Journal of Fluids and Structures 58 (2015) 20–39

Contents lists available at ScienceDirect

Journal of Fluids and Structures

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

On the verification and validation of a spring fabric

for modeling parachute inflation
Qiangqiang Shi a,n, Daniel Reasor b, Zheng Gao a, Xiaolin Li a, Richard D. Charles c
a
Department of Applied Mathematics and Statistics, Stony Brook University, Stony Brook, NY 11794–3600, United States
b
Test Support Squadron – Testing Techniques, Edwards AFB, CA 93524, United States.
c
Airdrop Technology Team, U.S. Army Natick Research Development and Engineering Center (NSRDEC), Natick, MA 01760, United States

a r t i c l e in f o abstract

Article history: A mesoscale spring–mass model is used to mimic fabric surface motion. Through coupling
Received 29 May 2014 with an incompressible fluid solver, the spring–mass model is applied to the simulation of
Accepted 25 June 2015 the dynamic phenomenon of parachute inflation. A presentation of a verification and
validation efforts is included. The present model is shown to be numerically convergent
Keywords: under the constraints that the summation of point masses is constant and that both the
Elastic membrane tensile stiffness and the angular stiffness of the spring conform with the material's Young
Spring model modulus and Poisson ratio. Complex validation simulations conclude the effort via drag
Parachute inflation force comparisons with experiments.
& 2015 Published by Elsevier Ltd.

1. Introduction

The thinness and lack of bending stiffness of canopy-like structures, such as parachutes, introduces a separation in scales
of local and global curvature, contact interactions, and two-sided fluid–structure interactions. Due to these complexities,
parachute modeling and simulation has traditionally been empirical in nature. Realistic and accurate analysis requires
sophisticated techniques in fluid–structure interaction (FSI) including computational fluid dynamics (CFD) and computa-
tional structure dynamics (CSD). Unfortunately these tools are not currently used to design aerodynamic deceleration
systems due to the computational resources required and the lack of validation of physics-based modeling efforts. However,
the computational platform based on the front tracking method (Glimm et al., 2000; Du et al., 2006) and the software
library FronTier (Fix et al., 2005; Glimm et al., 2007), with intelligent data structures and functionalities for geometrical and
topological evolution as well as the ability to handle the interaction between the moving interface and the fluid (both
incompressible and compressible) lends itself to fill these voids.
The authors' previous studies of Kim et al. (2013) and Li et al. (2013) discuss the application of the mesoscale model in effects to
mimic the dynamic motion of a fabric surface while coupled with an incompressible fluid solver for parachute inflating and
descending. The numerical simulations demonstrated good agreement with the experimental results both qualitatively and
quantitatively, but questions remain on the validity of the model and its relation to the continuum model of the fabric surface as an
elastic membrane. Van Gelder (1998) argued that the simple spring–mass model cannot be related to the continuum model for the
linear elastic membrane and therefore not suitable to represent the fabric surface. However, recently, Delingette (2008) proposed a
revision of the spring–mass model that includes the angular deformation energy where the force at each triangle vertex includes
spring forces through both its adjacent and opposite sides. This spring–mass model originates from the discretization of a St Venant

n
Corresponding author.

http://dx.doi.org/10.1016/j.jfluidstructs.2015.06.014
0889-9746/& 2015 Published by Elsevier Ltd.
 Q. Shi et al. / Journal of Fluids and Structures 58 (2015) 20–39 21

Kirchhoff membrane on a linear triangle element. Delingette demonstrated that with this modification, the force in the spring
model is indeed related to the strain and stress of the elastic membrane. For small displacements, as mentioned in Delingette
(2008), the behavior of the membrane matches exactly with that of the linear elastic material, which is verified numerically in this
paper. We will demonstrate that the difference between the spring–mass model used in our previous studies and Delingette model
can be bridged by a re-interpretation of the spring constant and by adding additional forces contributed by the opposite sides of the
vertex. We will show that the spring–mass model is numerically convergent under the constraints that the total mass is conserved
and that both the tensile stiffness and the angular stiffness of the spring conform with the material's Young modulus and Poisson
ratio. In addition, we will present an algorithm to compute the von-Mises stress of a fabric surface under strain.
The objective of this work is to employ the Delingette-modified spring–mass model in the simulation of parachute
inflation, which was also studied in our previous work (Kim et al., 2013) focusing upon the steady drag and velocity during
the terminal descent stage and the dependency of these variables on the parachute geometry, dimension, and payload. In
contrast to terminal descent, parachute inflation has a relatively short time duration, typically a few seconds. However, this
short period regime is paramount to the effectiveness of the deceleration system and modeling it accurately with physics-
based tools is difficult. Any malfunction, such as inversion, barber's pole, or jumper-in-tow, happens in this short stage could
have serious effect on the fate of the personnel and cargo delivery.
Parachute inflation is a complex aeroelastic phenomenon that involves complex aerodynamics and elastic structures. The
fact that the flow field and the canopy's geometric shape are interactive makes the inflation process a very difficult event to
model (Peterson, 1993). Researchers have studied parachute inflation with different methods including empirical analysis
(Potvin, 1998), semi-numerical simulation (Potvin et al., 2011), and through experiments (Potvin et al., 2011; Potvin and
McQuilling, 2011). During the inflation sequence, the canopy not only experiences extreme increase in internal volume and
large loading, but also involves geometric and material nonlinearities which make it a highly challenging event to model
from a dynamic structures perspective (Cochrane et al., 2010). Parachute inflation consists of a sequence of dynamically
animated stages. These stages have been summarized in Yu and Ming (2007) and Hamid and Desabrais (2002). For example,
during the inflation of a circular parachute, the canopy starts as a vertical tube with an open lower end. With very little drag,
air quickly rushes into the canopy tube due to the fast descending velocity of the system. Due to the limited volume inside
the canopy, the pressure at the apex point increases dramatically leading to a large pressure difference between the internal
and the external sides of the canopy. Meanwhile, the continued accumulation of air inside the canopy (inflation) results in
the expansion both vertically and horizontally. The duration of the inflation process depends on the orientation of the
parachute to the oncoming flow, altitude, and flight speed and ends when a sufficient amount of high pressure air fills the
canopy. The aerodynamic forces that act in opposite directions along the parachute determine its steady-state shape and the
final shape of the recirculating region (air bubble) inside (Potvin, 1998). In addition, the presence of the intense flow
separation outside the canopy, the strong turbulence near the edge of parachute canopy and the narrow space inside the
canopy before it is fully inflated result in computational challenges from the CFD perspective (Yihua et al., 2010).

