Uncertainty Quantification

Discontinuous Galerkin Methods for
Boussinesq-Green-Naghdi Equations
Nishant Panda ,Clint Dawson, Yao Zhang, Andrew B. Kennedy,
Joannes J. Westerink, Aaron Donnahue
Computational Hydraulics Group (CHG)
Institute for Computational Engineering and Sciences (ICES)
University of Texas (UT) at Austin
Department of Civil and Environmental and Earth Science
University of Notre Dame
IMUM 2013
September 18, 2013
What are Water Waves ?

Boussinesq theories
R-GN equations
Numerical Formulation using DG in 1D
verification and validation tests
Conclusions and future work
Outline
Boussinesq theories
R-GN equations
N.Panda
Green Naghdi Water Wave model and Numerical Simulation

Boussinesq theories
R-GN equations
Figure : A spectacular water wave.
N.Panda

Boussinesq theories
R-GN equations
So, What is a wave ? Very Crudely, a wave is

disturbance about a mean field
that propagates through a medium
and transports energy and not matter.
The obvious questions here being what causes a disturbance, and what
constitutes a medium? We will answer this in context of a water wave.
N.Panda

Boussinesq theories
R-GN equations
Water Waves schematic
Figure : Wave Characteristics.
N.Panda

Boussinesq theories
R-GN equations
Water Waves schematic cont.
Figure : Wave Characteristics contd.
N.Panda

Boussinesq theories
R-GN equations
Water waves propagating from deep water regions experience

significant transformations resulting in a rapid change in height,
speed and direction.
As depth decreases, waves become skewed about their crest with
marked steepening of the forward face until instability sets in
resulting in wave breaking.
Wave shoaling is described as the transformation of waves from near
shore zone until wave breaking.
N.Panda

Boussinesq theories
R-GN equations
While shallow assumptions are valid where the characterstic

wavelength (L) exceeds a typical depth (h0 ) by orders of magnitude
i.e kh0 1, non-linear near shore waves (where amplitude a and h0 are
comparable) have mostly been modeled through perturbation
techniques based on two non-dimensional parameters = kh0 and
= a/h0 first formulated by Boussinesq in 1872 and Rayleigh in
1876.
The smallness of is used to construct a polynomial representation
of the velocity field in the vertical co-ordinate which reduces a 3D
flow model to a 2D flow model. Moreover, the non-linear free
surface conditions are absorbed in the resulting equations which
makes it more tractable.
N.Panda

Boussinesq theories
R-GN equations
Some problems with Boussinesq theories
The scalings that are used in the perturbation analysis of Boussinesq

models can be severly restrictive. Wave shoaling is known to occur
when 1, while breaking is known to occur when 1.
Most of the Bossinesq models also assume irrotational flow field and
are hence valid up to the breaking point. Since, vortices are
generated from wave breaking, any model based on irrotational flow
will induce large errors in the velocity field.
N.Panda

Boussinesq theories
R-GN equations
Green - Naghdi
An alternate approach to the computation of shallow water

nonlinear dispersive waves lies in the Green-Naghdi formulation,
where a polynomial structure for the velocity field is retained
without irrotational assumption.
The mass and momentum equations were solved in a weighted
residual sense.
N.Panda

Boussinesq theories
R-GN equations
Currently there are two broad categories of Green Naghdi systems .

1 Yao et al desingned a new Rotational Boussinesq theory that
removes irrotational assumption by using higher order polynomial
expansion of velocity field on depth integrated continuity and euler
equations. These equations are built on work done by Madsen,Chen,
Nwogu etc.
2 A fully non-linear but weakly dispersive correction to Shallow
water equations derived by Bonneton et al, where the velocity field is
assumed to be irrotational.
N.Panda

Boussinesq theories
R-GN equations
Usually, in water wave theory one works with the non-dimensional Euler
equations for an incompressible fluid. The continuity equation reduces to
the free surface equation given by,
Z
+ .
udz = 0.
(1)
t
hb
where = (x, y , t) is the free surface. The non-dimensional momentum
equations, in Cartesian co-ordinates, are given by
u
u
+ u.u + w
+ P = 0.
t
z
2
w
w
P
+ 2 u.w + 2 w
+
+ g = 0.
t
z
z
N.Panda
(2)
(3)

Boussinesq theories
R-GN equations
Integrating (3) from z to , and assuming a zero gauge pressure at the

free surface, we find
Z
Z
Z
w
w
2
2
2
P(z) =
dz +
u.wdz +
w
dz + g ( z). (4)
t
z
z
z
z
where, = [/x, /y ]T , u = [u, v ]T and represents a
dimensionless wave number.
N.Panda

