Hitz and C.-D. Munz et al. / International Journal of Multiphase Flow 116 (2019) 250–269 261

Fig. 16. Total vaporised mass and volume change of C5-120 for different R∗0 .

Fig. 17. Volumetric expansion for R∗0 = 50 under different physical conditions.

Fig. 18. Comparison of HKR, KR and SHFM for case C5-120 and R∗0 = 50.
262 D. Dietzel, T. Hitz and C.-D. Munz et al. / International Journal of Multiphase Flow 116 (2019) 250–269
Fig. 19. Vaporisation coefficients λvap as a function of different characteristic quantities of our investigation.
results in a larger bubble surface which allows for a faster volume
expansion of the bubble (Fig. 18(a)). At the end of the simulation
time all vaporisation models predict the volumetric expansion
well. A better estimation of the boundary layer thickness δ can
improve the prediction of the volumetric expansion of the SHFM
for the early stages. This could be achieved by a dynamic deter-
mination of Cδ but further research would be required to find
suitable models and is beyond the scope of the current work.
5.3. Prediction of model coefficients
In the previous sections we have demonstrated that our simu-
lation strategy is valid for a wide range of parameters. The tables
in Appendix A list the corresponding calibration coefficients. The
main disadvantage of this approach is the need to iteratively cali-
brate the model coefficients for each combination of the initial pa-
rameters T0 , Rp and R∗0 . A truly predictive simulation requires the
modelling of the calibration coefficients and in this section, we dis-
cuss the relation of the model coefficients to the characteristic pa-
rameters of our study.
Fig. 19 depicts the vaporisation coefficient λvap as function of
R∗0 , Rp , T0 , Ja, R+ or t+ (Fig. 19(a)–(f), respectively). For clarity of
presentation, the diagrams show the results from T = 90 K and T
= 120 K only, but trends for the intermediate values reveal similar
trends and are therefore omitted from the figures.
• Dependence on R∗0 (Fig. 19(a)): For fixed physical conditions, λvap
scales inversely proportional with R∗0 . For the reference case,
the mass flux decreases monotonically with time/radius after
its early peak (cf. Fig. 12). The combination of the DNS solver
with the HKR closure yields – after a short transient stage at
the beginning of a simulation – a constant mass flux. Depend-
ing on the growth stages that are covered between R∗0 and R∗end ,
λvap scales this constant mass flux such that the mean value of
the reference solution is approached. At small R∗0 the depen-
dency can be reversed as the growth rate does not only de-
pend on the (highly transient) mass flux but also on the intertia
forces.
• Dependence on Rp (Fig. 19(b)): For small superheat ratios, no
obvious relation between Rp and λvap is found. The solid and
 D. Dietzel, T. Hitz and C.-D. Munz et al. / International Journal of Multiphase Flow 116 (2019) 250–269 263

Fig. 20. Comparison of estimated and calibrated λvap .