2. Improvement of mathematical model

The original spring model (Li et al., 2013) that serves as the basis for this effort is a simplified mesoscale model which
assumes that the force required to bend the surface is negligible and the force to stretch the surface is proportional to the
displacement from the equilibrium distance between adjacent mass points. In this model, the kinetic energy and the
potential energy of the triangulated mesh are given by

1 _ i j2 ; 1XN XN
T i ¼ mi jX V¼ κ ðjXi  Xj j  l0ij Þ2 ηij ; ð1Þ
2 4i¼1j¼1

where Xi is the position of the vertex i, κ is the spring constant, l0ij is the equilibrium length of the side shared by vertices Xi
and Xj , and ηij is a Boolean variable for adjacency. Applying the Lagrangian equation to L ¼ T  V, that is,
!
d ∂L ∂L
 ¼ 0; ð2Þ
dt ∂q_ j ∂qj

here qj refer to components of the coordinate, we can obtain the equation of motion for each vertex:
dX_i
mi ¼ Fi ; i ¼ 1; 2; …; N; ð3Þ
dt
where the force at each vertex point is
X
N   X
N
Fi ¼ ηij κ jXi Xj j l0ij eij ¼ ηij κ dlij eij ; ð4Þ
j¼1 j¼1
0
where eij is the unit vector from Xi to Xj , and we have defined dlij ¼ jXi  Xj j  lij as spring length elongation. Although this
model can simulate the dynamic motion of a fabric surface and exert correct tension and wrinkling of the surface, there is a
lack of proof on the relation between this model and the continuum model for an elastic membrane. Fig. 1 demonstrates the
mesh structure of such mode.
22 Q. Shi et al. / Journal of Fluids and Structures 58 (2015) 20–39

Fig. 1. The spring model on a triangulated mesh. Each vertex point in the mesh represents a mass point with point mass m. Each edge of a triangle has a
tensile stiffness. With the equilibrium lengths set during the initialization, the changing length of each side exerts a tensile spring force on the two
neighboring vertices in opposite directions. Each angle of a triangle has an angular stiffness which is set during the initialization. An additional tensile force
is generated when the angle is changed. Gore boundaries are modeled by curves with higher tensile stiffness.

Fig. 2. (a) Rest triangle T X0 whose vertices are Xi0 . (b) Deformed triangle T X whose vertices are Xi .

2.1. Delingette modification

The work of Delingette (2008) details the modeling efforts of a nonlinear membrane using a spring–mass model with a
triangulated mesh. This model is favored because of its relation to the elastic membrane model in continuum mechanics.
Illustrative features of Delingette's model are given in Fig. 2. The energy of membrane WðT X0 Þ that is required to deform a
single triangle T X0 with vertices fX10 ; X20 ; X30 g into its deformed position T X with vertices fX1 ; X2 ; X3 g consists of two parts:

 the energy of three tensile springs that prevent edges from stretching;
 the energy of three angular springs that prevent any change of vertex angles.

For a triangle in equilibrium T X0 , the initial states are given by area AX0 , angles αi, and lengths lij ði; j A 1; 2; 3; ja iÞ in
0

equilibrium while AX , βi and lij denote the area, angles, and lengths of the deformed triangle T X respectively.
0
The edge elongation can be written as dlij ¼ lij  lij . The potential energy is given (Delingette, 2008) as

  X
3
1 T X0 X
3
T
W T X0 ¼ κ ðdlij Þ2 þ γ i X0 dlij dlik ;
i ¼ 1
2 ij i ¼ 1
j ¼ ði þ 1Þ mod 3 j ¼ ði þ 1Þ mod 3
k ¼ ði þ 2Þ mod 3
 Q. Shi et al. / Journal of Fluids and Structures 58 (2015) 20–39 23

Fig. 3. Triangles T1 and T2 share Xi and Xj , the other vertices of triangles T1 and T2 are Xm and Xn respectively.

where

ðlij Þ2 ð2cot2 αk ðλ þ μÞ þ μÞ
0
T
κ ijX0 ¼
8AX0
is the tensile stiffness and

lik lij ð2cotαj cotαk ðλ þ μÞ  μÞ

0 0
T
γ i X0 ¼ ð5Þ
8AX0
is the angular stiffness, where i ¼ 1; 2; 3, j ¼ ði þ1Þ mod 3, and k ¼ ði þ 2Þ mod 3. γ and μ are the Lamé coefficients of the
material. These coefficients are simply related to the two physically meaningful parameters defined in planar elasticity for a
membrane, that is, Young's modulus E and the Poisson ratio ν (Gere, 2004):
Eν Eð1  νÞ
λ¼ and μ¼ :
1  ν2 1  ν2
Young's modulus quantifies the stiffness of the material, whereas the Poisson ratio characterizes the material
compressibility.
Through the application of Rayleigh–Ritz analysis the fabric surface, represented by the triangular mesh, should evolve
by attempting to minimize its membrane energy. Therefore, along the opposite derivative of that energy with respect to the
nodes of the system, that is, the deformed positions Xi :
  ∂WðT X0 Þ
Fi T X0 ¼ 
∂Xi
X3
T  X  Xi
¼ κ ijX0 dlij j
j ¼ 1;j a i
lij
3 
X X X
T T
þ γ k X0 dljk þ γ i X0 dlij k i
: ð6Þ
j ¼ 1;j a i
lik
k a j;k a i

The membrane deformation energy of the whole triangulation is the sum of the energy of each triangle. Thus, we obtain the
force at each vertex point as
X
N
Fi ¼ ηij Fij : ð7Þ
j¼1

As illustrated in Fig. 3, if vertices Xi and Xj are shared by two triangles T1 and T2, the other unshared vertices of triangles
T1 and T2 are denoted by Xm and Xn respectively, therefore we have
Fij ¼ ððκ Tij1 þ κ Tij2 Þdlij þ ðγ Ti 1 dlim þ γ Tj 1 dljm þ γ Ti 2 dlin þ γ Tj 2 dljn ÞÞeij
¼ κ~ ij dlij eij þ γ~ ij dlij eij ; ð8Þ

where κ~ ij ¼ κ Tij1 þ κ Tij2 , γ~ ij ¼ ðγ Ti 1 dlim þ γ Tj 1 dljm þ γ Ti 2 dlin þ γ Tj 2 dljn Þ=dlij and eij is the unit vector from Xi to Xj .
If the second term in Eq. (8) is neglected, Delingette's model is the same as the model used in Kim et al. (2013) except
that the spring constant varies if the corresponding initial triangles deviate from isosceles triangles. Numerical evidence
suggests that both the variation of κ~ ij and the modification from the second term (due to angular stiffness) are significant in
the above cases.

2.2. Stress analysis

The stress distribution of the parachute canopy fabric and gores during the process of inflation is valuable information
prior to the real-world test and evaluation process. The natural stresses are normal stresses directed parallel to the triangle
24 Q. Shi et al. / Journal of Fluids and Structures 58 (2015) 20–39

Fig. 4. The angle of each side with respect to the x-axis.

Fig. 5. Viscosity and streamline around parachute at time t ¼ 2 s computed by RNG k–ε model.

sides and the natural stress of the triangle in the spring–mass model is due to the stretching of each triangle side. Let τ1 ; τ2 ,
and τ3 be the natural stress on sides 1, 2, and 3 of a triangle respectively, we have

τi ¼ κ ðljk l0jk Þ; i ¼ 1; 2; 3; j ¼ ðiþ 1Þ mod 3; k ¼ ðj þ 1Þ mod 3; ð9Þ

where κ is the spring constant, ljk and ljk are the equilibrium and stretched length of the side shared by vertices Xj and Xk
0

respectively. This natural stress can be converted (Gere, 2004) into the stress in Cartesian coordinates on the plane of the
triangle. The Cartesian stress is a 2  2 tensor in the plane of the triangle
!
σ xx σ xy
σ¼ σ σ yy : ð10Þ
xy
 Q. Shi et al. / Journal of Fluids and Structures 58 (2015) 20–39 25

The conversion from natural stress to Cartesian stress is accomplished through a mapping matrix, that is
0
σ xx 1 0 c21 s21 s1 c1 1  1 0 τ1 1
B σ C B c2 s2 s c C B τ C
@ yy A ¼ @ 2 2 2 2A @ 2 A; ð11Þ
σ xy c23 s23 s3 c3 τ3
where ci ¼ cos θi ¼ dxi =ljk and si ¼ sin θi ¼ dyi =ljk are the trigonometric functions of the angle of each side with respect to the
x-axis, as shown in Fig. 4. The stresses in each of the two principal directions are obtained via diagonalization of the stress
tensor, that is
! !
σ xx σ xy σ1 0
¼T T  1; ð12Þ
σ xy σ yy 0 σ2

where the columns of the matrix T represent the principal direction vectors, and σ1 and σ2 are the solutions of the
characteristic equation:
 