Boussinesq theories
R-GN equations
Constructing the Boussinesq Green Naghdi Equations

Here an approximate velocity field given by
uu
=
N
X
n un (x, y , t)fn (q).
n=0
is inserted into the Euler equations to get arbitrary levels of

approximation.
1. Define a level of wave approximation, O(N ).
2. Insert the approximate velocity field into the free surface equation ,
retaining all the terms up to the desired level of approximation.
3. Insert the approximate velocity field into the pressure equation to
get P.
4. Insert the approximate velocity field into the horizontal momentum
equation.
N.Panda

Boussinesq theories
R-GN equations
Integrate in weighted residual sense, using the N + 1 basis functions used

in the approximate velocity field, i.e
Z
fm (
+u
.
u+w
+ P)dz
= 0.
m = 0...N
(5)
t
z
hb
where w
represents the approximate vertical velocity field which can be
determined from the approximate horizontal velocity field.
N.Panda

Boussinesq theories
R-GN equations
In this talk we focus mainly on the O(2 ) equations. The approximate

velocity field is given by
u
= u0 + 2 u1 f1 (q) + 2 u2 f2 (q),
(6)
w
= .u0 Hq u0 .hb ,
where q is a sigma-type co-ordinate given by q =
H = + hb (x, y ) is the total water depth.
N.Panda
z+hb
hb +
and

Boussinesq theories
R-GN equations
with these assumptions the Boussinesq-Green-Naghdi system has the

following surface elevation equation,
,t + .(u0 H + 2
2
X
um Hcm ) = 0.
(7)
m=1
The momentum equations are given by,

u0,t Hc1m
u0 .u0 Hc2m
g Hc3m
2
X
(un,t Hc4m un ,t c5m )
n=1
1
(.u0,t )H 3 c6m + .u0,t HH 2 c7m + (u0,t .hb )H 2 c8m
2

m
+ u0,t .hb Hc9m (.u0,t )H 2 hb c10
2
+ 2
2
X
m
m
[(un .u0 + u0 .un )Hc11
un .(u0 H)c12
]
n=1
N.Panda

Boussinesq theories
R-GN equations

m
m
+ 2 H 2 (.u0 )2 u0 .(.u0 ) (c13
+ hb c14
)
2

m
+ H 3 (.u0 )2 u0 .(.u0 ) c15

2
m
m
2 Hu0 .(u0 .hb )c16
2 H 2 [u0 .(u0 .hb )] c17
= 0,
(8)
m [0, 2].
N.Panda

Boussinesq theories
R-GN equations
Lets look at the 1D case

Am u0,t + Bm u1,t + Cm u2,t + am u0,xt + bm u0,xxt + cm u0 u0,x
2
+dm u0,x
+ em u0 u0,xx + fm u0,x u0,xx + gm u0 u0,xxx
+hm u02 + im u1 u0,x + jm u0 u1,x + km u2 u0,x + lm u0 u2,x

+nm u1 u0 + 0m u2 u0 + pm u1 + qm u2 + rm g x = 0
m = 0..2
along with the surface elevation equation :
t +
(1)
(1)
(u0 ( + h) + 2 u1 ( + h)g1 + 2 u2 ( + h)g2 ) = 0
x
N.Panda

Boussinesq theories
R-GN equations
At first glance, looking into such higher derivatives, you may ask why
DG. The reasons :
The R-GN equations have non-linear higher derivatives terms that
needs to be computed accurately, The noncoservative nature of the
equations can be handled efficiently through LDG.
DG methods offer an easy extension to unstructured grids. In fact,
this has been the sole reason why there arent any 2D
implementation of GN equations.
Built in stabilization mechanisms through efficient use of jumps.
N.Panda

Boussinesq theories
R-GN equations
N.Panda

Boussinesq theories
R-GN equations
Let = [0, L] be the spatial domain. Define a partition

0 = x1/2 < x3/2 < . . . < xJ+1/2 = L,
and define,

Ej = xj1/2 , xj1/2 ,

E = xj+1/2 ,
hj = xj+1/2 xj1/2 ,
(9)
h = max hj ,
j
to be the finite element, set of boundary points, element size and the
maximum element size respectively. Construct a set of test functions VhK
on the partition, consisting of piecewise polynomials of degree K :
VhK = {v : v |Ej PK (Ej )
N.Panda
j = 1, . . . , J}.
(10)

Boussinesq theories
R-GN equations
Let us denote,
+
v (xj+1/2
) = lim+ v (xj+1/2 + ),
0
v (xj+1/2
)
= lim+ v (xj+1/2 ).