dash-dotted lines in Fig. 19(b) represent cases with R∗0 = 5 and

R∗0 = 50, respectively. While λvap decreases with increasing Rp
at low R∗0 , it increases with increasing Rp at high R∗0 . The ref-
erence growth rate increases with increasing Rp at later time
intervals. Therefore, λvap also increases with increasing Rp in
the range of large R∗0 . At smaller R∗0 the reference growth rate
also increases with increasing Rp but the differences are less
pronounced. The inertia forces strongly affect the early growth
stage, and λvap can therefore not follow the trend observed for
higher R∗0 .
• Dependence on T0 (Fig. 19(c)): The calibrated coefficients λvap
decrease with decreasing T0 if R∗0 and Rp are kept constant. As
not only temperature changes but also the saturation pressure
and - for a constant Rp - the ambient pressure are functions
of temperature, and all these quantities affect the HKR (see Fig. 21. Vaporisation mass fluxes normalized by the vapour density and the factor
Eq. (6)), the correlation between λvap and T0 is difficult to as- A in Eq. (20) as a function of the dimensionless time t + . After an initial transient
sess. The driving pressure difference in Eq. (6) and therefore the the curves for all cases approximately converge to a single curve.
resulting mass flux typically decrease for smaller temperatures.
Here, we observe that the calibration of λvap also reduces the
is smaller than one. The opposite applies for large R∗0 and Fig. 20(b)
mass flux for lower temperatures.
confirms this theory. The red line in Fig. 20(a) is a least-squares fit
• Dependencies on Ja, R+ and t + (Fig. 19(a)–(f)): The coefficient
of the data points, which can be used to approximate λvap,pre based
λvap does not reveal any simple dependencies on either of the
on the initial estimation λvap,RPE with
characteristic numbers Ja, R+ and t + . Trends are observed if Rp
and R∗0 are fixed (Fig. 19(a)) or if Rp and T0 are fixed (Fig. 19(f)) λvap,pre = λvap,RPE · 1.27 · (1 − exp(−R∗0 · 0.086 ))0.61 . (28)
but in general the characteristic number can spread over sev-
eral orders of magnitude for similar vaporisation coefficients. The largest deviation between λvap,cal and λvap,pre occurs for R p =
100, T0 = 120 K and R∗0 = 5. The predicted coefficient and the re-
It is thus important to note here that there is no apparent cor- sulting volume expansion differ by approximately 45%. Fig. 20(c)
relation of λvap with any of the characteristic quantities that would summarizes these deviations for all test conditions. The error tends
allow for an a priori estimate of the model parameter and alter- to decrease for increasing initial radii and Fig. 21 gives one pos-
native procedures for modelling need to be sought. Similar find- sible reason for this trend. Here, the mass fluxes are normalized
ings are made for the KR and SHFM closures and are presented in with the coefficient A of Eq. (20) and with the density of the sat-
Appendix B. urated vapour in the bubble. The black lines show solutions from
A possible alternative modelling procedure for the calibration the RPE which converge after an initial highly transient stage into a
coefficients shall now be based on the reference solution. Since the single curve. The slope of this curve is approximately proportional
results follow similar trends for the different vaporisation mod- to (t + )−0.5 and remains roughly constant within the time interval
els we restrict this analysis to the HKR. With λcond = 0.99λvap , considered. The modelling of the early transients is rather compli-
Eq. (6) can be rearranged to yield the vaporisation coefficient as cated with the present method but the volumetric expansion (that
√ is our target quantity) is small for these stages and the accuracy
m˙  2π R
λvap,RPE = psat . (27) improves for the relaxed stages.
√ − 0.99 √pvap
Tliq Tvap
5.4. Sensitivity analysis for the model coefficients and DNS results
We obtain the mass flux m˙  , the vapour pressure pvap and the
vapour temperature Tvap from our reference solution at R = Rend . The preceding sections discussed the dependency of the vapor-
The saturation pressure psat is determined as a function of the isation model coefficients on the physical initial conditions and the
liquid temperature Tliq = T∞ . Fig. 20(a) depicts the ratio between initial bubble radii. This section discusses the sensitivity of the DNS
the calibrated coefficients λvap,cal (Appendix A) and the coefficients results to changes in the model coefficients for fixed initial condi-
λvap,RPE from Eq. (27). The estimates λvap,RPE differ from the cali- tions.
brated coefficients, but the deviations are similar for constant R∗0 . Fig. 22 depicts the bubble radius as a function of time for dif-
For small R∗0 the mass flux in the RPE is larger than the mass flux ferent physical parameters and vaporisation models. The red and
in the DG solution. As a result λvap,RPE is larger and λvap,cal /λvap,RPE black lines represent the calibrated results and results with a 10%
264 D. Dietzel, T. Hitz and C.-D. Munz et al. / International Journal of Multiphase Flow 116 (2019) 250–269

Fig. 22. Sensitivity of the bubble radius to a 10% change in the model coefficient. All cases use R∗0 = 50.