 σ xx  σ 1;2 σ xy 


 σ xy σ yy  σ 1;2  ¼ 0: ð13Þ

The von-Mises stress is commonly used to evaluate safety factors of material and is given by Eq. (14). The so-called safety
factors are defined as the ratio of the significant strength of the material to the von-Mises stress observed under the
application of interest. When this factor is observed to be a critical value (that corresponds to a high stresses application of
the fabric surface or gores) the parachute construction or material may need to be modified or a more appropriate model
used for that particular application:
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Σ vm ¼ σ 21 þ σ 22  σ 1 σ 2 : ð14Þ

3. Numerical methods

The numerical framework employed is based on the previous analysis and numerical verification in Li et al. (2013) and
Kim et al. (2013). These previous analyses proved that the spring–mass system without external force is a conservative
system and the motion of point mass in its tangential direction to the surface (or string on 2D) is purely oscillatory, and that
there exists an upper bound for the eigen frequency of the oscillation:
rffiffiffiffiffiffiffiffiffiffi
2M κ
ωmax r ; ð15Þ
m
where k is the spring constant, m is the point mass, and M is the maximum number of neighbors a vertex point in the spring
mesh can have. Using explicit time integration with the fourth order Runge–Kutta method, we showed that the energy
amplification factor per time step is ξ  1  72 ðωmax ΔtÞ6 , where Δt is the time step. Therefore to ensure stability and
2 1

accuracy, ωmax Δt o 0:1 is chosen for simulations discussed herein. The modification of the spring model due to Delingette
adds variation to the spring constant, but the general principle of the upper bound is still valid except that the maximum
spring constant in Eq. (15) should be substituted in for the determination of the upper bounds. This claim is numerically
verified in all the simulations. In addition to the Delingette modification of the spring model, several new components have
been added to the computational framework for realistic and efficient simulation of the parachute system. We briefly
discuss them in the following sub-sections.

3.1. Fluid solver and turbulence model

For personnel and cargo parachute, the speed of the surrounding flow is much smaller than the sound speed, therefore
the use of incompressible Navier–Stokes equation is appropriate:
Du
ρ þ ∇p ¼ μ∇2 u þ ρg; ð16Þ
Dt
where D=Dt ¼ ∂=∂t þ u  ∇ is the material derivative of the fluid. The incompressibility is given by
∇  u ¼ 0: ð17Þ
We solve this equation using the projection scheme (Chorin, 1968; Brown et al., 2001) and couple the fluid equation
with the canopy surface through the impulse method (Kim et al., 2013). We improved the accuracy of the fluid solver by
adapting the staggered grid algorithm in the projection step (Tau, 1994; Matyka, 2004; Langtangen et al., 2002; Lee and
LeVeque, 2003).
26 Q. Shi et al. / Journal of Fluids and Structures 58 (2015) 20–39

Fig. 6. Cross-sectional plot of eddy viscosity around parachute with different ymax. The left plot uses ymax ¼ 0:3 and the right plot uses ymax ¼ 0:4 on the
down side of the canopy and ymax ¼ 0:6 on the up side of the canopy.

Our treatment of the interface condition is similar to the immersed boundary method by Kim and Peskin (2006, 2009),
the difference is in the application of turbulent viscosity. We noted that the payload in Kim and Peskin's simulation is only
up to a few grams. We realized that the major obstacle in their simulation is due to the large Reynolds number, which
typically exceeds several millions for full-scale canopy (Johari and Desabrais, 2005). Therefore, eddy viscosity models based
on the Reynolds Averaged Navier–Stokes (RANS) equations are needed. The hydrodynamic behavior of a turbulent
incompressible fluid is governed by the RANS equations for the mean velocity U and pressure p:
∂U  h i
þU  ∇U ¼ ∇p þ ∇  ðν þ νT Þ ∇U þ ∇U T
∂t
∇  U ¼ 0; ð18Þ

where ν is the kinematic viscosity and νT is the turbulent eddy viscosity which is supposed to approximate the turbulent
fluctuations. In the Baldwin–Lomax (1978) algebraic model, νT ¼ lm Ω, where lm is the mixing length specified with a
2

damping function and Ω is the modulus of the mean rate of rotation tensor. The Baldwin–Lomax is relatively reliable since it
seldom produces completely unphysical results. Therefore, we adopt this model as a first step to test the parachute model.
However, this simple model tends to overpredict separation regions, thus not suitable for the cases with strong separation
such as the region around parachute canopies. The eddy viscosity computed by Baldwin–Lomax model around parachute
canopies with different maximum wake distance ymax are displayed in Fig. 6. A more realistic one is the RANS-based
turbulence model, or the k–ε model family, which automatically calculates the turbulence length scale (Wilcox et al., 1998).
In the standard k–ε model, the eddy viscosity is defined as
2
k
νT ¼ C μ ; ð19Þ
ε
where k is the turbulence kinetic energy and ε is the dissipation rate. To compute k and ε, two additional convection–
diffusion–reaction equations are needed:
 
∂k νT
þ∇  kU  ν þ ∇k ¼ P k  ε; ð20Þ
∂t δk
 
∂ε νT ε
þ∇  εU  ν þ ∇ε ¼ ðC 1 P k  C 2 εÞ; ð21Þ
∂t δε k

where P k ¼ ðνT =2Þj∇U þ∇U T j2 is the production of turbulent kinetic energy. For the standard k–ε model, the default values
of the involved empirical constants are C μ ¼ 0:09, C 1 ¼ 1:44, C2 ¼1.92, δk ¼ 1:0, δε ¼ 1:3. Although simple and efficient, the
standard model is unable to capture the effects of smaller scales of motion due to its single turbulence length scale. In order
to account for the different scales of motion, a mathematical technique called Re-Normalization Group (RNG) method
(Yakhot and Smith, 1992) was used to derive a turbulence model similar to the standard one, resulting in a modified form of
the ε equation:
 
∂ε νT ε 
þ∇  εU  ν þ ∇ε ¼ C 1 P k C n2 ε ; ð22Þ
∂t δε k

C μ η3 ð1  η=η0 Þ
C n2 ¼ C 2 þ ; ð23Þ
1 þ βη3
 Q. Shi et al. / Journal of Fluids and Structures 58 (2015) 20–39 27

Fig. 7. The reinforcement cables for the gores are modeled using the same spring system, but with different spring constants along the gore seams or
curved edges. The left plot shows fully opened canopy with expanded gores. After the inflation, the gore structure is clearly revealed. The right plot shows
the surface mesh and the details of the gore boundary curves.

η ¼ kS=ε, S is the modulus of the mean rate of strain tensor. The coefficients are derived explicitly in the RNG procedure and
also list here for the sake of completeness: C μ ¼ 0:0845, C 1 ¼ 1:42, C2 ¼1.68, δk ¼ 0:7194, δε ¼ 0:7194.
The implementation of k–ε model has no difference by solving other convection–diffusion–reaction equations, except
that two things need to be noticed. Firstly, one must specify an appropriate boundary condition to a solid wall for the
velocity and two turbulence parameters. Due to the extremely high Reynolds number in parachute simulation, it is
worthwhile to use the wall functions to bridge the viscosity-affected region and the fully turbulent region avoiding the need
for resolution of strong velocity gradients (Kuzmin et al., 2007). Another task is to avoid loss of positivity of k and ε due to
computational errors. This can be achieved by simply keeping the coefficients positive for the linearized equations without
touching any primitive variables (Lew et al., 2001). Fig. 5 displays the viscosity and velocity streamline computed with RNG
k–ε model.

