Toward Animating Water with Complex Acoustic Bubbles

Timothy R. Langlois Changxi Zheng Doug L. James

Cornell University Columbia University Stanford University

Frequency Deviation Vibration Magnitude

Figure 1: Complex acoustic bubbles: Our system is able to capture complex frequency effects due to bubbles’ shapes and positions. (Left)
Bubbles are colored blue/red if they are lower/higher than the theoretical Minnaert frequency for spherical bubbles, and depicts pitch rise
near the surface. (Right) Bubbles are colored (blue/red) based on their (small/large) vibration magnitude.

Abstract 1 Introduction
This paper explores methods for synthesizing physics-based bub- Liquids, and the sounds they make, are pervasive in our daily lives.
ble sounds directly from two-phase incompressible simulations of Whether it be a dripping faucet, a babbling brook, or your last glass
bubbly water flows. By tracking fluid-air interface geometry, we of water, you have most likely heard sounds generated by fluids re-
identify bubble geometry and topological changes due to splitting, cently. While there has been significant work on understanding how
merging and popping. A novel capacitance-based method is pro- fluids generate sound using bubbles, and breakthroughs in the visual
posed that can estimate volume-mode bubble frequency changes simulation of water, there are no existing methods for computing a
due to bubble size, shape, and proximity to solid and air interfaces. realistic audiovisual simulation of water with quality anywhere com-
Our acoustic transfer model is able to capture cavity resonance ef- parable to the purely visual component. For instance, current fluid
fects due to near-field geometry, and we also propose a fast precom- sound approaches do not even simulate acoustic bubbles realistically
puted bubble-plane model for cheap transfer evaluation. In addition, by modeling their evolution using two-phase liquid simulations, but
we consider a bubble forcing model that better accounts for bubble instead rely on single-phase flow solvers with ad hoc point-like
entrainment, splitting, and merging events, as well as a Helmholtz bubble generation techniques [Zheng and James 2009; Moss et al.
resonator model for bubble popping sounds. To overcome frequency 2010]. Such approximations are naturally much cheaper to compute,
bandwidth limitations associated with coarse resolution fluid grids, but, unfortunately, they have limited predictive value and ultimately
we simulate micro-bubbles in the audio domain using a power-law limited realism. In contrast, we seek to understand whether one can
model of bubble populations. Finally, we present several detailed ex- simulate bubbly flows with sound from first physical principles, and
amples of audiovisual water simulations and physical experiments what modeling challenges and trade-offs must be addressed.
to validate our frequency model.
In this paper, we explore a family of methods for sonifying detailed
two-phase liquid animations such as the one shown in Figure 1.
Keywords: acoustic bubbles, sound synthesis, acoustic transfer,
Given the complexity of liquid sound generation processes, there
fluid animation
are many details we consider (see Figure 2 for an overview of our ap-
Concepts: •Computing methodologies → Physical simulation; proach). Our approach begins with detailed multi-scale simulation
Continuous simulation; Massively parallel and high-performance of two-phase incompressible flow to resolve fine bubble geometry
simulations; needed for higher frequency sounds1 . Accurate modeling of surface
tension is needed to resolve bubble pinch-off and subsequent topol-
Permission to make digital or hard copies of all or part of this work for ogy changes. Individual bubble geometry is estimated and tracked,
personal or classroom use is granted without fee provided that copies are and we resolve bubble entrainment, merging, splitting and popping
not made or distributed for profit or commercial advantage and that copies processes at sub-ms time scales, and sub-mm length scales.
bear this notice and the full citation on the first page. Copyrights for com-
ponents of this work owned by others than the author(s) must be honored. Since the fluid flow is treated as incompressible for performance
Abstracting with credit is permitted. To copy otherwise, or republish, to post
on servers or to redistribute to lists, requires prior specific permission and/or ISBN: 978-1-4503-4279-7/16/07
a fee. Request permissions from [email protected]. c 2016 Copyright DOI: http://dx.doi.org/10.1145/2897824.2925904
1 Recall that Minnaert’s frequency model predicts f ≈ 6.52/d
held by the owner/author(s). Publication rights licensed to ACM. mm kHz
SIGGRAPH 2016 Technical Paper, July 24-28, 2016, Anaheim, CA (at STP) where dmm is the bubble diameter in mm.
 Fluid geometry Frequency Radiation Sound geometry affect its frequency, and by connecting frequency with
Simulation Solver frequency Solver Synthesis capacitance went on to compute frequencies for sphere-plane and
transfer
bubble tracking ellipsoid bubble geometries. Spratt et. al. [2015] verified this model
for simple nonspherical bubble geometries against a full acoustic
Figure 2: Overview of our system: From the fluid simulator, our scattering computation. We further extend this approach to support
method requires fluid surface geometry at each timestep, as well as frequency computations with general nonspherical bubble systems
bubble correspondences between timesteps. With this geometry, we and arbitrary air and solid interfaces.
compute each bubble’s frequency and acoustic transfer magnitude,
which are used to synthesize the bubble’s sound. Acoustic bubbles have been studied intensively, due to the impact
bubbles have on physical processes, and the necessity of passive
acoustic sensing of processes which are difficult to observe visually–
Leighton’s monumental work [1994] provides a definitive sum-
reasons, all acoustic fluctuations of the bubble/fluid/air system are mary. Acoustic bubbles represent a significant portion of ambient
handled using reduced-order vibration models. Toward this end, we ocean noise [Prosperetti 1988a; Prosperetti 1988b], and contribute
propose a method for detailed bubble frequency analysis based on to gas exchange between the ocean and the air and affect ocean
a “bubble capacitance” interpretation. Bubble pitch changes are per- albedo [Deane and Stokes 2010], which in turn affects climate [Bigg
ceptually important, but previous methods just model them using a et al. 2003; Leighton 1994]. They are important for “accurate quan-
parametric “chirp” as the bubble approaches the surface. In contrast, tification of a number of dynamic processes at the air-sea bound-
our method can resolve complex pitch shifts due to nonspherical ary, such as wave energy dissipation, gas exhange rates and the na-
bubble shapes, as well as nearby solid and air interfaces that can ture of rain” [Klusek and Lisimenka 2013; Longuet-Higgins 1990;
result in dramatic pitch decreases and increases, respectively (see Pumphery et al. 1989], such as using passive acoustic remote sens-
Figure 1). We estimate the bubble frequency using a boundary el- ing [Nystuen 1986; Ding and Farmer 1994; Means and Heitmeyer
ement method, and since we analyze many bubbles, we propose a 2002; Wilson and Makris 2008] and improvements (hopefully) in
fast amortized matrix solver which effectively exploits inter-bubble predictive computational models. Acoustic bubbles are even used by
similarities between matrix form factors and dense matrix solves. humpback whales during bubble net hunting [Leighton et al. 2007]!
To estimate a bubble’s sound pressure at a listener position, we must Fluid simulation has been a success story of computational physics,
faithfully estimate the surface-to-air acoustic transfer. To do so, we with widespread application in computer graphics and animation.
perform a standard frequency-domain boundary element analysis, However, most fluid simulators in graphics have focussed on single-
with surface vibration data input provided by our bubble-frequency phase free-surface flow [Stam 1999; Enright et al. 2002; Osher and
solver. Unlike previous methods, this accounts for the near-field Fedkiw 2006; Bridson 2008], and have only more recently tack-
scene geometry to capture container resonance effects, such as the led two-phase flow simulation [Hong and Kim 2005; Losasso et al.
characteristic rising pitch of a container being filled up with water. 2006; Boyd and Bridson 2012; Ando et al. 2015]. These works
For cheap, low-accuracy previews, we also propose a greatly simpli- almost exclusively aim for visually plausible simulations with large
fied bubble-plane transfer model, based on precomputing a lookup timesteps, and not for acoustic bubble computations. Because of
table indexed by the size and depth of proxy bubbles. their visual richness, a variety of methods have been developed
During the final sound synthesis phase, we simulate the bubble oscil- for simulating air bubbles [Hong and Kim 2003; Hong and Kim
lators using the estimated time-varying frequency. Bubble forcing 2005], and works range from tiny bubbles [Busaryev et al. 2012]
models are devised to account for bubble excitations during sur- to large volumes of bubbles [Zheng et al. 2009], and complex thin-
face entrainment, but also subsequent splitting and merging events. film interfaces [Da et al. 2015]. These methods lack the ability to
While prior works considered only Laplace pressure-jump forcing track individual acoustic bubbles while preserving their volumes
at entrainment, we leverage improved models of detailed surface- accurately—for example, Kim et al. [2007] artificially inflated bub-
tension modeling (recently proposed by Deane and Czerski [2008; bles to compensate for their volume change. However, bubble vol-
2010; 2011; 2011]). To model bubble popping sounds, we describe ume contributes significantly toward their estimated frequency, as
a Helmholtz resonator model. demonstrated in our work. Consequently, we have built upon Ger-
ris [Popinet 2003; Popinet 2009], a finite-volume-based multigrid
Since the fluid simulator’s spatial resolution restricts bubble sizes, solver used in computational fluid dynamics, to more accurately
our synthesized sounds are inherently frequency band-limited. simulate two-phase flows and track bubbles.
Therefore we propose a bandwidth extension scheme that performs
audio-domain simulation of micro-bubbles based on power-law Fluid sound simulation: Despite the importance of bubble sounds,
models of micro-bubble populations in breaking waves. relatively little work has been done on simulating them. Van den
Doel [2005] proposed a statistical method to generate bubble sounds.
Finally, our results include multiple examples of dripping, pouring, More recently, two works [Zheng and James 2009; Moss et al. 2010]
and splashing phenomena, as well as results from laboratory experi- have proposed more physically based methods. However, due to
ments to validate our capacitance-based frequency model. the sheer computational difficulty of predictive modeling of bubble
entrainment processes, these methods have relied on single-phase
liquid simulators with ad hoc stochastic models to estimate point-
2 Related Work like bubble creation rates and size distributions, with the unfortunate
consequence that bubble creation rates are either unrealistic [Moss
Fluid sound can be generated from a variety of sources, including et al. 2010] or must be laboriously hand-tuned for examples [Zheng
vortex based sounds [Howe 2002], fluid-structure interaction [Howe and James 2009]. Furthermore, these single-phase flows solvers
1998], shock waves, bubble popping [Deane 2013], drop im- can not estimate realistic nonspherical bubbles, and they lack re-
pact [Howe and Hagen 2011], and, the focus of this paper, harmonic alistic time-varying bubble frequencies. Additionally, Zheng and
bubble vibrations. Investigation of sounds produced by bubbles James [2009] use an acoustic transfer approximation which cannot
dates back almost a century, to the work of Lord Rayleigh [1917] capture scattering effects from enclosing solid interface, while Moss
and William Bragg [1920]. Minnaert calculated the frequency of et al. [2010] ignore acoustic radiation entirely.
isolated, spherical, harmonically vibrating bubbles [1933]. Stras-
berg [1953] described how a bubble’s shape and the surrounding Multi-frequency vibration: For small acoustic bubbles, it is usu-
 Creation Advection Splitting Merging Collapse