change in the model coefficient, respectively. In all cases, an in-

crease or decrease in the coefficient yields faster or slower bubble
growth. The maximum difference in the bubble radii between the
calibrated and the disturbed solutions is typically smaller than the
disturbance (usually about 2–5%, only in a specific case equal to
the disturbance) as the density of the fresh vapour also changes.
An increase in the mass flux as a result of a change in the coef-
ficient yields an increased density and the volume expansion re-
duces. The inverse effect is observed if the mass flux decreases.
The results show that a calibration of the vaporisation models is
needed for an accurate predictions of the bubble expansion but the
sensitivity to the exact value is moderate as a dampening rather
than an enhancement of a disturbance from the optimum value is
observed.
Fig. 23. Normalized bubble radius R∗ over time for C5-120 with R∗0 = 50. The DG
simulation with the 1D solver (black line) and the 3D solver (blue crosses) match
5.5. 3D simulation of single bubble growth
the reference solution (red line). (For interpretation of the references to colour in
this figure legend, the reader is referred to the web version of this article.)
As a final test case, we demonstrate the capabilities of the DG
solver in a truly three-dimensional system. Here we use case C5-
120 with an initial radius of R∗0 = 50. The HKR is taken for the clo- have discussed this effect in Sections 4.2 and 5.1. The right plane
sure of the interface mass transfer rate and the results from the 3D of the subfigures in Fig. 24 shows the characteristic velocity field
computation are compared to the 1D computation and to the refer- for single vapour bubble growth: the velocity peaks at the inter-
ence solution. Fig. 23 depicts the dimensionless bubble radius over face in the liquid phase and decreases towards the outside due
time. The three-dimensional case agrees very well with the one- to the spherical symmetry of the problem. The density fields in
dimensional simulation and with the reference solution. We also the bottom planes demonstrates that the sharpness of the inter-
show a three-dimensional representation of the bubble growth in face is guaranteed. Last but not least, mass and energy conser-
Fig. 24. The four subfigures from the top left to the bottom right vation needs to be ensured due to the non-conservative charac-
depict different simulation times ranging results from t = 0.0 μs to ter of the level-set approach. Fig. 25 illustrates the relative change
t = 1.0 μs. The left plane of each subfigure illustrates the pressure of the total mass and the total energy throughout the simula-
field. We observe a pressure wave propagating away from the in- tion. Both increase by less than 1% demonstrating that the present
terface as the bubble grows. Note, that the color bar scales with three-dimensional solver reasonably conserves mass and energy.
a maximum pressure of p = 3 bar which indicates that the global The results show that the DG solver developed here provides a
pressure level rapidly relaxes towards the ambient pressure. We useful tool that can be used in future studies for the computation
 D. Dietzel, T. Hitz and C.-D. Munz et al. / International Journal of Multiphase Flow 116 (2019) 250–269
Fig. 24. Three-dimensional representation of the bubble growth.
Fig. 25. Mass and energy conservation.
of multi-bubble dynamics and their interactions leading to spray
break-up under flashing conditions.
6. Summary and outlook
The goal of the present investigation was the simulation of sin-
gle vapour bubble growth with a fully compressible two-phase
solver. We have used the HLLP Riemann solver coupled to a discon-
tinuous Galerkin method and a third order Runge–Kutta scheme to
discretise the Euler equations in space and time, respectively. The
solver conserves a sharp interface between the liquid and vapour.
We have compared three different models to estimate the vapor-
265
isation mass flux. The first model is the Hertz–Knudsen relation
(HKR) which is based on the difference between saturation pres-
sure and vapour pressure. In the second approach we apply the
kinetic relation (KR) according to the original formulation of the
HLLP Riemann solver. It estimates the mass flux depending on the
deviation of the local entropy to the saturation entropy. The last
approach determines the vaporisation mass flux based on a sub-
grid scale heat flux and is denoted as sub-grid heat flux model
(SHFM). Here we have assumed a parabolic temperature profile in
the liquid within the boundary layer and have derived the tem-
perature gradient from the first derivative at the interface. These
models are calibrated to match the bubble radius of a reference
solution for wide pressure, temperature and time ranges as the ul-
timate objective of the DG code is its use to predict spray break-up
under flash boiling conditions. The operating conditions are char-
acteristic for superheated cryogenic oxygen issued in combustion
chambers near vacuum.
We have observed similar results using the HKR and the KR.
The instantaneous character of the reference mass flux is not cap-
tured but the volumetric expansion of bubbles and the total va-
porised mass are well predicted. We have found uncertainties only
at the smallest time scales. Typical length and time scales vary a
lot within our parametric study and the calibrated model coeffi-
cients change accordingly. For an increasing temperature the cal-
ibration promotes vaporisation but a general dependency on the
time scales or pressure ranges is not obvious. The SHFM qualita-
tively matches the reference solution for the mass flux well. How-
ever, the bubble radius and the volumetric expansion match to the
reference case only in the last part of a simulation. A more sophis-
ticated representation of the model coefficient may improve the
266 D. Dietzel, T. Hitz and C.-D. Munz et al. / International Journal of Multiphase Flow 116 (2019) 250–269

