Sensitivity Analysis for Shallow Water Equations - Advanced Study ...
Sensitivity Analysis for Shallow Water Equations - Advanced Study ...
Sensitivity Analysis for Shallow Water Equations - Advanced Study ...
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
<strong>Sensitivity</strong> <strong>Analysis</strong> <strong>for</strong> <strong>Shallow</strong> <strong>Water</strong> <strong>Equations</strong><br />
Vani Cheruvu<br />
<strong>Advanced</strong> <strong>Study</strong> Program Research Review<br />
National Center <strong>for</strong> Atmospheric Research<br />
June 14, 2007<br />
Vani Cheruvu (ASP Research Review) Adjoint <strong>Sensitivity</strong> <strong>Analysis</strong> 06/14/2007 1 / 30
Organization of the talk<br />
Motivation<br />
Adjoint <strong>Sensitivity</strong> <strong>Analysis</strong> <strong>for</strong> SWE<br />
◮ Direct <strong>Sensitivity</strong> Method<br />
◮ Adjoint <strong>Sensitivity</strong> Method<br />
◮ SWE and its Adjoint<br />
◮ Spectral Element Method<br />
Spectral finite volume method <strong>for</strong> SWE<br />
Wavelet based limiter <strong>for</strong> discontinuous Galerkin Method<br />
◮ Gibbs phenomenon<br />
◮ Limiters<br />
Summary<br />
Vani Cheruvu (ASP Research Review) Adjoint <strong>Sensitivity</strong> <strong>Analysis</strong> 06/14/2007 2 / 30
Vani Cheruvu (ASP Research Review) Adjoint <strong>Sensitivity</strong> <strong>Analysis</strong> 06/14/2007 3 / 30
Vani Cheruvu (ASP Research Review) Adjoint <strong>Sensitivity</strong> <strong>Analysis</strong> 06/14/2007 4 / 30
Chennai’s coastline be<strong>for</strong>e and after tsunami<br />
Vani Cheruvu (ASP Research Review) Adjoint <strong>Sensitivity</strong> <strong>Analysis</strong> 06/14/2007 5 / 30
Chennai’s coastline be<strong>for</strong>e and after tsunami<br />
Vani Cheruvu (ASP Research Review) Adjoint <strong>Sensitivity</strong> <strong>Analysis</strong> 06/14/2007 6 / 30
Chennai’s Marina beach after tsunami<br />
Vani Cheruvu (ASP Research Review) Adjoint <strong>Sensitivity</strong> <strong>Analysis</strong> 06/14/2007 7 / 30
Chennai’s Marina beach after tsunami<br />
Vani Cheruvu (ASP Research Review) Adjoint <strong>Sensitivity</strong> <strong>Analysis</strong> 06/14/2007 8 / 30
Why an Adjoint Equation?<br />
Diffusion equation defined in the domain x ∈ (0, L) is<br />
Objective function :<br />
where r = δ(x − x T ).<br />
Adjoint Equation<br />
Objective function:<br />
where f = Mδ(x − x s ).<br />
D ∂2 C<br />
+ f = 0, C(0) = C(L) = 0<br />
∂x 2<br />
J =<br />
J =<br />
∫ L<br />
0<br />
r(x)C(x) dx<br />
D ∂2 C ∗<br />
∂x 2 + r = 0<br />
∫ L<br />
0<br />
f (x)C ∗ (x) dx<br />
Vani Cheruvu (ASP Research Review) Adjoint <strong>Sensitivity</strong> <strong>Analysis</strong> 06/14/2007 9 / 30
Derivation<br />
∫ L<br />
0<br />
[ ] [<br />
C ∗ D ∂2 C<br />
∂x 2 + f dx = D C ∗ ∂C ] L [ ] ∂C<br />
∗ L<br />
− D<br />
∂x<br />
0<br />
∂x C 0<br />
∫ L<br />
+<br />
[CD ∂2 C ∗ ]<br />
∂x 2 + C ∗ f dx = 0<br />
0<br />
J =<br />
∫ L<br />
0<br />
[CD ∂2 C ∗<br />
∂x 2<br />
]<br />
+ C ∗ f dx +<br />
∫ L<br />
0<br />
rCdx<br />
J =<br />
∫ L<br />
0<br />
C<br />
[D ∂2 C ∗<br />
∂x 2 + r ]<br />
dx +<br />
∫ L<br />
0<br />
fC ∗ dx<br />
Vani Cheruvu (ASP Research Review) Adjoint <strong>Sensitivity</strong> <strong>Analysis</strong> 06/14/2007 10 / 30
<strong>Shallow</strong> <strong>Water</strong> <strong>Equations</strong><br />
Objective Function<br />
∂η<br />
∂t + ∂(Hu′ + Ūη)<br />
∂x<br />
∂u ′<br />
∂t + ∂(Ūu′ + gη)<br />
∂x<br />
= 0<br />
= 0<br />
J =<br />
∫ T ∫ L<br />
0<br />
0<br />
r(η, u ′ ; x, t)dxdt<br />
t= time; x = distance along the channel; η = depth of flow;<br />