Figure 3: Bubble Tracking: We use colors to denote different bubble ids bi . There are five types of bubble actions between timesteps. The
left figure in each column denotes timestep i, and the right figure denotes timestep i + 1. Creation or entrainment: no bi overlap with bi+1 .
Advection: bi+1 overlaps with one value of bi . Splitting: one bubble id bi overlaps with ≥ 2 different bubble ids, bi+1 . Merging: two different
bi ’s overlap with one bi+1 . Collapse: no bi+1 overlaps with a bi .

ally assumed that they vibrate at a single frequency. However, it is fluid and defines the density and viscosity as
possible that other vibration modes can radiate sound at different fre-
quencies. Lamb [1932] investigated higher-order vibration modes ρ(c) = cρ1 + (1 − c)ρ2
using a first-order perturbation analysis, and concluded that they do µ(c) = cµ1 + (1 − c)µ2
not radiate efficiently enough to be important. A series of papers by
Longuet-Higgins [1989a; 1989b; 1991] showed that in certain situa- The advection equation for the density (2) is then replaced with an
tions, resonant coupling can occur between shape modes and the vol- equivalent equation for the volume fraction
ume mode, and higher vibration modes can radiate as a monopole, ∂t c + ∇ · (cu) = 0.
causing a single bubble to produce sound at several frequencies.
The recent work in graphics by [Moss et al. 2010] proposed using We use the open-source Gerris solver [Popinet 2003; Popinet 2009],
a related multi-frequency zonal-harmonic vibration model. While which is based on a finite volume method implementation with an
multi-frequency radiation has been observed in certain specific ex- adaptive octree data structure, parallel multigrid Poisson solver, and
perimental conditions [Chicharro and Vazquez 2014], it is believed an accurate surface tension model. Gerris enables high-fidelity sim-
that under real-world forcing conditions these coupling effects are ulation of bubble shapes and motion, with good volume preservation
less important [Longuet-Higgins 1991], and arguments have been properties, as well as realistic bubble formation from surface entrain-
given that shape-mode coupling accounts for a very small portion ment. However, despite its merits, there are significant spatial and
of the total air-domain sound radiation of bubbles [Longuet-Higgins temporal resolution requirements in order to resolve bubble entrain-
1990]. Longuet-Higgins himself also mentions that his model is ment and subsequent topological events, and thin interfaces. In our
not valid for realistic pinch-off scenarios, which would require fully simulations, we use spatial resolutions on the order of [0.625mm,
nonlinear equations [Longuet-Higgins 1989b]. 5mm], timesteps (determined by CFL conditions) on the order of
[30µs, 250µs], and spatial adaptation near interfaces. Nevertheless,
Experimental evidence to support multi-frequency bubble sounds is our ability to resolve tiny bubbles below 1 mm is limited. Also,
also lacking. Medwin and Beaky [1989] analyzed a large number Gerris has no way to preserve thin films when bubbles come into
of cases of single- and multi-bubble events (over 2000) generated contact with each other or the air surface, so merging events are over
in a wave tank. They laboriously classified four types of bubble estimated and surface popping is immediate.
radiation. The large majority of events show no signs of multiple
frequencies. The one type of event (“type D”) that does show multi- 4 Bubble Identification and Tracking
ple frequencies is speculated to be caused by interference from two
bubbles. Therefore, while it is possible for a single bubble to radiate
The link between the fluid simulation and our sound pipeline is the
at multiple frequencies, it seems to be rare, and we ignore it in the
interface geometry, consisting of the surface of each bubble and
present study.
the enclosing fluid volume. We describe here how this is imple-
mented in Gerris, but note that any fluid simulation method which
3 Fluid Simulation can provide this geometry will be compatible with our method.
Our representation of the interface geometry is given by c on the
Due to the dependence of a bubble’s frequency on its size, shape, octree grid. Specifically, bubbles are connected components of cells
and position, accurate fluid simulation is necessary for realistic where c < 1, which are identified and uniquely numbered with a
sound. We solve the incompressible, variable-density Navier-Stokes flood fill algorithm. To track bubbles between time steps, a new
equations with surface tension. variable bi is defined at timestep i and initialized to the number
of each bubble. The bi variable is advected between time steps.
ρ (∂t u + u · ∇u) = −∇p + ∇ · (2µD) + σκδs n (1) Overlaps between the current bi+1 and the previous bi advected
∂t ρ + ∇ · (ρu) = 0 (2) from the last timestep are used to correlate the bubbles between
time steps. There are several situations:
∇·u=0 (3)
1. New bubble: No bi touching bi+1
with u the fluid velocity, ρ the fluid density, µ the dynamic viscosity, 2. Advected bubble: One value of bi touching one value of bi+1
and D the deformation tensor defined as Dij = (∂i uj + ∂j ui )/2. 3. Split bubble: Two or more values of bi+1 overlap one of bi
The surface tension force is only nonzero on the interface, which the 4. Merged bubbles: Two or more values of bi overlap one of bi+1
Dirac distribution δs signifies. σ is the surface tension coefficient, 5. Collapsed bubble: No bi+1 touching a value of bi
κ is the surface curvature (not related to our κ in equation (4)), and
n is the surface normal. These are illustrated in Figure 3. The CFL condition ensures that in-
terface fragments do not move more than one grid cell per timestep,
For two-phase flows, densities ρ1 , ρ2 and viscosities µ1 , µ2 are and we did not observe any bubbles getting lost during tracking.
given for the first and second fluids respectively. The continuous Marching cubes [Lewiner et al. 2003] is used to generate interface
volume fraction field c(x, t) is used to denote the volume of the first geometry for each bubble, and for the enclosing fluid volume.
 Frequency
5 Bubble Frequency Estimation (Hz)
1440