results for the volumetric expansion and gives potential for an ex-
plicit model without a-priori calibration.
Generally, we have shown that the HLLP Riemann solver can be
used for the numerical simulation of single vapour bubble growth.
A coupling to different vaporisation models and a suitable calibra-
tion of model coefficients can be based on simple a-priori compu-
tations using a reference case and the DG solver will allow for a
prediction of the volumetric bubble expansion. The solver is stable
in the most cases we have investigated even with high pressure
ratios between vapour and liquid. This model will be used in fu-
ture investigation for the simulation of multiple bubble growth in
superheated liquid jets.
Acknowledgements
The authors acknowledge the financial support by the Deutsche
Forschungsgemeinschaft (DFG) as part of the Collaborative Re-
search Center SFB TRR 75 “Droplet dynamics under extreme ambi-
ent conditions” held jointly by the University of Stuttgart and the
University of Technology Darmstadt.
Appendix A. Coefficients of the different vaporisation models
This section lists the calibrated coefficients for the HKR, the KR
and the SHFM for each combination of T0 , Rp and R∗0 . If no value is
given, no stable solution was obtained.

Table A.1
λvap for R p = 5 and R p = 10.
R∗0 C5-90 C5-100 C5-110 C5-120 C10-90 C10-100 C10-110 C10-120

1 0.050 0.071 0.093 0.110 0.045 0.064 0.084 0.093

5 0.056 0.071 0.084 0.093 0.052 0.065 0.080 0.083
10 0.056 0.064 0.075 0.083 0.055 0.065 0.076 0.082
25 0.042 0.045 0.050 0.057 0.047 0.052 0.058 0.063
50 0.030 0.031 0.033 0.038 0.035 0.037 0.042 0.046

Table A.2
λvap for R p = 50 and R p = 100.
R∗0 C50-90 C50-100 C50-110 C50-120 C100-90 C100-100 C100-110 C100-120

1 – 0.058 0.075 0.082 – – 0.073 0.081

5 0.053 0.064 0.072 0.074 0.052 0.058 0.071 0.073
10 0.054 0.065 0.074 0.076 0.054 0.059 0.073 0.075
25 0.050 0.055 0.059 0.064 0.051 0.054 0.060 0.064
50 0.039 0.041 0.045 0.049 0.039 0.041 0.045 0.049

Table A.3
cent for R p = 5 and R p = 10.

R∗0 C5-90 C5-100 C5-110 C5-120 C10-90 C10-100 C10-110 C10-120

1 3793.1 907.02 278.16 119.15 4278.8 1012.0 333.72 139.34

5 4447.5 1114.60 424.96 193.12 4844.3 1340.4 476.09 231.92
10 4529.4 1250.74 495.66 224.82 4796.2 1375.0 503.63 256.33
25 5611.6 1886.19 747.02 350.78 5806.8 1881.2 724.47 340.54
50 8185.8 2746.51 1207.39 553.95 8152.1 2554.7 1085.46 507.25

Table A.4
cent for R p = 50 and R p = 100.

R∗0 C50-90 C50-100 C50-110 C50-120 C100-90 C100-100 C100-110 C100-120

1 – – 374.95 – – – 386.42 –
5 5261.9 1480.5 555.79 296.51 5369.6 1659.3 562.68 303.33
10 5193.0 1467.8 576.02 288.19 5159.7 1639.1 585.23 294.85
25 5683.7 1825.8 721.44 368.89 5627.7 1883.7 768.68 372.18
50 7667.4 2603.3 1032.56 524.86 7684.8 2612.0 1036.04 531.14

Table A.5
Cδ for R p = 5 and R p = 10.