3.2. Modeling of parachute gores

Most parachute canopies are made in patches called gores. Parachute gores are stitched by the reinforcement cables. The
reinforcement cables are important structures which can have substantial impact on the parachute's interaction with the
surrounding fluid. The gore structures in the parachute system can also affect the stability of the parachute motion. To
accurately model the aerodynamic motion of a parachute, a mathematical model which reveals the geometry as well as the
material strength of the canopy surface and the gore stitches must be employed. In the present model, the reinforcement
cables are treated as curves embedded in the canopy surface. Young's modulus for the fabric surface and the reinforcement
cable differ by assigning different spring constants to the surface mesh and to the gore curves. The insertion of reinforced
gore boundaries as stiffened interior curves in the canopy surface mesh is demonstrated in Fig. 1. Fig. 7 demonstrates that
when fully inflated, the spring system reveals both the patches and the indentation of the reinforcement cables.

3.3. Modeling of canopy porosity

It has long been known that permeability is an important material property that influences canopy performance; it is
paramount to accurately capture fabric dynamics, total drag, and drift distance. A finite permeability will make the
parachute much more stable while still maintain the balance of the payload. The permeability of the parachute material
often plays a vital role in the parachute design. A state-of-art tuning from an impervious material to a highly permeable
fabric can make a parachute from a wandering sloth into a plummeting stabilizer. We have considered three different
implementations of porosity for the parachute including Tezduyar and Sathe (2007), Kim and Peskin (2006), and Tutt
28 Q. Shi et al. / Journal of Fluids and Structures 58 (2015) 20–39

Fig. 8. The difference between the collision handling for the fluid interface and the fabric surface. The upper two plots show the topological bifurcation of
fluid interface. The lower two plots show the repulsion of the fabric collision.

(2006). We have chosen to implement the two latter methodologies because they share a similar framework of
computational mesh and are compatible with the front tracking environment employed.
For Kim and Peskin's method, porosity is considered as leaking pores in the immersed elastic boundary. Let β be the
density of pores and there are β ds pores in the interval ðs; s þ dsÞ along the surface. If each pore has an aerodynamic
conductance equal to γ, the flux through the pore is γ ðp1  p2 Þ where p1 and p2 are the pressures on the two sides of the
boundary, then the flux through the interval ðs; s þ dsÞ of the boundary is given by βγ ðp1  p2 Þ ds.
Tutt's algorithm is closest to the front tracking implementation. Tutt used the Ergun equation to describe the magnitude
of porous flow velocity at a given pressure difference based on two coefficients in the following equation:

ΔP μ ρ
¼  vf þ  v2f ¼ a  vf þb  v2f ;
L K1 K2

where

ε3  D2
K1 ¼ ;
150  ð1  εÞ2

ε3  D
K2 ¼
1:75  ð1  εÞ

are referred to as the viscous and inertial factors respectively. D is defined as the characteristic length, ε is the porosity and is
equal to the ratio of the void and total volume. vf, μ, and ρ are fluid velocity, viscosity, and density respectively. In the
FronTier-airfoil (Kim et al., 2013; Li et al., 2013) code used here, the pressure difference is computed after the velocity
projection into the divergence-free space. We can then use the pressure difference on two sides of the canopy to compute vf
and add it to the advection solver as flux from the immersed elastic surface.

3.4. Collision handling

The original FronTier code was designed for the study of fluid interface instability problems. For these problems, collision
and contact of surfaces are resolved by merging or bifurcation. However, fabric surface can neither merge nor bifurcate,
therefore we need to have functions which can carefully deal with the repulsive contact of structure surfaces. Parachute
collision/contact is handled through the standard FronTier library with major revisions such as functions to handle the non-
manifold surface and three dimensional curves (not as boundary of a surface, such as the spring chords). Fig. 8 shows the
difference of collision handling between the fluid–fluid interface and the fabric–fabric surface.
 Q. Shi et al. / Journal of Fluids and Structures 58 (2015) 20–39 29

Table 1
GPU and CPU computing time of three dimensional spring model. Spring models with three different mesh sizes (80  80  20, 160  160  40,
320  320  80) were tested with both pure CPU code and hybrid (CPU and GPU) code. The hybrid code is 15–20 times faster than the pure CPU code for
the computation intensive part. In general, the overall speed of the hybrid code is 7–10 times faster than the pure CPU code for the spring–mass
model part.

Mesh Total time (s) Per step (s)

Without-GPU With-GPU CPU GPU

80  80  20 1345 152 1.71 0.12

160  160  40 8115 804 5.02 0.25
320  320  80 40 291 5349 11.63 0.55

Table 2
A Dell precision T7600 workstation with dual NVIDIA quadro graphics cards was used to set up the test environment.

Hardware CPU Dual eight core XEON E5-2687W, 3.1 GHz

64 GB DDR3
32 kB  16 L1 Cache, 256 kB  16 L2 Cache, 20 MB
 2 L3 Cache
GPU Dual Quadro 6000 with 14 multiprocessor, 448 cores, 1.15 Hz
6 GB global memory, 64 kB constant memory
48 kB shared memory and 32 768 registers per multiprocessor

Software OS Fedora 18 with kernel 3.9.2-200.fc18.x86_64

Compiler gcc version 4.7.2
CUDA CUDA Toolkit 5.0

3.5. Global indexing of surface mesh

Global indexing is a new feature that is added to the computational framework. The original FronTier (Du et al., 2006) had
to deal with frequent surface mesh optimization and topological reconstructions. This makes the parallelization based on
the matching of global index very difficult. As a result, the original FronTier library relied on floating point matching for
parallel communication. The floating point matching is not completely reliable therefore more complicated algorithms were
implemented as reinforcement. However for fabric-like surface, especially when a spring–mass model is used, the inter-
connectivity and proximity of the interface marker points should not be changed. Therefore, global indexing is ideal for the
parallel communication of the surface information as employed in this work.

3.6. Modularized functionalities

Modularization is emphasized in our code development. The parachute module is an independent application program.
This new module consists of four components:

(1) the initialization module;

(2) the ODE module for the spring system;
(3) the PDE module for fluid dynamics;
(4) the FronTier library for the interface geometry handling.

3.7. GPU accelerated spring module

Graphics Processing Unit (GPU) computing (Kirk and Hwu, 2010) is to use the GPUs together with CPUs to accelerate a
general-purpose scientific and engineering application. GPU computing can offer dramatically enhanced application
performance by offloading computation-intensive portions of the programming code to the GPU units, while the remainder
of the code still runs on the CPU. Joint CPU/GPU applications constitute a powerful combination because CPUs consist of a
few cores optimized for serial processing, while GPUs consist of thousands of smaller, more efficient cores designed for
massive parallel calculations. Serial portions of the code with logical comparison run on the CPU while floating point
operation intensive parallel portions of the code run on the GPU. The total operation time recorded by the CPU clock and
time for GPU intensive computational part are collected in Table 1 in seconds. From Table 1 we can conclude that the
application of the GPU has a clear advantage in the computation of spring model. To fully take advantage of the GPU system,
further optimization are needed to the fluid solver. Table 2 shows the hardware structure of our computer.
30 Q. Shi et al. / Journal of Fluids and Structures 58 (2015) 20–39

Fig. p
9.ffiffiffiffiffiffiffiffiffiffiffiffiffi
(Top) partial enlargement of the string chords with different mesh sizes and (bottom) shape of the string chords at dimensionless time
τ ¼ t keq =m ¼ 823, where t ¼ 2 s is the real time, m ¼ 5  10  4 kg is the total mass, keq ¼ k=ns  81 N=m is the equivalent spring constant of series springs
and ns is the number of spring segments in the string. From left to right the total number of points in the string are 60, 123, 248, 498. The total mass of the
string chord as well as the payload at the lower end is kept constant in the simulations.