After a brief introduction to bubble vibrations (§5.1), we describe a

new model for estimating the instantaneous bubble frequency (§5.2,
1280
§5.3), and an efficient algorithm that amortizes computation across
many bubbles (§5.4). k

8cm
1040
5.1 Bubble Basics

The equations of spherical bubble vibration were originally pro-

880
posed by Minnaert [1933] and are described in detail in Chapter 3
of [Leighton 1994]. Briefly, when bubbles are formed, they vibrate,
creating pressure waves which travel through the fluid and then pass
through the fluid-air interface into the air. The simple harmonic 720
oscillator model is similar to a linear spring-damper-mass system, 8cm
where the spring forces are provided by internal gas pressure and Figure 4: Spatially varying bubble frequency (in Hz) depicted
surface tension, and the mass is due to the surrounding liquid. We for a spherical bubble of 3mm radius (Minnaert frequency of ap-
use the “volume-pressure” frame, so that the infinitesimal volume proximately 1100 Hz). The pitch lowers as the bubble nears rigid
pulsation of a bubble v = V (t) − V0 (where V0 is the average walls (left, right, and bottom), and rises sharply as the bubble nears
volume) satisfies the fluid surface (top). Even in this small 8cm-by-8cm tank, the
mv̈ + αv̇ + κv = p(t) (4) bubble’s frequency can differ by over 700 Hz depending on position.
Note that for this figure, frequencies were only sampled 3.6mm from
where p(t) is a forcing term in units of pressure. The equation is the boundary (wall or surface), and extrapolated.
usually divided through by m and written as
p(t)
v̈ + 2β v̇ + ω 2 v = where γ is the polytropic index; we use γ = 1.4 for air. Evaluating
m
2 κ
the derivative (5) we obtain the bubble stiffness,
with ω = . For a bubble of equivalent radius r, β =
√ m
ωδ/ δ 2 + 4, with δ = δ(ω, r) = δrad + δvis + δth where γP0
κ= . (6)
√ 3γ−1
V0
ωr 4µ ψ − 3 − 3(γ−1)
δrad = δvis = δth = 2 The bubble air pressure P0 is the sum of the hydrostatic pressure
c ρωr2 ψ−4
(approximately Patm + ρgd for a bubble at depth d), and surface
16 Gth g 2σ
with ψ = 9(γ−1) 2 , c is the speed of sound in the fluid, µ tension ( R for a spherical bubble). To support pressure and sur-
ω 0
is the liquid’s shear viscosity, ρ is the fluid density, γ is the gas’ face tension for nonspherical bubbles in complex fluid domains, we
3γp
heat capacity ratio, Gth = 4πρDfg is the thermal damping constant obtain a reliable approximation of P0 from our two-phase fluid sim-
ulator. Observe that κ only depends on the internal gas and surface
at resonance, Dg is the gas’ thermal diffusivity, g is gravitational tension, which is consistent with Minnaert’s derivation [Leighton
acceleration, and pf is the hydrostatic pressure of the liquid. A full 1994]. Interestingly, note that κ is incapable of producing increasing
derivation is provided by Leighton [1994]. pitch as a bubble rises (it actually predicts the opposite dependence
Minnaert computed ω for a spherical bubble. However, nearby fluid- on d), and thus the mass factor is responsible for this effect.
air and fluid-solid surfaces, as well as bubble shape, can affect ω, The effective mass (m) can be derived similar to Strasberg [1953],
and are the reason for the familiar pitch shift as bubbles approach but with consideration for complex fluid domains. The key idea is
the fluid surface. In the following sections, we describe one of to equate the oscillator kinetic energy 12 mv̇ 2 , for a given harmonic
our main contributions: a method for accurately computing ω (by bubble volume velocity v̇, with the kinetic-energy volume integral of
computing κ and m) for complex geometries, and thereby providing the surrounding fluid, Wf , to determine m. Wf is also the work that
more realistic sound. the bubble does to the fluid. In order to compute the fluid kinetic
energy, we exploit the fact that bubbles are acoustically compact
5.2 Frequency Model low-frequency sources 2 , and thus the surrounding fluid’s vibration
is well approximated by incompressible, irrotational flow modeled
Extending Strasberg’s [1953] derivation, we seek to better estimate using Laplace’s equation. Specifically, let the surrounding fluid’s
the bubble frequency ω, by more accurately modeling the scene- particle velocity be ∇φ, where φ is the velocity potential which
specific effective stiffness, κ, and effective mass, m. We consider a satisfies
bubble’s infinitesmal volume pulsation v(t) in the volume-pressure
frame (4), in the undamped case (α = 0). ∇2 φ(x) = 0, x ∈ Ω,
The effective stiffness (κ) accounts for internal gas and surface subject to a pressure-like Dirichlet boundary condition on the bub-
tension effects, and is the simplest of the two values. The stiffness ble, φ = φb , and other suitable boundary conditions elsewhere
in the volume-pressure frame is the rate of change of bubble pressure (discussed later). Therefore we can express the effective mass as
with volume,
dp 2Wf
 Z
2 ρ

κ=− . (5) m= 2 = 2 (∇φ)2 dΩ .
dv v̇ v̇ 2 Ω
Assuming a polytropic gas law, the reference volume V0 and pres-
sure P0 are related to the modified volume V = V0 +v and pressure 2 Note that for a bubble at 1 atm pressure, we have f R ≈ 3m/s, or
P = P0 + p by ωR ≈ 19m/s, so that kR = ωr/c ≈ 0.013 << 1. Also note that
P0 V0γ = P V γ , λ/R ≈ 114.
 φ=0
Γa

∇2 φ = 0

Ω φ=1 (a) ω = 1 (b) ω = 1.14 (c) ω = 1.27 (d) ω = 1.72

Γb C=1 C = 1.31 C = 1.61 C = 2.96
∂n φ = 0 Figure 6: Capacitance-based frequency estimation for a rising
Γr bubble: We recover increasing “chirp-like” frequency and capaci-
Figure 5: Interior Laplace BVP for bubble capacitance: We use tance (normalized) as the bubble (initial radius R = 5.8 mm) nears
the solution’s ∂n φ gradient on the bubble boundary Γb to compute the surface (a-d). The rising pitch produced as the bubble’s wa-
the bubble capacitance using (9), and on the air boundary Γa to ter layer (lamella) thins corresponds to a thin-plate capacitor of
evaluate acoustic radiation (in §6). increasing thinness.

Converting the volume integral to a boundary integral using Green’s

first identity we obtain
Z Z Z
(∇φ)2 dΩ + φ∇2 φ dΩ = − φ ∂n φ dS,
Ω
| Ω {z } ∂Ω
(a) ω = 1 (b) ω = 0.95 (c) ω = 0.91 (d) ω = 0.85
≈0
t=0 t = 3ms t = 6ms t = 13ms
where the minus sign is due to normals pointing out of the fluid Figure 7: Shape-dependent bubble frequency is demonstrated
domain Ω (and into air or solid domains). Denoting the bubble, air, here for a simulated bubble undergoing natural shape changes. Fre-
and rigid surfaces as Γb , Γa , and Γr , the mass integral becomes quency is normalized. Our bubble frequency estimation method
Z Z Z  can resolve musical pitch fluctuations occuring on semi-tone mag-
ρ nitudes, on the order of 10ms.
m=− 2 φ ∂n φ dS + φ ∂n φ dS + φ ∂n φ dS
v̇ Γb Γa Γr

= mb + ma + mr .
condition (φb = 1), the fluid-air surface as a conductor at zero po-
The terms can be evaluated by making use of the boundary condi- tential (φa = 0), and rigid interfaces as insulators (∂n φr = 0). Our
tions for the generalized problem (see Figure 5). The air-interface generalized “bubble capacitor” boundary value problem (BVP) is
ma contribution is zero since φ = 0 on Γa , as the acoustic pressure shown in Figure 5.
p = −ρ ∂φ ∂t
= −iωρφ = 0 there. The rigid-surface mr contribu-
tion is also zero: the acoustic particle velocity must be zero in the Given the solution to Laplace’s equation φ̂ for this BVP, we compute
normal direction on the rigid boundary, and thus ∂n φ = 0 on Γr . the capacitance as
Therefore, the only mass contribution arises from the bubble term
m = mb . Z
1
C=− ∂n φ̂ dS. (9)
5.3 Capacitance Interpretation of Bubble Frequency 4π Γb