R∗0 C5-90 C5-100 C5-110 C5-120 C10-90 C10-100 C10-110 C10-120

1 0.993 1.074 1.288 1.826 1.367 1.495 1.774 2.620

5 0.175 0.221 0.305 0.437 0.218 0.281 0.400 0.607
10 0.088 0.121 0.174 0.248 0.103 0.141 0.208 0.307
25 0.050 0.074 0.102 0.139 0.052 0.081 0.113 0.162
50 0.037 0.055 0.076 0.103 0.039 0.057 0.080 0.113
 D. Dietzel, T. Hitz and C.-D. Munz et al. / International Journal of Multiphase Flow 116 (2019) 250–269 267

Table A.6
Cδ for R p = 50 and R p = 100.

R∗0 C50-90 C50-100 C50-110 C50-120 C100-90 C100-100 C100-110 C100-120

1 1.750 1.942 2.367 3.524 1.801 2.336 2.454 3.640

5 0.289 0.374 0.514 0.810 0.299 0.419 0.525 0.845
10 0.130 0.178 0.246 0.389 0.130 0.190 0.255 0.397
25 0.062 0.089 0.127 0.190 0.062 0.092 0.128 0.193
50 0.041 0.061 0.087 0.129 0.042 0.062 0.088 0.132

Appendix B. Dependency of cent and Cδ on initial conditions
and characteristic numbers
Fig. B.1 depicts the dependencies of the coefficient cent in the
KR on R∗0 , Rp , T0 , Ja, R+ and t + . Eq. (7) shows that the mass flux
scales inversely proportional to cent and similar trends as observed
for the HKR would be expected for 1/cent .
• Dependency of cent on R∗0 (Fig. B.1(a)): With increasing R∗0 , the
model coefficient increases such that the asymptotic mass flux
decreases. As discussed in Section 5.3 this results from the tran-
sient decrease of the reference mass flux with time/radius.
• Dependency on Rp (Fig. B.1(b)): The model coefficient depends
only weakly on the superheat ratio Rp . Again, the trends and
conclusions are similar to those of the HKR in Section 5.3 con-
sidering the inverse proportionality of the mass flux on cent .
• Dependency on T0 (Fig. B.1(c)): We observe a decrease of the
coefficient with increasing temperature if R∗0 and Rp are fixed.
As observed for the HKR, the model coefficient scales such that