0
Then, we can define the jump and average of v at the endpoints of Ej as :

+
|v (xj+1/2 )| = v (xj+1/2
) v (xj+1/2
),

(11)

1
+
v (xj+1/2
) + v (xj+1/2
) .
v (xj+1/2 ) =
2
For any v VhK , we can write v as
v=
J X
K
X
vij i (x),
(12)
j=1 i=0
where {i } is a basis for PK . In this paper we chose i = Pi , where Pi is

the normalized Legendre polynomial
N.Panda

Boussinesq theories
R-GN equations
Given u h VhK , all derivatives of u h are calculated in an LDG sense

described below. Define :
h = uxh ,
B(h , w ) = Luh (w ),
where B : VhK VhK R is the bi-linear form and Luh : VhK R is the
linear form given by
X
B(h , w ) =
(h , w )Ej ,
j
Luh (w ) =

(u h , wx )Ej + uh , [|w |] E ,
(13)
where w VhK and ( , ) denotes the standard L2 inner product. In a

h
h
similar fashion, we compute uxx
, uxxx
and so on.
N.Panda

Boussinesq theories
R-GN equations
Looking ahead, let us define the following bi-linear form :

X h

B u h , w =
u , w E + |u h | , [|w |] E ,
j
(14)
Where 0. Note, uh = F (u h , u h+ ) is the single valued flux function

evaluated at the edges of Ej . Various flux functions can be found in the
DG literature. The simplest flux is the average flux given by :

+
F (uj+1/2
, uj1/2
) = u(xj+1/2 ) .
(15)
N.Panda

Boussinesq theories
R-GN equations
some insight through linearized equations

Before we begin to solve the complete equations
it is better to look at

the linearized versions. The linearized O 2 R-GN equations can be
written as:
2
X
Amn (hb ) un,t + [Bm0 (hb ) u0,xt + Cm0 (hb ) u0,xxt ] + gm x = 0,
n=0
m = 0 . . . 2.
t +
2
X
(Dn (hb ) un ),x = 0.
n=0
To keep our analysis simple we choose the shifted Legendre polynomials

in (6) which decouples u1 and u2 above and hence it is sufficient only to
look at the following equation:
u0,t c0 hb2 u0,xxt + g x = 0,
t + hb u0,x = 0.
N.Panda
(16)

Boussinesq theories
R-GN equations
discrete stability of the linearized equations : ENERGY

method
Let us rewrite (16) as a system of first order (in space) equations :
r u0,xt = 0,
u0,t
c0 hb2 r,x
(17a)
+ g ,x = 0,
,t + hb u0,x = 0.
(17b)
(17c)
For simplicity let us assume u(0) = u(L) = 0. Adding (17c) and (17b)
and subtracting (17a) after multiplication by g , hb u0 and c0 hb3 u0,x
respectively and integrating from 0 to L we get:
g (,t , ) + hb (u0,t , u0 ) + c0 hb3 (u0,xt , u0,x ) = 0.
Hence, to show stability of the numerical method it is sufficient to show

h
h
h
g ,th , h + hb u0,t
, u0h + c0 hb3 u0,xt
, u0,x
+ = 0.
such that maintains the energy when integrated in time.
N.Panda

Boussinesq theories
R-GN equations
Working with the discrete versions of (17a) , (17b) and (17c) and
following the steps above after performing integration by parts, we get
for an element Ej :

h
h
h
g ,th , h E + hb u0,t
, u0h E + c0 hb3 u0,xt
, u0,x
+ Ej = 0.
(18)
E
j
where, Ej is given by:

Z
ghb
h u0h
r h u0h
xj+1/2
h u0h x
j1/2
xj+1/2
rh u0h x
j1/2
Ej
c0 hb3
Ej
xj+1/2
u0h h x
j1/2
xj+1/2
u0h r h x
j1/2
!
,
(19)
!
.
Adding over the elements we get:

h
h
h
g ,th , h + hb u0,t
, u0h + c0 hb3 u0,xt
, u0,x
+ = 0,
N.Panda
(20)

Boussinesq theories
R-GN equations
where = I + II + B.T . We see that I, II are given by:

X

,
I = ghb
|u0h | h h + | h | u0h u0h
Ei
II = c0 hb3
X

h
h

|r h |
u0 u0h + |u0h |
r rh .
Ei
and similar expression for the boundary terms B.T . Here, Ei represents
the set of interior edges. F
N.Panda