Since the velocity potential is constant on the surface of the bubble,

Since this formula only requires ∂n φ̂ on Γb , we can solve for
φ = φb , it can be taken outside the integral,
each bubble’s capacitance using a boundary integral formulation
ρ
Z
ρ
Z of Laplace’s equation. We discuss an optimized bubble capacitance
m=− 2 φ ∂n φ dS ≈ − 2 φb ∂n φ dS. solver in §5.4. Finally, our approach reproduces and generalizes the
v̇ Γb v̇ Γb frequency models of Minnaert and Strasberg, and supports nontriv-
R ial frequency estimation in complex fluid geometries (see Figure 6),
Then since the volume velocity is v̇ = − Γ ∂n φ dS, it follows and for complex bubble shape changes (see Figure 7).
b
that m = ρφb /v̇. The bubble is a uniform-potential conductor-like
surface, so Strasberg noticed that treating φ as electrostatic potential
and v̇ as flux, the ratio v̇/φ is mathematically equivalent to 4π times 5.4 Fast Amortized Solution of Capacitance BVP
the surface’s electrostatic capacitance, C. Therefore, the effective
mass can be identified as We now describe an efficient method for estimating bubble capac-
ρ itances that exploits common computations between the bubbles
m= . (7) for speed. We can use the boundary element method (BEM) to
4πC
solve the capacitance BVP using established codes for the interior
By combining equations (6) and (7), the bubble frequency can be Laplace problem with mixed boundary conditions (BCs) [Banerjee
given in terms of its capacitance, and Butterfield 1981].

κ 4πγP0 BEM BVP matrix structure: After discretizing the direct boundary
ω2 = = C. (8) integral equation formulation of the interior Laplace problem for
m ρV0
the velocity potential associated with a single bubble, we arrive at
the linear matrix problem,
Bubble capacitance BVP: To compute the capacitance we can in-
terpret the bubble as a conductor with a unit potential boundary Hφ = Gv,
 0.002
for fast evaluation of A−1 b for different bubbles. By exploiting
0.001 the Sherman-Morrison-Woodbury formula for U [Hager 1989], the
Relative Error

0.000 matrix inverse can be written as

0.001  
0.002 X Y
A−1 =
0.003 Z U
0.004 where
0.4
−1
0.2 X = Gbb − BD−1 C , (10)
Relative Error

0.0 Y = −XBD −1
, (11)
0.2 −1
U =D (I − CY ), and (12)
0.4
0.6 Z = −U CG−1
bb . (13)
1000 2000 3000 4000 5000
Frequency (Hz)
By carefully exploiting common subexpressions and cached LU fac-
Figure 8: Solver Error: For one timestep of the pouring faucet torization of the large D
block, the product x = A−1 b can be eval-
example, we meshed the domain at a high resolution (5mm maxi- uated in O (na + ns )2 flops for small bubbles of size nb = O(1).
mum edge length), and computed frequency and transfer for all the Please see Appendix A for explicit algorithmic details. In our exam-
bubbles. Although we use aggressive mesh simplification, the rela- ples, we observe that the amortized capacitance solver accelerates
tive errors for our frequency solver compared to the high resolution bubble frequency estimation by 10-12x for small bubbles, while for
results are very small (top). Pressure magnitude errors (bottom) are very large bubbles the speed-up may only be 3x, but still worthwhile.
tolerable in our range of interest, and increase at higher frequencies
as expected.
5.5 Adaptive Meshing

The generation of meshes used for BEM anal-

with the following block structure, ysis should be discussed briefly, since adap-
" # ! " # ! tive meshing is needed to keep our frequency
Hbb Hba Hbs φb Gbb Gba Gbs vb
solve times manageable. Detailed meshes
Hab Haa Has φa = Gab Gaa Gas va ,
Hsb Hsa Hss φs Gsb Gsa Gss vs
are obtained from the fluid simulator using
Marching Cubes, which we adaptively sim-
where φ is the BEM vector of potential values and v is the BEM plify using Quadric Error Simplification [Garland and Heckbert
vector of (outward) normal derivative values, ∂n∂φ
; here the three 1997]. However, we must properly resolve interfaces and inter-
boundary regions are denoted by b (bubble), a (air), and s (solid). surface gaps as bubbles approach the fluid surface or container walls
Applying the Capacitance BVP boundary conditions (see Figure 5), in order to compute accurate capacitance values. Therefore we use a
we arrive at the linear system Ax = b with the block structure, sizing function that ensures each triangle’s maximum edge length is
less than its distance to the nearest bubble. The inset shows an illus-
trative example. Although we use aggressive mesh simplification,
    
Gbb Gba Hbs vb Hbb
 Gab Gaa Has   va  = Hab  φb ,

our errors (shown Figure 8) are still tolerable.
Gsb Gsa Hss −φs Hsb
| {z } | {z } | {z } 6 Bubble Acoustic Transfer
A x b

where the unknowns are vb ∈ Rnb , va ∈ Rna , and φs ∈ Rns . Solv- We now describe two acoustic transfer solvers that estimate the pres-
sure amplitude at the listener’s position(s) from an acoustic bubble
R system provides three quantities: vb which helps approxi-
ing this
vibrating with unit pressure: (1) a BEM-based solver that includes
mate Γ ∂n φ̂ dS ≈ aTb vb and thus (9) for the bubble capacitance;
b scene geometry (§6.1), and (2) a fast but very approximate solver
va which describes the free surface vibration, and will be used for that uses a proxy bubble-plane transfer model (§6.2). In practice,
acoustic radiation modeling (in §6); and φs which is ignored since we sample frequency and bubble-to-ear transfer values at a fixed
it is not needed by our application. rate during the lifetime of the bubble; in our implementation these
In our implementation we use the BEM++ software library [Smigaj solves are done every 1ms, and interpolated with a cubic spline.
et al. 2012] to evaluate the various G and H matrix blocks; we used These transfer values are used for sound synthesis later in §8.
Galerkin discretization (which is more robust to meshing irregulari-
ties than collocation schemes) with constant elements for Neumann 6.1 BEM-based Acoustic Transfer Solver
data (on Γs ), and piecewise linear elements for Dirichlet data (on
Γb and Γa ). We now describe how to approximate realistic sound amplitudes
from an harmonically vibrating fluid surface. Since sound scattering
Fast amortized matrix solver: For a given frame, we solve the and resonance effects from external geometry (such as the walls of
matrix problem many times for many bubbles. A naive LU-based a glass container) can introduce perceptually important and pitch-
evaluation of x = A−1 b would require O (nb + na + ns )3 flops.


dependent amplitude variations (such as when a glass is filled up
If we denote the four blocks of A by with water) we seek to include near-field scene geometry in our
  transfer solver that has been neglected in prior works [Zheng and
Gbb B
A= , James 2009].
C D
Mathematically we approximate solutions to the exterior wave radi-
we observe that the huge lower-right “domain” submatrix D re- ation problem specified by the Helmholtz equation
lating the self-effect form factors for the air and solid boundaries
(Γa ∪ Γs ) is constant across problems. We can exploit this fact (∇2 + k2 )p(x) = 0, x ∈ Ωa ,
 ∂n φ = va ∇2 φ + ka2 φ = 0
Γa
Ωa