Fig. B.1. Vaporisation coefficients cent as a function of different characteristic quantities of our investigation.
268 D. Dietzel, T. Hitz and C.-D. Munz et al. / International Journal of Multiphase Flow 116 (2019) 250–269
Fig. B.2. Vaporisation coefficients Cδ as a function of different characteristic quantities of our investigation.
vaporisation is enhanced for an increase in temperature. How-
ever, the entropy as the key quantity in Eq. (7) also changes
with T0 and a direct correlation is not apparent.
Fig. B.2 shows the resulting correlations for the SHFM. In gen-
eral, an increase in Cδ yields a decrease in the mass flux. However,
the SHFM yields a similar transient behaviour as the reference so-
lution and the correlations of Cδ in Fig. B.2 deviate from those ob-
tained for the HKR and the KR, respectively. This is a consequence
of the R-dependency of the mass flux given by Eqs. (8)–(10).
• Dependency of Cδ on R∗0 (Fig. B.2(a)):Eqs. (9) and (10) show that
the mass flux decreases with increasing R∗0 . The high coeffi-
cients observed in Fig. B.2(a) at small R∗0 avoid an overpredic-
tion of the mass flux when the radii and approximated bound-
ary layer thicknesses are rather small.
• Dependencies on Rp and T0 (Fig. B.2(b) and B.2(c)): The trends
Cδ = f(R p ) and Cδ = f(T0 ) are explained by the same physical
effects. Fig. B.2(b) and B.2(c) show that Cδ increases with Rp
and T0 , respectively. This is particularly true for small R∗0 (solid
lines). Both, for increasing T0 and increasing Rp , the temper-
ature difference T = [Tliq − Tsat ( p∞ )] increases. According to
Fig. 13, the pressure in the bubble rapidly approaches the am-
bient pressure such that T = [Tliq − Tsat ( p∞ )] reasonably ap-
proximates the temperature difference in Eqs. (9). Thus, an in-
crease in T0 and Rp would result in an increase of m˙  . At
the same time, the bubble radius and accordingly the bound-
ary layer estimation decrease with increasing T0 and Rp which,
again, allows for higher mass fluxes. The increase of Cδ avoids
nonphysically high mass fluxes.
Both, the KR and the SHFM do not reveal any clear functional
dependence of the coefficients on the characteristic parameters R+ ,
t + and Ja.
References
Avdeev, A., Zudin, Y., 2005. Inertial-thermal governed vapor bubble growth in highly
superheated liquid. Heat Mass Tranf. 41, 855–863.
 D. Dietzel, T. Hitz and C.-D. Munz et al. / International Journal of Multiphase Flow 116 (2019) 250–269
Bar-Kohany, T., Levy, M., 2016. State-of-the-art review of flash-boiling atomization.
Atom. Sprays 26 (12), 1259–1305.
Bell, I., Wronski, J., Quoilin, S., Lemort, V., 2014. Pure and pseudo-pure fluid ther-
mophysical property evaluation and the open-source thermophysical property
library coolprop. Ind. Eng. Chem. Res. 53, 2498–2508.
Brennen, C.E., 1995. Cavitation and Bubble Dynamics. Oxford University Press.
Chang, C.-H., Deng, X., Theofanous, T.G., 2013. Direct numerical simulation of inter-
facial instabilities: a consistent, conservative, all-speed, sharp interface method.
J. Comput. Phys. 242, 946–990.
Fechter, S., Jaegle, F., Schleper, V., 2013. Exact and approximate Riemann solvers at
phase boundaries. Comput. Fluids 75, 112–126.
Fechter, S., Munz, C., Rohde, C., Zeiler, C., 2017. A sharp interface method for com-
pressible liquid-vapor flow with phase transition and surface tension. J. Comput.
Phys. 336, 347–374.
Fechter, S., Munz, C., Rohde, C., Zeiler, C., 2018. Approximative Riemann solver for
compressible liquid vapor flow with phase transition and surface tension. Com-
put. Fluids 169, 169–185.
Fechter, S., Munz, C.D., 2015. A discontinuous Galerkin-based sharp-interface
method to simulate three-dimensional compressible two-phase flow. Int. J. Nu-
mer. Methods Fluids 78, 413–435.
Fedkiw, R., Aslam, T., Merriman, B., Osher, S., 1999. A non-oscillatory Eulerian ap-
proach to interfaces in multidimensional flows (the ghost fluid method). J. Com-
put. Phys. 152, 457–492.
Harten, A., Lax, P.D., van Leer, B., 1983. On upstream differencing and Godunov-type
schemes for hyperbolic conservation laws. Soc. Ind. Appl. Math. 25 (1), 35–61.
Hertz, H., 1882. Ueber die verdunstung der flüssigkeiten, insbesondere quecksilber.
Ann. Phys. Chem. 253, 177–193.
Hindenlang, F., Gassner, G.J., Altmann, C., Beck, A., Staudenmaier, M., Munz, C.D.,