Table 3
Initial configuration of the 2-D and 3-D simulations

Mesh size Stiffness (N/m) Number of points Average mass (g)

Swing
Case 1 502 5000 60 0.008333
Case 2 1002 10 000 123 0.004065
Case 3 2002 20 000 248 0.002016
Case 4 4002 40 000 498 0.001004

Drum
Case 1 153 1000 266 1.428571
Case 2 303 1000 990 0.383838
Case 3 603 1000 3790 0.100264
Case 4 1203 1000 14 841 0.025605

4. Numerical results

The system of equations for the spring model is nonlinear and an analytical proof of convergence for this model is very
difficult. Therefore the proof of convergence through numerical mesh refinement is presented here, e.g., with fixed initial
and boundary conditions being used, the displacement, the energy (kinetic energy and potential energy), and the total
length (2D) or area (3D), which are functions of time, are convergent under mesh refinement. It is not surprising that the
convergence rate is of first order due to the equation of each vertex point in the spring mesh involves only its immediate
neighbors and that convergence is not ideal when the surface is compressed or wrinkled.
Proving analytical convergence for the coupled spring–mass model fluid solver system is even more difficult than the
spring–mass model itself; as a result, coupled simulation results are compared directly with experiments. One of the main
interests of this effort is the short time history response of the parachute during inflation where the drag force can be
validated with experimental data. It is demonstrated that the coupled method discussed above captures important
properties of the parachute inflation force response.

4.1. Verification of numerical convergence

One crucial step in assessing the mathematical validity of the spring–mass model is convergence under mesh refinement.
The following question is, if the system is convergent, to what continuum model and partial differential equations the
discrete computation of the spring–mass model would converge? The numerical results reveal that the spring model is
 Q. Shi et al. / Journal of Fluids and Structures 58 (2015) 20–39 31

Fig. 10. Convergence test results of spring model on string chord. In this test the string chord is fixed at one end while the other end has a payload and is
free to move. The simulations are on the sequences with 60, 123, 248, 498 points. The total mass M ¼Nm is a constant in the simulation. The upper left plot
shows the string lengths during the four simulations and the upper right plot is the Cauchy error of the sequences. The lower plots are for the total kinetic
energy of the system.

Table 4
Convergence tests of spring model for a string chord. In the computational sequences, the total mass of the string chord is fixed. As the number of points
increases, the point mass is reduced accordingly. Cauchy error is calculated on two consecutive mesh sequences. Columns el, ek, and ep are errors of total
length, total kinetic energy and total spring potential energy respectively. The numerical results show the first order convergence on each of them.

Mesh size el ek ep

50 and 100 0.00374 0.01664 0.03464

100 and 200 0.00184 0.00821 0.01726
200 and 400 0.000918 0.00406 0.008614

convergent under the conditions that the total mass of the fabric surface is kept constant and that both the spring tensile
stiffness and angular stiffness conform with Young's modulus and the Poisson ratio of the material (Delingette, 2008).
A sequence of numerical simulations of a string with fixed boundary in a two-dimensional (2-D) domain and a
membrane with fixed boundary in a three-dimensional (3-D) domain are carried out. In both 2-D and 3-D simulations, the
spring constant of the string (2-D) or the membrane (3-D) is conformed with the material's Young modulus and Poisson
ratio. The total mass, which is the summation of all point mass M total ¼ Nm, is kept constant where N is the total number of
points and m is the mass of each mass point. Therefore, as the computational mesh is refined and the total number of mass
points increases, the point mass m is decreased in proportion to the reciprocal of N.
The simulations are computed in domains of 1  1 m2 in 2-D and 1  1  1 m3 in 3-D. In the first test case, the dynamic
motion of a swinging string with one end fixed is simulated. The initial length of the string is 0.583 m and a weight of 0.5 g
as shown in Fig. 9. The total number of grid points is increased and the point masses for each experiment varied accordingly.
The details of each simulation are presented in Table 3. The total length and total kinetic energy of the string as a function of
time and their Cauchy errors are displayed in Fig. 10 and the numerical errors are shown in Table 4. From Table 4, it is clear
32 Q. Shi et al. / Journal of Fluids and Structures 58 (2015) 20–39

Fig. 11. Convergence test of the drum membrane under mesh refinement. From left to right the computational mesh of the domain are 15, 30, and 60. The
total mass of the membrane is kept constant in the simulations. The upper three plots show the membrane position at t ¼ 1 s and the lower plots show
the membrane position at t ¼ 2 s.

Table 5
Convergence tests of spring model for a fabric drum. In the computational sequences, the total mass of the membrane is fixed. As the number of points
increases, the point mass is reduced accordingly. Cauchy error is calculated on two consecutive mesh sequences. Columns eA, ek, and ep are errors of total
area, total kinetic energy and total spring potential energy respectively. The numerical results show the first order convergence on each of them.

Mesh size eA ek ep

15 and 30 0.02528 1.91810 1.97967

30 and 60 0.01507 1.21784 1.21772
60 and 120 0.00604 0.53550 0.52837

Fig. 12. Convergence test results of spring model on membrane. The boundary of the membrane is fixed. The simulations are on the sequences with 153,
303, 603 and 1203 mesh for the computational domain. The total mass M ¼ Nm is a constant in the simulation. The left plot shows the total area during the
four simulations and right plot shows the Cauchy errors of the sequences.

that the errors in the Cauchy sequence are reduced approximately by half each time as we reduce the average mesh size by
half. This indicates that the sequence is convergent to first order.
In the second case, a circular vibrating membrane with radius of 0.4 m and a total weight of 380 g is simulated. The
membrane is linearly perturbed initially from the center to the fixed boundary. Table 5 shows the membrane at t ¼ 1 s and
t ¼ 2 s in a sequence of three different mesh refinements. The refinement experiments include four levels of refinements
corresponding to the mesh sizes of 153, 303, 603, and 1203. The details of each simulation are presented in Table 3. The errors
of the total area in the Cauchy sequence are shown in Fig. 12 and Table 5. The convergence of the membrane is time
synchronized as in the string chord case. From Fig. 12 and Table 5, we can see that the errors of Cauchy sequence are reduced
approximately by half in each step as the average mesh size is reduced by half. This implies that the membrane spring–mass
system is also convergent in first order.
 Q. Shi et al. / Journal of Fluids and Structures 58 (2015) 20–39 33

Table 6
Young's modulus and Poisson ratio verification. In this table, Ed and νd are respectively Young's modulus and Poisson ratio in the current model with
Delingette correction, and E0 and ν0 are respectively Young's modulus and Poisson ratio in the previous model.