∂n φ = 0
Γr

Figure 9: Exterior Helmholtz BVP for acoustic radiation

Figure 11: Interpolating velocity BC data for a rising bubble in
a square container (top view): (Left) velocity BCs from the interior
where k = ω/ca , and ca is the speed of sound in air. As boundary
solve geometry (on Γa ) are (Right) interpolated onto the mesh (of
conditions, we impose a vibration BC on the fluid-air interface,
Γa ∪ Γr ) for the exterior Helmholtz BVP.
and a no-vibration BC on the rigid scene geometry (see Figure 9).
Following our bubble frequency solve, the vibration of the fluid-air
surface is known3 from ∂n φ̂ on Γa , which in turn becomes input
boundary data for the air-domain acoustic transfer estimation. At runtime, instead of solving the exterior BVP, we compute the
equivalent spherical radius of the bubble, and the distance to the
To support arbitrary scene geometry, we use established BEM fluid surface. These values are used to lookup the closest set of
solvers for the exterior Helmholtz radiation BVP. In our implemen- multipole coefficients in our database, which are evaluated at the
tation, we use the BEM++ implementation [Smigaj et al. 2012] A listening position. When performing lookups in our database, we
major practical task is generating a suitable scene mesh with BC first find the closest radius, then find the closest depth sample with
data at each time step. We first mesh the fluid and external container that radius. It took 2.2 hours to construct the database using 32 cores.
together to generate detailed geometry for the external radiation Runtime evaluation times for each example are given in Table 1.
problem (see Figure 11). Since these meshes are typically too de-
tailed for efficient BEM analysis, we again use adaptive meshing (as
in §5.5) to decimate the mesh. We further restrict the largest edge
7 Bubble Forcing
length so as to sufficiently resolve the smallest wavelength; in our
examples, we have used a 3cm limit which can resolve λa = 17.2cm There are several mechanisms which can drive bubble vibrations.
at the upper 2kHz range we simulate. Finally the interior capaci- Previous fluid sound work in graphics has used Laplace pressure
tance BVP and exterior transfer BVP may have different meshes for forcing, which approximates entrainment forcing by a pressure jump
the fluid-air interface, and therefore we interpolate the previously due to surface tension. The additional pressure, pσ = 2σ
R
(R is the
mean bubble radius), provides an initial impulse to the bubble os-
obtained velocity-like solution data, va (i.e., ∂n φ̂), from the interior
cillator (possibly smoothed in time). However, estimates of the
mesh to the mesh of the exterior fluid-air interface Γa . The pres-
Laplace pressure jump, as well as hydrostatic pressure and shape
sure BC on that interface is the same up to a scaling factor, since
ρ ρ mode coupling effects, show that they represent a minor (< 10%)
∂n pf = ρfa ∂n pa = ρfa ∂n φ̂. amount of the total forcing [Pumphrey and Ffowcs Williams 1990].
A recent set of papers from Deane and Czerski [2008; 2010; 2011;
6.2 Fast Bubble-Plane Proxy Transfer 2011] propose a family of models based on neck collapse (for en-
trainment and splitting events) and neck expansion (for merging
Transfer computations can be expensive for detailed water surfaces events) where surface tension effects account for the majority of
and scenes. For many applications and “fast preview” renderings, forcing (summarized in Figure 12).
we can use a cheap transfer model based on a simplified bubble- Entrainment: The forcing of bubbles released from an underwa-
plane geometry (see Figure 10). ter tube were analyzed in [Deane and Czerski 2008; Czerski and
To do this, we precompute a lookup ta- Deane 2010]. As the bubbles separate from the nozzle, a conical
ble of transfer values for spherical bubbles neck forms with very sharp curvature (causing high surface ten-
of various radii (r) at various depths (d) sion) at the tip. At separation, surface tension causes the neck to
beneath a planar water surface. We sam- rapidly shoot into the bubble, quickly decreasing the bubble’s vol-
ple bubble radii from 0.25mm to 1cm in ume, which forces oscillations. A similar situation happens during
0.25mm increments, and depths down to bubble entrainment: a neck is formed as a bubble separates from
8cm in 0.5mm increments. For each ra-
dius and depth, we solve the interior fre-
quency BVP and exterior radiation BVP. Figure 10: Fast proxy
Then we fit a multipole expansion [Zheng transfer model
and James 2010] to the bubble’s exterior radiation data, and only
store the multipole coefficients (as opposed to storing all the bound-
ary data). Further, we assume the bubble’s response is cylindrically
symmetrical, so we only store the zonal multipole coefficients.

3 We refer to the pressure gradient as a velocity, but they are merely pro-

portional: ∂n p = −iωρVn for a harmonic vibration of velocity Vn . Since

(a) (b) (c)
the normal velocity is continuous across the fluid-air interface, −iωVn =
∂n pa /ρa = ∂n pf /ρf , and so up to a multiplicative constant density factor Figure 12: Bubble forcing: There are 3 types of forcing events we
we can say that ∂n pa and ∂n pf are equal velocity-like quantities. model: (a) entrainment, (b) splitting, and (c) merging.
the fluid surface. This geometry is illustrated in Figure 12a. Their
time-dependent forcing model based on a conical neck is
p
9κσηpin,0 1 + η 2 2
f (R, t) = − t , (14) Figure 13: Bubble frequency extension model: (Left) a bubble
4ρR3
collapses before it has finished oscillating, resulting in an audible
where κ is the gas polytropic index, σ is the surface tension co- sample-and-hold frequency artifact. (Right) To improve the approx-
efficient, pin,0 is the equilibrium pressure in the bubble, R is the imation we extrapolate the frequency using a fitted exponential.
equilibrium radius, ρ is the fluid density, and η is the slope of the
neck cone (η = tan(θ)). We have multiplied their original equa-
tion by ρR2 to transform to the correct units (pressure). We sample method to integrate the oscillator equations, and did not need any
the neck angle uniformly between [20o , 55o ], which corresponds special treatment to handle abrupt frequency changes.
to η ∈ [0.36, 1.43]. It was observed that this forcing function is
valid for approximately 300-400 µs for 2mm bubbles (Minnaert fre- Simultaneous events: A single bubble can undergo multiple
quency = 1.6 kHz, period = 625 µs). For a bubble
with period τ , events during a single timestep of our simulation. For example,
we use this forcing function for the min 12 τ, 600µs , and set it to a near-surface bubble could split and one of the daughter bubbles
0 afterwards. could touch the surface and disappear. Similarly, a bubble could
be entrained and merge with another bubble during one time step.
Splitting: For bubble splitting events, the same mechanism is pro- To handle these volume changes and missed events correctly, we
posed in [Czerski and Deane 2011] as the main source of forcing. monitor each bubble’s volume during bubble tracking. If there is a
Both child bubbles are driven by the neck that forms during pinchoff sudden increase, we add a new bubble entrainment event and merge
(illustrated in Figure 12b). We again use

equation (14), and again it immediately. When there is a sudden volume decrease, we add a
assume it is active for min 12 τ, 600µs for each bubble. The same split event followed immediately by a collapse event.
sampling method for η is also used.
Merging: For bubble coalescence events, a similar mechanism is 8.1.1 Frequency Extension
proposed in [Czerski 2011]. When two bubbles merge, very strong
surface tension is generated at the merge point. This causes the For various reasons such as resolution limits or the idealized damp-
bubble surface to expand, rapidly increasing the volume of the new ing model, simulated bubbles may reach the fluid surface and col-
bubble (see Figure 12c). Their model is lapse before they are done oscillating. This singularity can limit
our ability to resolve the bubble’s chirp-like frequency response. To
6σκpin,0 2 avoid such artifacts, we fit a small exponential model, aect , to the
f (t, R) = t , (15)
ρR3 frequency samples, and use it to extrapolate the bubble’s frequency
in time (see Figure 13). Parameters a and c are calculated to ensure
where again we have multiplied their original equation by ρR2 to C0 and C1 continuity.
transform to pressure. The time that this forcing function is active
is labeled tlim , and is defined as the amount of time it takes for 8.1.2 Bubble Popping Sound Model
the expanding radius to reach a fixed fraction (sampled uniformly
between [0.4, 0.8]) of the smaller bubble radius. An arbitrary mod- Bubbles that “pop” at the surface have their oscillators die out, but
ulation function is also added, resulting in the forcing function in reality there is a characteristic chirp-like popping sound due to the
   small pressurized cavity that briefly rings like a so-called Helmholtz
1 1 t − tlim 6κσp0 2 resonator [Spiel 1992]. These sounds are quiet, but occupy a part
f (t, R) = − tan−1 3 t .
2 π tlim ρR3 of the audio spectrum which is otherwise quiet in our model due
to the predominantly lower pitch of bubbles resolvable by our fluid
We again limit the forcing to min (tlim , 600µs). simulator. The physical process of bubble bursting and aerosol gen-
eration is actually terribly complicated [Lhuissier and Villermaux
8 Sound Synthesis 2012]. We use a simplified model for the sound produced by burst-
ing bubbles proposed for remote acoustic sensing of important bub-
The last part of our pipeline (after fluid simulation, bubble identi- ble properties, such as bubble cap film thickness [Deane 2013].
fication and tracking, frequency estimation, and radiation analysis) The model gives the time dependent frequency of a bubble pop as
is to synthesize the resulting sound at the listening position. While (equation 5 in [Deane 2013])
conceptually similar to [Zheng and James 2009], there are several
different and important details in our approaches. s  
c 3π 2 ut
fH (t) = R sin , 0 ≤ t ≤ tmax ,
8.1 Audio Synthesis Details 2π 16V R