2012. Explicit discontinuous Galerkin methods for unsteady problems. Comput.
Fluids 61, 86–93.
Houim, R., Kuo, K., 2013. A ghost fluid method for compressible reacting flows with
phase change. J. Comput. Phys. 85, 865–900.
Hu, X., Adams, N., Iaccarino, G., 2009. On the HLLC Riemann solver for interface
interaction in compressible multi-fluid flow. J. Comput. Phys. 228, 6572–6589.
Knudsen, M., 1950. Kinetic Theory of Gases. London Methuen.
Kopriva, D.A., 2009. Implementing Spectral Methods for Partial Differential Equa-
tions. Springer Science + Business Media B.V..
Lamanna, G., Kamoun, H., Weigand, B., Manfletti, C., Rees, A., Sender, J., Os-
chwald, M., Steelant, J., 2015. Flashing behavior of rocket engine propellants.
Atom. Sprays 25 (10), 837–856.
Lamanna, G., Kamoun, H., Weigand, B., Steelant, J., 2014. Towards a unified treat-
ment of fully flashing sprays. Int. J. Multiph. Flow 58, 168–184.
Lauer, E., Hu, X., Hickel, S., Adams, N., 2012. Numerical modelling and investigation
of symmetric and asymmetric cavitation bubble dynamics. Comput. Fluids 69,
1–19.
Lee, H.S., Herman Merte, J., 1996. Spherical vapor bubble growth in uniformly su-
perheated liquids. Int. J. Heat Mass Transf. 39, 2427–2447.
269
Lee, J., Son, G., 2017. A sharp-interface level-set method for compressible bubble
growth with phase change. Int. Commun. Heat Mass Transf. 86, 1–11.
Manfletti, C., 2014. Laser ignition of an experimental cryogenic reaction and control
thruster: pre-ignition conditions. J. Propul. Power 30 (4), 925–933.
Mikic, B., Rohsenow, W., Griffith, P., 1970. On bubble growth rates. Int. J. Heat Mass
Transf. 13 (4), 657–666.
Persad, A.H., Ward, C.A., 2016. Expressions for the evaporation and condensation
coefficients in the Hertz–Knudsen relation. Chem. Rev. 116, 7727–7767.
Plesset, M., 1949. The dynamics of cavitation bubbles. J. Appl. Mech. 16, 277–282.
Plesset, M., Zwick, S., 1954. The growth of vapor bubbles in superheated liquids. J.
Appl. Phys. 25 (4), 493–500.
Plesset, M.S., Properetti, A., 1977. Bubble dynamics and cavitation. Annu. Rev. Fluid
Mech. 9, 145–185.
Prosperetti, A., 2017. Vapor bubbles. Annu. Rev. Fluid Mech. 49, 221–248.
Prosperetti, A., Plesset, M.S., 1978. Vapour-bubble growth in a superheted liquid. J.
Fluid Mech. 85, 349–368.
Rayleigh, L., 1917. On the pressure developed in a liquid during the collapse of a
spherical cavity. Philos. Mag. J. Sci. 34, 94–98.
Robinson, A., Judd, R., 2004. The dynamics of sperical bubble growth. Int. J. Heat
Mass Tranf. 47, 5101–5113.
Saurel, R., Boivin, P., Metayer, O.L., 2016. A general formulation for cavitating, boiling
and evaporating flows. Comput. Fluids 128, 53–64.
Saurel, R., Petitpas, F., Abgrall, R., 2008. Modelling phase transition in metastable
liquids: application to cavitating and flashing flows. J. Fluid Mech. 607, 313–350.
Sazhin, S., 2014. Droplets and Sprays. Springer.
Schrage, R.W., 1953. A Theoretical Study of Interphase Mass Transfer. Columbia Uni-
versity Press, New York.
Sher, E., Bar-Kohany, T., Rashkovan, A., 2008. Flah-boiling atomization. Prog. Energy
Combust. Sci. 34, 417–439.
Shusser, M., Ytrehus, T., Weihs, D., 20 0 0. Kinetic theory analysis of explosive boiling
of liquid droplet. Fluid Dyn. Res. 27, 353–365.
Sonntag, M., Munz, C.D., 2014. Shock capturing for discontinuous Galerkin methods
using finite volume subcells. In: Fuhrmann, J., Ohlberger, M., Rohde, C. (Eds.),
Finite Volumes for Complex Applications VII-Elliptic, Parabolic and Hyperbolic
Problems. Springer International Publishing, Cham, pp. 945–953.
Sussman, M., Smereka, P., Osher, S., 1994. A level set approach for computing solu-
tions to incompressible two-phase flow. J. Comput. Phys. 114, 146–159.
Tanaka, N., Furkawa, K., Suemori, S., Matsuo, T., Nishida, M., Yasutake, A., 2011. The
greening of spacecraft reaction control systems. Mitsubishi Heavy Ind. Tech. Rev.
48 (4), 44–50.
Theofanous, T., Patel, P., 1976. Universal relations for bubble growth. Int. J. Heat
Mass Tranf. 19 (4), 425–429.
Toro, E.F., 2009. Riemann Solvers and Numerical Methods for Fluid Dynamics.
Springer-Verlag, Berlin Heidelberg.
Zein, A., Hantke, M., Warnecke, G., 2010. Modeling phase transition for compressible
two-phase flows applied to metastable liquids. J. Comput. Phys. 229, 2964–2998.