Young's modulus results ðGPaÞ Poisson Ratio results

Input Ed E0 Input νd ν0

Group 1
Case 1 0.1 0.096301 0.111815 0.14 0.141887 0.494577
Case 2 0.2 0.192606 0.223641 0.14 0.141920 0.494593
Case 3 0.3 0.288865 0.335500 0.14 0.141910 0.494562
Case 4 0.4 0.385234 0.447318 0.14 0.141913 0.494577
Case 5 0.5 0.481520 0.559100 0.14 0.141905 0.494565

Group 2
Case 1 0.1 0.095736 0.108928 0.22 0.223342 0.501076
Case 2 0.2 0.191491 0.217858 0.22 0.223332 0.501088
Case 3 0.3 0.287065 0.326758 0.22 0.223382 0.501037
Case 4 0.4 0.382989 0.435726 0.22 0.223349 0.501068
Case 5 0.5 0.478419 0.544658 0.22 0.223365 0.501068

Group 3
Case 1 0.1 0.095345 0.107268 0.30 0.305362 0.508593
Case 2 0.2 0.190661 0.214516 0.30 0.305342 0.508582
Case 3 0.3 0.286053 0.321794 0.30 0.305343 0.508601
Case 4 0.4 0.381401 0.429016 0.30 0.305359 0.508577
Case 5 0.5 0.476729 0.536299 0.30 0.305355 0.508587

NWA: Numerical results With Angular stiffness; NWOA: Numerical results WithOut Angular stiffness.

Fig. 13. Young's modulus (left) and Poisson's ratio (right). We tested the spring model by stretching the fabric surface to different lengths. The numerical
results show that the spring–mass model catches the fabric's Young's Modulus and Poisson ratio nicely in the linear regime of strain.

4.2. Verification of Young's modulus and Poisson's ratio

Young's modulus, also known as the tensile modulus or elastic modulus, is a measure of the stiffness of an elastic
material and is a parameter used to characterize materials. It is defined as the ratio of the stress along an axis over the strain
along that axis in the range of stress in which Hooke's law is valid. In solid mechanics, the slope of the stress–strain curve at
any point is called the tangent modulus. The tangent modulus of the initial, linear portion of a stress–strain curve is called
Young's modulus. Young's modulus, E, can be calculated by dividing the tensile stress by the tensile strain in the elastic
portion of the stress–strain curve:

tensile stress σ F=A0 Fl0

E¼ ¼ ¼ ¼ ; ð24Þ
tensile strain ϵ Δl=l0 A0 Δl

where E is Young's modulus, F is the force exerted on the object under tension, A0 is the original cross-sectional area through
which the force is applied, Δl is the change of length of the object from the original length l0 of the object.
Poisson's ratio is the negative ratio of transverse strain to axial strain. When a material is stretched, it usually tends to
contract in the directions transverse to the direction of stretching. The Poisson ratio is the ratio of relative contraction to
34 Q. Shi et al. / Journal of Fluids and Structures 58 (2015) 20–39

Fig. 14. Simulation of a table cloth draping under the action of the gravity. The fabric constraint automatically adjusts the regions ofpthe
ffiffiffiffiffiffiffiffifficloth. The spring
model of the fabric gives a realistic motion of the cloth. The characteristic eigen frequency for the fabric model in this simulation is k=m ¼ 1000 and the
friction constant is 0.1.

Fig. 15. The von-Mises formula (Eq. (14)) is used to calculate the fabric stress in the spring model. The left plot shows the von-Mises stress of a rectangular
membrane when pulled from the four corners. Similarly, the right plot shows the von-Mises stress of the triangular membrane pulled from its three
vertices.

relative stretching:
dϵtrans dϵy
ν¼  ¼ ; ð25Þ
dϵaxial dϵx
where ν is Poisson's ratio, ϵtrans is the transverse strain and ϵaxial is the axial strain. Theoretically, in the case of small
deformations Poisson's ratio ν was computed by Eq. (26) and in the case of large deformation it was computed by Eq. (27):
Δw=w0
ν¼  ; ð26Þ
Δl=l0

ν ¼  logð1 þ Δl=l0 Þ ð1 þ Δw=wÞ: ð27Þ

To verify that the spring model can catch isotropic elastic material's Young modulus and Poisson ratio, we carried out a
set of simulations by stretching fabric surfaces with different Young's modulus and Poisson ratios. These simulations start
with a fabric surface which has its original length l0 ¼ 0:1 m and original width w0 ¼ 0:02 m. The fabric is then pulled along
the direction of the longer side of it with a distributed force.
Firstly, three groups of simulations which stretch fabric surfaces with Poisson ratios of 0.14, 0.22, and 0.30 are carried out.
Each group with five different values of Young's modulus, that is 0.1 GPa, 0.2 GPa, 0.3 GPa, 0.4 GPa, and 0.5 GPa. The fabric
surface is pulled at one end. Young's modulus and Poisson ratio are calculated from the deformation of the surface and the
force added on it. The results are summarized in Table 6 which shows that the spring–mass model nicely reproduces the
values of Young's modulus and the Poisson ratio from the input. Poisson's ratio in Table 6 is computed by Eq. (26).
Next, a group of simulations, stretching the fabric surface whose Young's modulus and Poisson ratio are fixed at 0.5 GPa
and 0.14 respectively, are carried out. The total number of triangles of the triangulations used in these simulations changes
from 1143, 1590, 2468, to 4520. The strain of the elongation changes from 0.002, 0.004, 0.006, 0.008 to 0.01. The measured
Young's modulus and Poisson ratio from these simulations are demonstrated in Fig. 13.
The numerical solutions imply that the revised spring–mass model based on the derivation by Delingette (2008) can
accurately simulate an isotropic elastical membrane in the linear region with strain up to 0.01. The error between the spring
 Q. Shi et al. / Journal of Fluids and Structures 58 (2015) 20–39 35

Fig. 16. Von-Mises stress on the parachute canopy during its inflation. The red color shows regions with high stress. The figure shows that the areas near
the canopy–string connection points are the most stressful part of the parachute during inflation. (For interpretation of the references to color in this figure
caption, the reader is referred to the web version of this paper.)

Fig. 17. Drag force time history during the inflation phase of a C-9 personnel parachute. The experimental data is provided by Dr. Jean Potvin at St. Louis
University in a private communication, where YPG004, YPG005 and YPG008 represent three independent jump experiments.

model and continuum model increases when elongation is too large and the deformation reaches the nonlinear regime. For
parachute simulation, the strain of the canopy, even during the most dynamic phase of inflation, should still be in the linear
regime. Therefore the spring model is an excellent model for such simulations.
It should be mentioned that in the case in which the fabric surface is compressed and it adjusts itself with wrinkles, the
convergence is not obvious. In some cases, a small perturbation of the initial condition may result in substantially different
folding and wrinkling patterns. How to describe such case and define its mathematical convergence remain as an open
question to the model. Fig. 14 shows the draping of a table cloth due to gravitational force on a circular table. The right plot
shows the wrinkled edge of the cloth.

4.3. Computation of stress

The geometrical deformation of a fabric surface is the major source of stress on the material. The stress causes surface
tension in parachute canopy and exerts a normal component of force which the parachute canopy interacts with the
surrounding fluid to produce drag of the deceleration. The stress is also an important engineering variable in the parachute
safety design. Therefore accurate computation of stress will not only help to understand the fluid dynamics of the parachute,
but also to ensure that the material will not be ripped apart during the deceleration. The stress has been computed on a
rectangular and triangular fabric surface being stretched from point forces at their corners. The magnitude of von-Mises
stress on each individual triangle of the spring mesh is given in Fig. 15. In the rectangular fabric simulation, the total mass of
the membrane is 13 g and the spring constant is 1000 N/m. A total of 5461 mass points are in the triangulated mesh. The
color plot shows that the largest stress is near the pulling corner while the central part is relatively non-stressed. The right
36 Q. Shi et al. / Journal of Fluids and Structures 58 (2015) 20–39

Fig. 18. Angled deployment of C-9 parachute with the flow. The deployment starts with a 151 angle between the initial parachute and the direction of flow.
The parachute experiences only slight asymmetry of the canopy. The plots show the parachute at (from left to right) t ¼ 0 s, t ¼ 1:5 s and t ¼ 3:0 s.