Culling silent bubbles: To avoid unnecessary frequency and radi- where u is the velocity of the retracting bubble film, R is the bubble
ation solves for inaudible bubbles, we cull silent bubbles. Specifi- cap radius, and V is the bubble volume. tmax is the time it takes
cally, we do not perform any solves for bubbles that are older than for the film to retract fully. The film retraction velocity can be
−ln(.01)/β, where β is estimated from the bubble’s equivalent estimated from the length of time the film has been draining for
spherical radius and Minnaert frequency. before it nucleates. We define the minimum drain time tmin as the
time it would take for the film thickness to reach R/10.
Oscillator tracking: To avoid discontinuities in the synthesized
sound, we continue oscillators through split and merge events. Dur- When a bubble reaches the surface in our simulation, we uniformly
ing split events, the parent bubble’s oscillator continues to the largest sample the drain time between [tmin , 20ms] to define u and syn-
child bubble. During merge events, the largest parent bubble’s os- thesize a cosine chirp with frequency fH (t). We modulate the pop
cillator continues to the child bubble. We use the standard RK4 sound with an ad-hoc function chosen to match experimental data
 Air Mic

Hydrophone

Figure 14: Bubble popping sounds add additional high-frequency

content as shown here by comparing spectrograms of sounds pro-
duced (Left) without and (Right) with the popping sound model Figure 15: Container effects: Container effects can be strong in
based on [Deane 2013]. The simulation example is the pouring the air (top), while in a simultaneous hydrophone recording (bot-
faucet. tom) the waveform of the same entrained bubble is much cleaner.
Resonances of the container can be seen as lines in the spectrum,
which continue after the bubble has popped.
from [Deane 2013], namely,
   
ln(.00001) 2 t Results were computed on a heterogeneous cluster of 31 nodes,
mod(t) = exp t atan .
tmax π 2tmax where each node’s core count ranged from 8 to 64 (Intel Xeon
X5355 and Xeon X7560 processors), as well as on the NSF clus-
The amplitude of the pop is not well defined. Given the maximum ter SuperMIC (two Intel Xeon E5-2680 processors per node). We
absolute value of the corresponding bubble sound smax , we choose report runtimes and simulation statistics in Table 1.
to scale the pop sound so that it has a maximum magnitude of
asmax , where a is uniformly sampled between [.001, .03]. The
effect of the popping sound model is shown in Figure 14. 9.1 Discussion of Tank Effects

8.1.3 Bubble Bandwidth-Extension Scheme Container effects such as reverberance can be significant. An exam-
ple is shown in Figure 15. When a single bubble is entrained in a
glass fish tank, strong echo can be seen in the air microphone signal.
Given the resolution limits of our fluid simulator, bubbles below
However, the sound recorded by a hydrophone of the same event
a certain length scale can not be resolved correctly, resulting in
looks much closer to a theoretical damped harmonic oscillator.
a band-limited frequency response of the bubble oscillator model.
Experimental studies of bubble populations in breaking waves have
established various bubble size statistics, and have shown that the 9.2 Validation
number of tiny bubbles tends to follow power-law models [Deane
and Stokes 2002]. To artificially extend the frequency response We performed several experiments to validate our frequency model.
of our renderings, we optionally seed audio-domain bubble events
from a power-law distribution as follows. Single bubble entrainment: We recorded several bubbles en-
trained by droplets, and simulated a similar entrainment case. We
We sample tiny bubbles in the audio domain, based on simulated are able to capture the characteristic frequency chirp of the bubble
larger bubbles. Specifically, for each entrained simulation bubble as it rises (see Figure 16).
with radius rparent ≥ 2mm, we assume that the impact which
created this bubble also generated other smaller bubbles with radii Underwater bubble creation: Using a syringe and plastic tubing,
rtiny ∈ [0.1mm, 1mm]. The number of artificial bubbles gen- we also released underwater bubbles and recorded their emissions.
erated for each simulation bubble is uniformly sampled between As the bubble moves away from the tube (rigid surface), there is
[0, 3000 ∗ rparent ]. Given the simulation bubble’s creation time t, a slight pitch increase, but we do not see the characteristic chirp
the start times for each of the artificial bubbles are uniformly sam- because the bubble finishes vibrating before it reaches the surface.
pled from [t−0.1, t] (because the tiny bubbles are created during the
impact). The sizes of the tiny bubbles are sampled from a -3/2 power
law, consistent with observations of the distribution of bubbles be-
low the Hinze scale. Finally, since we have no geometry or positions
for these artificial bubbles, we base their amplitude on the parent
bubble. Given the parent bubble’s transfer magnitude pparent , we
set the transfer value for a tiny bubble to 50 pparent (rtiny )1/3

8.2 Sound Synthesis Summary

In summary, our sound synthesis pipeline is very similar to previous

work, with the important details that we need to track oscillators
through split and merge events to avoid discontinuities, and simul-
taneous events need to be treated correctly. We proposed several
audio domain methods to help add missing detail from our simula-
tions, including frequency extension, a bubble popping model, and
a microbubble model. The three latter models are all optional.

9 Results and Discussion Figure 16: Single Bubble Entrainment: A simulated bubble en-
trainment event (top) produces a similar spectrum to a recorded
Please see our accompanying video for visual and audio results. entrainment event (bottom).
 Frequency Radiation Proxy
Example Domain Simulation time # of bubbles / solve time Amortized solve time evaluation
size (cm) Length (s) (hours / cores) # of solves (hours) speedup (hours) time (hours) % culled
Dripping Faucet 8 x 8 x 24 9.0 97 / 32 153 / 965 .005 0.81 .07 0.09 6.4
Pouring Faucet 8 x 8 x 24 8.5 402 / 64 331521 / 585311 23 7.35 52 3.8 71.2
Water Step 8 x 24 x 24 4.5 1000 / 96 420134 / 483654 20 4.85 44 20.0 71.8
Dam Break 16 x 16 x 32 2.64 394 / 64 114471 / 121646 2.9 4.72 15 0.66 72.4
Armadillo Drop 16 x 16 x 32 4.0 293 / 64 13981 / 11653 .245 3.82 3.4 0.58 87.9

Table 1: Results. Our fluid simulations used different numbers of cores, which are reported above. The frequency and radiation solves are
massively parallel, and were computed using 680 cores. Proxy transfer evaluation was done on a single core, but could be parallelized easily.

Figure 18: Dripping Faucet

Figure 17: Underwater bubble release: A bubble released from

an underwater tube shows a slight frequency rise as it moves away
from the rigid tube. The simulation (top) matches well with experi-
ment (bottom).

A simulated scenario of a bubble moving away from an underwater

tube produces a similar effect.

9.3 Large Results

Dripping faucet: Our dripping faucet example has a relatively

small number of bubbles. It clearly demonstrates the pitch rise as
bubbles approach the surface, as well as the benefit of our frequency
extension model. Figure 19: Dam Break

Dam break: There is a loud, low frequency sound produced by the

large tubular bubble in the dam break example. The sloshing sounds
in this example are convincing. The dam break clearly demonstrates
the benefit of our microbubble model, which adds higher frequency
texture.
Water step: The water step has the most bubbles of all our simula-
tions. While it has a fairly constant sound spectrum, the differences
of our multiple models can be seen. This example also demon-
strates the importance of transfer, as it sounds very dissonant when
rendered without transfer.
Pouring faucet: We use the same domain as the dripping faucet, but
this time use a constant 1cm radius stream to fill the container. This
example most clearly shows the importance of transfer, allowing us
to capture the characteristic pitch shift as the container fills.
Figure 20: Water Step
Armadillo drop: For our last example, we dropped a water shaped
armadillo into a pool of water. While this example is fun, it also
highlights some of the deficiencies of our system. Even with our
microbubble model, there is not much of an impact sound during droplet impacts, instead of on larger bubbles entrainment times, may
the initial impact of the armadillo. A microbubble model based on help.
 Figure 21: Pouring Faucet Figure 22: Armadillo