Boussinesq theories
R-GN equations

From the above expressions
it is easy
to see that if we choose u0h = u0h ,

h
| in the interior and u0h = 0,
h = h and rh = r h 11 |u0,t
h
h
h
h
= and r = r at the boundaries 0 and L respectively, we get:
=
1
d X h 2
c0 hb3 11
|u0 | ,
2
dt
(21)
Ei
and integrating in time we get the desired stability.
N.Panda

Boussinesq theories
R-GN equations
With this in mind, lets look at the complete non-linear equations The
R-GN equations can be written as :
= Rhs ,
(22a)
L [s0 ] = Rhsu0 ,
(22b)
s1 = Rhs1 ,
(22c)
s2 = Rhs2 ,
(22d)
where = t , s0 = u0,t , s1 = u1,t , s2 = u0,t ; and L is an elliptic operator
2
C x
given by A + B x
2 , where A,B,C are :
A = H g0 2 hx x H g0 ,
B = 2 H 2 Hx g0 2 hb,x H 2 (
g0 s0 ) + 2 H 2 hb,x s0 ,
(23)
2 3
C=
H (
g0 0 ).
2
N.Panda

Boussinesq theories
R-GN equations
The weak formulation of the R-GN equations (22) is then to find :

h VhK ,
s0h VhK ,
s1h VhK ,
s2h VhK ,
(24)
r h VhK ,
p h VhK ,
where r h , p h approximate s0,x and s0,xx respectively, such that,
N.Panda

Boussinesq theories
R-GN equations

B h , = L1 () ,

Bs s0h , + Br r h , + Bp p h , = L2 () ,

B s1h , = L3 () ,

B s2h , = L4 () ,
(25a)
(25b)
(25c)
(25d)
where B is defined in (14). Bs , Br and Bp are given by :

X

As0h , w E ,
Bs s0h , w =
j

Br r h , w =
Br h , w
Ej
(26)
X

Bp p h , w =
Cp h , w E ,
j
N.Panda

Boussinesq theories
R-GN equations
where A, B and C are defined in (23). To eliminate, r h and p h we define

the following equations :
X
X

r h, w E =
(s0h , wx )Ej + s0h , [|w |] E ,
j
X
j
j
h
p ,w

Ej
r h , wx

Ej

+ rh , [|w |] E 11 h[|s|] , [|w |]iE .
(27)
The constant 11 is chosen to be of O(1/h).
N.Panda

Boussinesq theories
R-GN equations
Why do we need a penalty in the time derivatves ? t = = Rhs our

Rhs is not resolved due to the discontinuous functions, high
non-linearity and improper knowledge of the spaces. Hence will not be
resolved correctly and this will lead to instability.
N.Panda

Boussinesq theories
R-GN equations
The Ghost of Aliasing errors
p - Aliasing. The task of computing higher order derivatives upto the

desired level of accuracy is very challenging. The R-GN equations are
demanding in the sense that it requires all the derivatives to be computed
correctly. Consider, a function f = sinx. Project f into a linear DG space.
Then Compute fxx and fxxx using LDG.
N.Panda

Boussinesq theories
R-GN equations
Figure : Second Order Derivative.
N.Panda

Boussinesq theories
R-GN equations
Figure : Third Order Derivative.
N.Panda

Boussinesq theories
R-GN equations
One solution: Take k = 4.
Figure : Higher Order Space.
N.Panda

Boussinesq theories
R-GN equations
Linear standing waves

The R GN equations as such dont have any known exact analytic
solutions. However it is known that for horizontal bottoms, a linear
standing wave solution exists. We choose a linear standing wave given by
a/hb = 0.01, and impose wall boundary conditions and the following
initial conditions :
(x, 0) = acos kx,
u0 (x, 0), u1 (x, 0), u2 (x, 0) = 0.
(28)
where a, and k represents the amplitude, and wave-number (2/L)

respectively. The linearized Boussinesq equation for a standing wave
admits an exact solution given by
= acos(kx) cos(t).
N.Panda

Boussinesq theories
R-GN equations
Figure : L2 error convergence plots for the linearized equations.
N.Panda

Boussinesq theories
R-GN equations
Transformation of wave train over a submerged shoal
In this experiment, a wave train propagates towards a submerged

trapezoidal shoal.
Linear behavior is exhibited before the bar, while non-linear shoaling
causes steepening as the waves interact with the slope. Complex
multi-frequency waves are generated after the bar as bound
harmonics are released in deeper water at the top of the bar.
This experiment tests both the linear dispersion(after the bar) and
the non-linear characteristics of the model.
N.Panda

Boussinesq theories
R-GN equations
Initial Configuration
The initial wave train has a period of Tp = 2.02s and wave height
2a = 2cm. The mean water depth is hb = 0.4m.
Figure : Intial configuration for validation case
N.Panda