Fig. 19. Angled deployment of C-9 parachute with the flow. This sequence starts with a 601 angle between the initial parachute and the direction of the
flow. In this case, the canopy skirt is dangerously wrapped at the lower side of the canopy. The plots show the parachute at (from left to right)
t ¼ 0 s; t ¼ 1:5 s and t ¼ 3:0 s.

plot shows the stress computed using Eq. (14). In the second case, the total mass is also 13 g with 1123 mass points. Fig. 16
shows the most stressful areas of a C-9 canopy during the inflation stage of the parachute. Here the peaks in stress are
located at the canopy–string connection points.

4.4. Comparison of drag in inflation

Parachute inflation is one of the most crucial stages for the deceleration system. Both experimental data and numerical
simulation results showed that there exists a time period during which the parachute canopy experiences a peaked drag force
that can be as high as 3–4 times of the payload. Potvin recorded the drag force personally from his jumps. The experimental
results (Fig. 17) reveal that the peak of the drag force climbed during the first second after the canopy deployment. The drag
force reached about 700 lbs at roughly t ¼ 1:3–1:5 s. At this time the air from bottom opening rushes to fill the volume of the
canopy causing expansion in both vertical and horizontal directions. The ballooning canopy exerts the majority of the drag force
to all the stiffer chords. In the simulations, the parachute drag force is recorded as the sum of forces on each chord. The total
force as a function of time is compared with the experimental data by Potvin. We have carried out simulations with different
initial shapes of the canopy. The flat and partially closed canopies show similar peaked drag at t ¼ 1:2–1:6 s, but the canopy
with an angled initial condition demonstrated a peak at the later time of t ¼ 2 s. Early in the simulations the drag force appears
oscillatory for several seconds following the highest peak. The dynamic evolution of the canopy geometry at these times
corresponds to the oscillatory motion of the horizontal motion of the string chord and the breathing motion of the canopy. Such
oscillation is reduced when a transverse damping force is added to the string chord.

4.5. Simulation of angled drop

The majority of parachute malfunctions occur during the inflation sequence. One of the most harmful malfunctions is the
canopy “inversion” which occurs when one or more gore sections near the skirt of the canopy blow between the suspension
lines on the opposite side of the parachute and then catch air (Adams, 2010). That portion then forms a secondary lobe with
 Q. Shi et al. / Journal of Fluids and Structures 58 (2015) 20–39 37

Fig. 20. Inversion of the parachute canopy during an angled drop. The alignment of the parachute started with a 751 angle with the direction of the
velocity. A complete inversion occurs at t ¼ 2 s. The two plots are views of the inverted canopy from different directions.

Fig. 21. The left plot is the pressure along a horizontal grid line across the two skirt sides of the parachute canopy and the right plot is the pressure along
the vertical grid line through the center of the canopy.

the canopy inverted. The condition may work out or may become a complete inversion i.e. the canopy turns completely
inside out (Butler and Crowe, 1999). Inversion during parachute inflation is dangerous as it can completely shut up the inlet
of the canopy and prevent the creation of an air volume under the canopy, thus reduces the drag force to essentially zero
and results in a free fall.
Numerical solution becomes a valuable tool for the parachute design if computer simulation can reveal and predict
malfunctions of the parachute canopy during the deployment. A group of different drops in which the initial alignments of
the parachute form different angles with the direction of the fluid velocity are simulated here. Fig. 18 shows the case in
which the alignment of canopy–string–payload forms a 151 angle with the fluid velocity. In this case, the canopy only
slightly loses its symmetry during the inflation, but the inflation is normal. The total adjustment to vertical fall takes a
longer time, but the opening of the canopy is just on time. In the case of the parachute forming a 601 angle with the flow, as
shown in Fig. 19, the side of the parachute facing the flow is dented and wrapped up and the opening time is increased. In
the case in which the parachute forms a 751 angle with the flow, the complete inversion of the canopy happens at
approximately t ¼ 2:0 s. Fig. 20 shows such inverted canopy.
The inflation force comes from the pressure difference between the two sides of the canopy which is obtained from the
projection method as an complementary variable. In Fig. 21, we plot the typical pressure profile long a horizontal grid line
and the pressure along a vertical grid line. Our impulse method (Kim et al., 2013) captures the momentum resulting from
the pressure difference and add it to the impulse of the spring mass points. Fig. 22 shows the vector fields of velocity and the
magnitude of vorticity around the canopies both with and without vent. It should be noted that the eddy viscosity does not
cover the wake of flow in the parachute simulation, especially at the wake of the flow after passing the canopy. A more
accurate turbulent model may be needed to resolve the flow leaving the parachute canopy.

5. Summary and discussion

A spring model for the simulation of fabric surface is discussed based on the modifications of Delingette. New components
to the parachute module have been included including the distinction of gore boundary, more sophisticated collision handling
and porosity. A dramatic computational speed up the spring model simulation is achieved through heterogeneous
38 Q. Shi et al. / Journal of Fluids and Structures 58 (2015) 20–39

Fig. 22. Velocity and magnitude of vorticity of fluid flow around parachute canopy in the simulations of the G-11 parachute without vent and C-9
parachute with vent.

computations where the most expensive portion of calculation is performed by the GPU. Numerical experiments verified first
order convergence of the spring model under the conditions that the total mass of the membrane is kept constant and that
both the tensile stiffness and angular stiffness of the spring conform with the material's Young's modulus and the Poisson
ratio. However, the convergence is weak in the case when the fabric is under compression and forms wrinkles.
The verification of the spring model on the material properties of the fabric membrane gives exciting results. Both
Young's modulus and Poisson ratio agree with the theory excellently in the linear regime in which the relative displacement
Δl=l is less than 10 percent. The difference increases as it reaches the nonlinear regime in which the relative displacement
exceeds 10 percent. However, since both parachute canopy and string chords are stiff materials and the relative stretches are
both smaller than 10 percent even during the most dynamic inflation phase, the spring model is sufficient for the simulation
of parachute deceleration system.
Our comparison of the inflation drag with the experimental data by Potvin shows agreement on the peak drag force at
approximately t ¼ 1:2–1:6 s. However, simulations demonstrate oscillatory variation of the drag force, or after-shock. By adding
transverse damping of string chord such oscillation can be reduced, but do not fully explain the discrepancies between
experiment and simulations. Future work will focus on the effects due to porosity and the wake of parachutist body or payload.
Parachute malfunctions during the inflation stage are dangerous and our numerical simulations show that parachute canopy
can undergo inversion when the parachute is dropped with an angle exceeding 601 between the parachute and the free fluid
stream velocity. The turbulent viscosity in our current simulations is from the approximate Baldwin–Lomax algebraic model
which has its limited validity. We are currently engaged in the implementation and comparison of more sophisticated turbulent
models. We plan to document such a study in our new papers. The initial states of parachute in our simulations are either from
flat or half-folded. The current fluid solver and interaction method still face challenge when the space under the folded canopy
is too small. The accuracy could be increased if adapted mesh refinement is added.