10 Conclusion ing asymptotic forcing information, but assume idealized geometry

and involve undetermined parameters. It is therefore interesting to
We have explored many stages of audiovisual fluid simulation, and explore data-driven approaches for automatically extracting forc-
identified the need for, and proposed, numerous sound simulation ing models and their parameters. Laboratory experiments reveal
models. We believe that increased resolution of liquid sound gener- a wide-range of surface related sound events which do not radiate
ation mechanisms will also lead to improved visual fidelity of fluids effectively underwater (as measured by a hydrophone), but that pro-
in computer animation. duce significant high-frequency sound contributions. Methods for
One surprising finding of our bubble frequency model was the large approximating these processes, either in the geometric domain or
variations predicted based on spatial proximity to boundaries (see audio domain, are necessary to improve sound realism. Finally,
Figure 4), whereas the nonspherical pitch variations were compar- there are numerous important potential applications of fluid-sound
atively more modest and less perceptually important for sound ren- synthesis in other fields of science and engineering that should be
dering. As Figure 7 demonstrates, there were some stong pitch explored.
variations due to shape, but these seemed rare. Perhaps the most im-
portant aspect of nonspherical bubbles was their ability to conform Acknowledgements
to the fluid-air interface and produce large pitch increases for rising
bubbles due to the thin-plate capacitor effect (see Figure 6).
We would like to acknowledge the help of Arthur Sams on early
Our frequency model provides a way to capture the complex fre- BEM computations, Andrew Wolfers for laboratory measurement
quency effects that acoustic bubbles exhibit, and we hope that it of fluid sounds, Hurf Sheldon for Cornell Measurement Lab sup-
will lead to better audiovisual simulation of water in the future, as port, Joel Gross and Pu Zhang for fluid sound discussions, Landon
well as more accurate methods for passive acoustic sensing. Our Boyd and Robert Bridson for sharing MultiFLIP software, the de-
frequency model can be solved efficiently, and is surprisingly robust velopers of the Gerris open source software, academic licensing
to coarse meshes (see Figure 8 top). of Maxwell Renderer, and early collaborations with Phaedon Sinis,
Katherine Breeden, Lindsey Williams (who first suggested Gerris)
Limitations and future work: Far from solving the problem out- and Pat Hanrahan. We thank the anonymous reviewers for their con-
right, the current study identifies many challenges, limitations, and structive feedback. We acknowledge funding and support from the
opportunities for future work in realistic audiovisual fluid simu- National Science Foundation (HCC-0905506, CAREER-1453101,
lation. Our formulations calculate frequency and radiation based DGE-1144153), Intel, donations from Pixar, Adobe, and XSEDE
on an independent bubble assumption; however, bubbles can af- compute resources. Initial phases of this research were conducted in
fect each other to produce frequency-coupled vibrations and in conjunction with the Intel Science and Technology Center–Visual
turn change their acoustic emissions. Our frequency-domain trans- Computing. Any opinions, findings, and conclusions or recommen-
fer model can capture resonance effects in containers and thereby dations expressed in this material are those of the authors and do
improve the sound quality over previous work [Zheng and James not necessarily reflect the views of the National Science Foundation
2009; Moss et al. 2010]. But it misses perceptually significant time- or others.
domain reverberation effects. In addition, the container sound is
missing, which can be an important source of fluid-solid coupling References
sounds. Also, we only model the dependence of ω on local geom-
etry, but bubble damping β may also be affected. Acoustic radia-
tion, while perceptually crucial, is computationally expensive. Our A NDO , R., T HUEREY, N., AND W OJTAN , C. 2015. A stream
bubble-plane proxy is one attempt at performance improvement, but function solver for liquid simulations. ACM Transactions on
other techniques such as fast-multipole methods, FDTD with PML Graphics (TOG) 34, 4, 53.
on the GPU, or geometric acoustics are interesting future work di-
rections. BANERJEE , P. K., AND B UTTERFIELD , R. 1981. Boundary el-
ement methods in engineering science, vol. 17. McGraw-Hill
Our sound model is inherently band-limited due to resolution- London.
limited fluid simulation. Post-processing effects such as popping
and microbubble models can help. But it is also desired to incorpo- B IGG , G. R., J ICKELLS , T. D., L ISS , P. S., AND O SBORN , T. J.
rate more advanced multi-scale models in our system. The bubble 2003. The role of the oceans in climate. International Journal of
forcing models we use are relatively simple, typically only captur- Climatology 23, 10, 1127–1159.
B OYD , L., AND B RIDSON , R. 2012. MultiFLIP for energetic two- H ONG , J.-M., AND K IM , C.-H. 2005. Discontinuous fluids. ACM
phase fluid simulation. ACM Transactions on Graphics (TOG) Transactions on Graphics (TOG) 24, 3, 915–920.
31, 2, 16.
H OWE , M. S., AND H AGEN , N. A. A. 2011. On the impact noise
B RAGG , S. W. H. 1920. The World of Sound. G. Bell and Sons of a drop falling on water. Journal of Sound and Vibration 330,
Ltd. 4, 625–635.
B RIDSON , R. 2008. Fluid simulation for computer graphics. CRC H OWE , M. S. 1998. Acoustics of Fluid-Structure Interatctions.
Press. Cambridge Press.
B USARYEV, O., D EY, T. K., WANG , H., AND R EN , Z. 2012. Ani- H OWE , M. S. 2002. Theory of Vortex Sound. Cambridge Press.
mating bubble interactions in a liquid foam. ACM Transactions
on Graphics (TOG) 31, 4, 63. K IM , B., L IU , Y., L LAMAS , I., J IAO , X., AND ROSSIGNAC , J.
2007. Simulation of bubbles in foam with the volume control
C HICHARRO , R., AND VAZQUEZ , A . 2014. The acoustic signature method. In ACM Transactions on Graphics (TOG), vol. 26, ACM,
of gas bubbles generated in a liquid cross-flow. Experimental 98.
Thermal and Fluid Science 55, 221–227.
K LUSEK , Z., AND L ISIMENKA , A. 2013. Acoustic noise gener-
C ZERSKI , H., AND D EANE , G. B. 2010. Contributions to the
ation under plunging breaking waves. Oceanologia 55, 4, 809–
acoustic excitation of bubbles released from a nozzle. The Jour-
836.
nal of the Acoustical Society of America 128, 5, 2625–2634.
L AMB , H. 1932. Hydrodynamics. Cambridge University Press.
C ZERSKI , H., AND D EANE , G. B. 2011. The effect of coupling on
bubble fragmentation acoustics. The Journal of the Acoustical L EIGHTON , T., F INFER , D., G ROVER , E., AND W HITE , P. 2007.
Society of America 129, 1, 74–84. An acoustical hypothesis for the spiral bubble nets of humpback
C ZERSKI , H. 2011. A candidate mechanism for exciting sound whales and the implications for whale feeding. Acoustic Bulletin
during bubble coalescence. The Journal of the Acoustical Society 22, 1, 17–21.
of America 129, 3, EL83–EL88. L EIGHTON , T. G. 1994. The Acoustic Bubble. Academic Press.
DA , F., BATTY, C., W OJTAN , C., AND G RINSPUN , E. 2015. Dou-
L EWINER , T., L OPES , H., V IEIRA , A. W., AND TAVARES , G.
ble bubbles sans toil and trouble: discrete circulation-preserving
2003. Efficient implementation of marching cubes cases with
vortex sheets for soap films and foams. ACM Transactions on
topological guarantees. Journal of Graphics Tools 8, 2 (decem-
Graphics (TOG) 34, 4, 149.
ber), 1–15.
D EANE , G. B., AND C ZERSKI , H. 2008. A mechanism stimulat-
L HUISSIER , H., AND V ILLERMAUX , E. 2012. Bursting bubble
ing sound production from air bubbles released from a nozzle.
aerosols. Journal of Fluid Mechanics 696 (4), 5–44.
The Journal of the Acoustical Society of America 123, 6, EL126–
EL132. L ONGUET-H IGGINS , M. S. 1989. Monopole emission of sound by
D EANE , G. B., AND S TOKES , M. D. 2002. Scale dependence asymmetric bubble oscillations. Part 1. Normal modes. Journal
of bubble creation mechanisms in breaking waves. Nature 418, of Fluid Mechanics 201, 525–541.
6900, 839–844. L ONGUET-H IGGINS , M. S. 1989. Monopole emission of sound by
D EANE , G. B., AND S TOKES , M. D. 2010. Model calculations asymmetric bubble oscillations. Part 2. An initial-value problem.
of the underwater noise of breaking waves and comparison with Journal of Fluid Mechanics 201, 543–565.
experiment. The Journal of the Acoustical Society of America
L ONGUET-H IGGINS , M. S. 1990. An analytical model of sound
127, 6, 3394–3410.
production by raindrops. Journal of Fluid Mechanics 214, 395–
D EANE , G. B. 2013. Determining the bubble cap film thickness of 410.
bursting bubbles from their acoustic emissions. The Journal of
L ONGUET-H IGGINS , M. S. 1991. Resonance in nonlinear bubble
the Acoustical Society of America 133, February 2013, EL69–75.
oscillations. Journal of Fluid Mechanics 224, 531.
D ING , L., AND FARMER , D. M. 1994. Observations of breaking
surface wave statistics. Journal of Physical Oceanography 24, 6, L OSASSO , F., S HINAR , T., S ELLE , A., AND F EDKIW, R. 2006.
1368–1387. Multiple interacting liquids. In ACM Transactions on Graphics
(TOG), vol. 25, ACM, 812–819.
E NRIGHT, D., M ARSCHNER , S., AND F EDKIW, R. 2002. Anima-
tion and rendering of complex water surfaces. ACM Transactions M EANS , S. L., AND H EITMEYER , R. M. 2002. Surf-generated
on Graphics (TOG) 21, 3, 736–744. noise signatures: a comparison of plunging and spilling breakers.
The Journal of the Acoustical Society of America 112, 2, 481–
G ARLAND , M., AND H ECKBERT, P. S. 1997. Surface simplifi- 488.
cation using quadric error metrics. In Proceedings of the 24th
Annual Conference on Computer Graphics and Interactive Tech- M EDWIN , H., AND B EAKY, M. M. 1989. Bubble sources of the
niques, ACM Press/Addison-Wesley Publishing Co., New York, Knudsen sea noise spectra. The Journal of the Acoustical Society
NY, USA, SIGGRAPH ’97, 209–216. of America 86, 3, 1124–1130.