Boussinesq theories
R-GN equations
(a) x = 10.5m
(b) x = 12.5m
N.Panda

Boussinesq theories
R-GN equations
(c) x = 13.5m
(d) x = 15.7m
N.Panda

Boussinesq theories
R-GN equations
(e) x = 17.3m
(f) x = 19.0m
Figure : Comparison of experimental and numerical surface elevation variation

at different locations around the shoal
N.Panda

Boussinesq theories
R-GN equations
Wave reflection of solitary wave from a vertical wall
Solitary wave reflection exhibits complex non-linear and dispersive

phenomena and has been used as a validation case for numerous
numerical models based on Boussinesq-Green-Naghdi equations.
Experimental
Experimental observations have revealed that solitary waves emerging
from a collision, in addition to having experienced changes in their
phases, were trailed by a dispersive wave train. Moreover, for large
amplitudes, the maximum run-up was observed to be higher than those
determined from linear theory.
N.Panda

Boussinesq theories
R-GN equations
Figure : Intial configuration for the validation case

N.Panda

Boussinesq theories
R-GN equations
Figure : Maximum Surface elevation Vs the initial amplitude

N.Panda

Boussinesq theories
R-GN equations
A discontinuous Galerkin method was investigated for the

Green-Naghdi equations in resolving near shore water waves.
The numerical scheme was validated in 1D against challenging
experiments.
Linear stablity was derived and non-linear stablity was observed.
Future work will extend these equations to include moving shoreline,
wave breaking and wave generation and absorption to construct a
surf zone model.
Harry Potter seems more real than fast CPU times right now.
N.Panda

Boussinesq theories
R-GN equations
THANK YOU !!
N.Panda

Uncertainty Quantification

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Uncertainty Quantification

Uploaded by

Copyright:

Available Formats

Discontinuous Galerkin Methods for

September 18, 2013

What are Water Waves ?

What are Water Waves ?

Numerical Formulation using DG in 1D

verification and validation tests

Conclusions and future work

Green Naghdi Water Wave model and Numerical Simulation

What are Water Waves ?

Figure : A spectacular water wave.

Green Naghdi Water Wave model and Numerical Simulation

What are Water Waves ?

So, What is a wave ? Very Crudely, a wave is

Green Naghdi Water Wave model and Numerical Simulation

What are Water Waves ?

Water Waves schematic

Figure : Wave Characteristics.

Green Naghdi Water Wave model and Numerical Simulation

What are Water Waves ?

Water Waves schematic cont.

Figure : Wave Characteristics contd.

Green Naghdi Water Wave model and Numerical Simulation

What are Water Waves ?

Water waves propagating from deep water regions experience

Green Naghdi Water Wave model and Numerical Simulation

What are Water Waves ?

While shallow assumptions are valid where the characterstic

Green Naghdi Water Wave model and Numerical Simulation

What are Water Waves ?

Some problems with Boussinesq theories

The scalings that are used in the perturbation analysis of Boussinesq

Green Naghdi Water Wave model and Numerical Simulation

What are Water Waves ?

An alternate approach to the computation of shallow water

Green Naghdi Water Wave model and Numerical Simulation

What are Water Waves ?

Currently there are two broad categories of Green Naghdi systems .

Green Naghdi Water Wave model and Numerical Simulation

What are Water Waves ?

Green Naghdi Water Wave model and Numerical Simulation

What are Water Waves ?

Integrating (3) from z to , and assuming a zero gauge pressure at the

Green Naghdi Water Wave model and Numerical Simulation

What are Water Waves ?

Constructing the Boussinesq Green Naghdi Equations

n un (x, y , t)fn (q).

is inserted into the Euler equations to get arbitrary levels of

Green Naghdi Water Wave model and Numerical Simulation

What are Water Waves ?

Integrate in weighted residual sense, using the N + 1 basis functions used

Green Naghdi Water Wave model and Numerical Simulation

What are Water Waves ?

In this talk we focus mainly on the O(2 ) equations. The approximate

Green Naghdi Water Wave model and Numerical Simulation

What are Water Waves ?

with these assumptions the Boussinesq-Green-Naghdi system has the

The momentum equations are given by,

(un,t Hc4m un ,t c5m )

Green Naghdi Water Wave model and Numerical Simulation

What are Water Waves ?

+ H 3 (.u0 )2 u0 .(.u0 ) c15

Green Naghdi Water Wave model and Numerical Simulation

What are Water Waves ?

= lim+ v (xj+1/2 ).