Acknowledgments

We would like to thank Dr. Joseph Myers of the ARO for fostering the collaborative relationship between authors at Stony
Brook University and the Army scientists, and Dr. Michael Kendra for the support with Air Force Summer Faculty Fellowship
 Q. Shi et al. / Journal of Fluids and Structures 58 (2015) 20–39 39

to Edwards AFB. We would like to give special thanks to Jean Potvin at St. Louis University for providing the experimental
data on the drag of the C-9 parachute, Alec Dyatt for giving us valuable suggestions on the direction of the research, and
Keerti Bhamidipati for many fruitful discussions while Xiaolin Li was working at the Edwards AFB. Also thanks to Yiyang
Yang for helping to make some of the figures. This work is supported in part by the US Army Research Office under the
award W911NF0910306, W911NF1410428 and the ARO-DURIP Grant W911NF1210357.

References

Adams, Douglas S., 2010. Lessons learned and flight experience from planetary parachute development. In: The Seventh International Planetary Probe
Workshop (IPPW7).
Baldwin, B.S., Lomax, H., 1978. Thin layer approximation and algebraic model for separated turbulent flow. In: AIAA 16th Aerospace Sciences Meeting,
January.
Brown, D., Cortez, R., Minion, M., 2001. Accurate projection method for the incompressible Navier Stokes equations. Journal of Computational Physics 168,
464–499.
Butler Jr., Manley C., Crowe, Michael D., 1999. The design, development and testing of parachutes using the bat sombrero slider. In: The 15th CEAS/AIAA
Aerodynamic Decelerator Systems Technology Conference.
Chorin, A.J., 1968. Numerical solution of the Navier Stokes equations. Mathematics of Computation 22, 745–762.
Cochrane, Cedric, Lewandowski, Maryline, Koncar, Vladan, 2010. A flexible strain sensor based on a conductive polymer composite for in situ measurement
of parachute canopy deformation. Sensors 10 (14248220), 8291–8303.
Delingette, Herve, 2008. Triangular springs for modeling nonlinear membranes. IEEE Transactions on Visualization and Computer Graphics 14 (March (2)),
723–731.
Du, Jian, Fix, Brian, Glimm, James, Jia, Xicheng, Li, Xiaolin, Li, Yunhua, Wu, Lingling, 2006. A simple package for front tracking. Journal of Computational
Physics 213, 613–628.
Fix, B., Glimm, J., Li, X., Li, Y., Liu, X., Samulyak R., Xu, Z., 2005. A TSTT integrated frontier code and its applications in computational fluid physics. Journal of
Physics: Conference Series 16, 471–475.
Gere, James M., 2004. Mechanics of Materials, 6th edition. Bill Stenquist, Belmont.
Glimm, J., Grove, J.W., Li, X.-L., Tan, D.C., 2000. Robust computational algorithms for dynamic interface tracking in three dimensions. SIAM Journal on
Scientific Computing 21, 2240–2256.
Glimm, J., Fix, B., Li, X.-L., Liu, J.-J., Liu, X.-F., Liu, T.-S., Samulyak, R., Xu, Z.-L., 2007. Front tracking under TSTT. Proceedings of the IGPP-CalSpace Conference;
Palm Springs, California, 26-30 March 2006. Astronomical Society of the Pacific, San Francisco.
Hamid, J., Desabrais, K.J., 2002. Aerodynamics of parachute opening ADA411095, ARO37594.6-EG. Mechanical Engineering Department, Worcester
Polytechnic Institute, pp. 1–131.
Johari, Hamid, Desabrais, Kenneth J., 2005. Vortex shedding in the near wake of a parachute canopy. Journal of Fluid Mechanics 536, 185–207.
Kim, Y., Peskin, C.S., 2006. 2-D parachute simulation by the immersed boundary method. SIAM Journal on Scientific Computing 28, 2294–2312.
Kim, Y., Peskin, C.S., 2009. 3-D parachute simulation by the immersed boundary method. Computers & Fluids 38, 1080–1090.
Kim, J.-D., Li, Y., Li, X.-L., 2013. Simulation of parachute FSI using the front tracking method. Journal of Fluids and Structures 37, 101–119.
Kirk, David B., Hwu, Wen-mei W., 2010. Programming Massively Parallel Processors: A Hands-On Approach. Morgran Kaufmann, Burlington.
Kuzmin, Dmitri, Mierka, Otto, Turek, Stefan, 2007. On the implementation of the k–epsilon turbulence model in incompressible flow solvers based on a
finite element discretisation. International Journal of Computing Science and Mathematics 1 (2), 193–206.
Langtangen, Hans Petter, Mardal, Kent-Andre, Winther, Ragnar, 2002. Numerical methods for incompressible viscous flow. Advances in Water Resources 25
(8), 1125–1146.
Lee, Long, LeVeque, Randall J., 2003. An immersed interface method for incompressible Navier–Stokes equations. SIAM Journal on Scientific Computing 25
(3), 832–856.
Lew, Adrian J., Buscaglia, Gustavo C., Carrica, Pablo M., 2001. A note on the numerical treatment of the k–epsilon turbulence model. International Journal of
Computational Fluid Dynamics 14 (3), 201–209.
Li, Y., I-Liang, J.-D., Kim, Chern, Li, X.-L., 2013. Numerical method of fabric dynamics using front tracking and spring model. Communications in
Computational Physics 14 (5), 1228–1251.
Matyka, Maciej, 2004. Solution to two-dimensional incompressible Navier–Stokes equations with simple, simpler and vorticity-stream function
approaches. Driven-lid cavity problem: solution and visualization. arXiv:0407002,2004 [physics.ins-det].
Peterson, Carl W., 1993. The fluid physics of parachute inflation. Physics Today 46, 32.
Potvin, Jean, McQuilling, Mark, 2011. The Bi-model: Using cfd in Simulations of Slowly-Inflating Low-Porosity Hemispherical Parachutes. AIAA Paper 2011-
2542, May 2011.
Potvin, J., Bergeron, K., Brown, G., Charles, R., Desabrais, K., Johari, H., Kumar, V., McQuilling, M., Morris, A., Noetscher, G., Tutt, B., 2011. The Road Ahead: A
White Paper on the Development, Testing and Use of Advanced Numerical Modeling for Aerodynamic Decelerator System Design and Analysis. AIAA
Paper 2011-2501, May.
Potvin, J., 1998. Parachute Inflation. McGraw-Hill 1998 Yearbook of Science and Technology. McGraw-Hill Education, New York.
Tau, Eric Yu, 1994. A second-order projection method for the incompressible Navier–Stokes equations in arbitrary domains. Journal of Computational
Physics 115 (1), 147–152.
Tezduyar, Tayfun E., Sathe, Sunil, 2007. Modelling of fluid–structure interactions with the space–time finite elements: solution techniques. International
Journal for Numerical Methods in Fluids 54 (6–8), 855–900.
Tutt, B.A., 2006. The application of a new material porosity algorithm for parachute analysis. In: The Ninth International LS-DYNA Users Conference.
Van Gelder, A., March 1998. Approximate simulation of elastic membranes by triangulated spring meshes. Journal of Graphics Tools 3 (March (2)), 21–41.
Wilcox, David C., et al., 1998. Turbulence Modeling for CFD, vol. 2. DCW industries la Canada, CA.
Yakhot, Victor, Smith, Leslie M., 1992. The renormalization group, the epsilon-expansion and derivation of turbulence models. Journal of Scientific
Computing 7 (1), 35–61.
Yihua, Cao, Qianfu, Song, Zhuo, W.U., Sheridan, John, 2010. Flow field and topological analysis of hemispherical parachute in low angles of attack. Modern
Physics Letters B 24, 1707–1725.
Yu, Li, Ming, Xiao, 2007. Study on transient aerodynamic characteristics of parachute opening process. Acta Mechanica Sinica 23, 627–633.