H AGER , W. W. 1989. Updating the inverse of a matrix. SIAM M INNAERT, M. 1933. On musical air-bubbles and the sounds of
review 31, 2, 221–239. running water. Philosophical Magazine 16, 104, 235–248.
H ONG , J.-M., AND K IM , C.-H. 2003. Animation of bubbles M OSS , W., Y EH , H., H ONG , J.-M., L IN , M. C., AND M ANOCHA ,
in liquid. In Computer Graphics Forum, vol. 22, Wiley Online D. 2010. Sounding liquids: Automatic sound synthesis from
Library, 253–262. fluid simulation. ACM Trans. Graph. 29, 3 (July), 21:1–21:13.
N YSTUEN , J. A. 1986. Rainfall measurements using underwater A Fast Amortized BEM Solver Details
ambient noise. The Journal of the Acoustical Society of America
79, 4, 972–982. Here we provide details (from §5.4) on how to efficiently evaluate
x = A−1 b in our frequency solver for the common case of multiple
O SHER , S., AND F EDKIW, R. 2006. Level set methods and dynamic bubbles (with nb  na + nb ). First, to clarify, we avoided the
implicit surfaces, vol. 153. Springer Science & Business Media. expensive inverse in the U block of A−1 expression (12) using
P OPINET, S. 2003. Gerris: a tree-based adaptive solver for the the Sherman-Morrison-Woodbury low-rank update formula [Hager
incompressible euler equations in complex geometries. Journal 1989] to obtain
of Computational Physics 190, 2, 572–600.
−1
U = D − CG−1
bb B (16)
P OPINET, S. 2009. An accurate adaptive solver for surface-tension- −1 −1 −1 −1 −1
driven interfacial flows. Journal of Computational Physics 228, =D +D C (Gbb − BD C) BD (17)
16, 5838–5866.
| {z }
X

P ROSPERETTI , A. 1988. Bubble dynamics in oceanic ambient = D−1 (I + C XBD −1 −1

| {z }) = D (I − CY ). (18)
noise. In Sea Surface Sound, B. Kerman, Ed., vol. 238 of NATO −Y
ASI Series. Springer Netherlands, 151–171.
The bubble-independent solver setup constructs the air-solid domain
P ROSPERETTI , A. 1988. Bubble-related ambient noise in the ocean.
form-factor matrix D, and the LU factorization for D−1 . Then
The Journal of the Acoustical Society of America 84, 3, 1042–
given a specific bubble geometry, we construct the remaining blocks
1054.
of A, and then compute
P UMPHERY, H., C RUM , L., AND B JØRNØ , L. 1989. Underwater
sound produced by individual drop impacts and rainfall. J. Acoust. G−1
bb ← LU Solver(Gbb ) (19)
Soc. Am. 85, 1518–1526. T1 ← D C −1
(nb LU solves for C RHS) (20)
P UMPHREY, H., AND F FOWCS W ILLIAMS , J. 1990. Bubbles as X ← LU Solver(Gbb − BT1 ). (then discard T1 ) (21)
sources of ambient noise. IEEE Journal of Oceanic Engineering
15, 4, 268–274.
 
bb
Then given the RHS vector (in block form) b = we can form
R AYLEIGH , L. 1917. VIII. On the pressure developed in a liquid bas
 
during the collapse of a spherical cavity. Philosophical Magazine xb
x = A−1 b = as follows. Since
Series 6 34, 200, 94–98. xas
S MIGAJ , W., A RRIDGE , S., B ETCKE , T., P HILLIPS , J., AND
xb = Xbb + Y bas = X(bb − BD−1 bas ), (22)
S CHWEIGER , M. 2012. Solving boundary integral problems
with bem++. ACM Transactions on Mathematical Software. we can form xb via (here t∗ are temporary variables)
S PIEL , D. E. 1992. Acoustical measurements of air bubbles burst-
ing at a water surface: Bursting bubbles as helmholtz resonators. t1 ← D−1 bas (LU solve) (23)
Journal of Geophysical Research: Oceans (1978–2012) 97, C7, t2 ← bb − Bt1 (24)
11443–11452.
xb ← Xt2 . (LU solve) (25)
S PRATT, K. S., L EE , K. M., W ILSON , P. S., AND W OCHNER ,
M. S. 2015. On the resonance frequency of an ideal arbitrarily- Second, observe that
shaped bubble. Proceedings of Meetings on Acoustics 20, 1.
xas = Zbb + U bas = U (−CG−1
bb bb + bas ) (26)
S TAM , J. 1999. Stable fluids. In Proceedings of the 26th annual con-
ference on Computer graphics and interactive techniques, ACM so we can form xas = U t2 = D−1 (I − CY )t2 as
Press/Addison-Wesley Publishing Co., 121–128.
S TRASBERG , M. 1953. The pulsation frequency of nonspherical t1 ← G−1
bb bb (LU solve) (27)
gas bubbles in liquids. The Journal of the Acoustical Society of t2 ← bas − Ct1 (28)
America 25, 3, 536–537.
t3 ← Y t2 = −X(B(D−1 t2 ))) (2 LU solves) (29)
VAN DEN D OEL , K. 2005. Physically based models for liquid t4 ← t2 − Ct3 (30)
sounds. ACM Transactions on Applied Perception 2, 4 (Oct.),
534–546. xas ← D−1 t4 (LU solve) (31)
W ILSON , J. D., AND M AKRIS , N. C. 2008. Quantifying hurricane Assuming (for simplicity) that nb = O(1) and N = na + ns , then it
destructive power, wind speed, and air-sea material exchange involves only O(N ) new form factor computations per bubble (in-
with natural undersea sound. Geophysical Research Letters 35, stead of O(N 2 )), and it costs O(N 2 ) to construct this A−1 transfor-
10. mation (instead of O(N 3 )), and O(N 2 ) to apply it to form A−1 b.
Z HENG , C., AND JAMES , D. L. 2009. Harmonic fluids. ACM The main per-bubble cost is nb + 3 applications of the D−1 LU
Trans. Graph. 28, 3 (Aug.), 37:1–37:12. solver.

Z HENG , C., AND JAMES , D. L. 2010. Rigid-Body Fracture Sound

with Precomputed Soundbanks. ACM Transactions on Graphics
29, 4 (July), 69:1–69:13.
Z HENG , W., YONG , J.-H., AND PAUL , J.-C. 2009. Simulation of
bubbles. Graphical Models 71, 6, 229–239.