u ′ = discharge per unit width; r = user-defined measuring function<br />
r = 1 2 [η(x 0, t) − ¯η(x 0 )] 2 δ(x − x 0 )<br />
Vani Cheruvu (ASP Research Review) Adjoint <strong>Sensitivity</strong> <strong>Analysis</strong> 06/14/2007 11 / 30
Direct <strong>Sensitivity</strong> Method<br />
The basic problem is solved<br />
A reference value of the objective function J 0 is computed<br />
A flow variable is perturbed at the boundary at the selected<br />
perturbation time<br />
The associated objective function value is computed again<br />
This procedure is repeated <strong>for</strong> sufficiently many selections of time<br />
This yields the temporal evolution of the sensitivities<br />
Vani Cheruvu (ASP Research Review) Adjoint <strong>Sensitivity</strong> <strong>Analysis</strong> 06/14/2007 12 / 30
Adjoint <strong>Sensitivity</strong> Method<br />
Derivation of 1D Adjoint <strong>Equations</strong><br />
Adjoint variables : φ(x, t) and ψ(x, t)<br />
Multiply the continuity equation by φ(x, t) and the momentum<br />
equation by ψ(x, t)<br />
The sum of these two products is then integrated over space and time<br />
Flow control is desired, hence the variation with respect to η and u ′<br />
Variation of the objective function with respect to η and u ′<br />
Add the variations<br />
Choose initial condition <strong>for</strong> the adjoint problem<br />
Choose boundary values <strong>for</strong> the adjoint variables<br />
Vani Cheruvu (ASP Research Review) Adjoint <strong>Sensitivity</strong> <strong>Analysis</strong> 06/14/2007 13 / 30
<strong>Shallow</strong> water equations and its Adjoint<br />
Adjoint<br />
∂η<br />
∂t + Ū ∂η<br />
∂x + ¯V ∂η<br />
∂y + H ∂u′<br />
∂x + H ∂v ′<br />
∂y<br />
∂u ′<br />
∂t + Ū ∂u′<br />
∂x + ¯V ∂u′<br />
∂y + g ∂η<br />
∂x<br />
∂v ′<br />
∂t + Ū ∂v ′<br />
∂x + ¯V ∂v ′<br />
∂y + g ∂η<br />
∂y<br />
= 0<br />
= 0<br />
= 0<br />
∂φ<br />
∂τ − Ū ∂φ<br />
∂x − ¯V ∂φ<br />
∂y − g ∂ψ 1<br />
∂x − g ∂ψ 2<br />
∂y + ∂r<br />
∂η<br />
= 0<br />
∂ψ 1<br />
∂τ − Ū ∂ψ 1<br />
∂x − ¯V ∂ψ 1<br />
∂y − H ∂φ<br />
∂x + ∂r<br />
∂u ′ = 0<br />
∂ψ 2<br />
∂τ − Ū ∂ψ 2<br />
∂x − ¯V ∂ψ 2<br />
∂y − H ∂φ<br />
∂y + ∂r<br />
∂v ′ = 0<br />
Vani Cheruvu (ASP Research Review) Adjoint <strong>Sensitivity</strong> <strong>Analysis</strong> 06/14/2007 14 / 30
Solution of <strong>Shallow</strong> <strong>Water</strong> <strong>Equations</strong><br />
Contour lines of Gaussian Hill; 1000 time steps with ∆t = 0.0005<br />
400<br />
350<br />
300<br />
250<br />
200<br />
150<br />
100<br />
50<br />
50 100 150 200 250 300 350 400<br />
Vani Cheruvu (ASP Research Review) Adjoint <strong>Sensitivity</strong> <strong>Analysis</strong> 06/14/2007 15 / 30
Solution of Adjoint <strong>Shallow</strong> <strong>Water</strong> <strong>Equations</strong><br />
Contour lines of Gaussian Hill; 500 time steps with ∆t = 0.0005<br />
400<br />
350<br />
300<br />
250<br />
200<br />
150<br />
100<br />
50<br />
50 100 150 200 250 300 350 400<br />
Vani Cheruvu (ASP Research Review) Adjoint <strong>Sensitivity</strong> <strong>Analysis</strong> 06/14/2007 16 / 30
Spectral Element Method<br />
The domain [a, b] is partitioned into N x non-overlapping intervals<br />
called elements.<br />
I j = [x j−1/2 , x j+1/2 ], j = 1, . . . , N x , and ∆x j = (x j+1/2 − x j−1/2 )<br />
I<br />
j!1<br />
I j<br />
I j+1<br />
x x x x<br />
j!3/2<br />
j!1/2<br />
j+1/2<br />
j+3/2<br />
Map every Element I j onto the reference element [−1, +1].<br />
Introduce a local coordinate ξ ∈ [−1, +1] s.t.,<br />
ξ = 2 (x − x j)<br />
∆x j<br />
, x j = (x j−1/2 + x j+1/2 )/2 ⇒ ∂<br />
∂x = 2<br />
∆x j<br />
∂<br />
∂ξ .<br />
Vani Cheruvu (ASP Research Review) Adjoint <strong>Sensitivity</strong> <strong>Analysis</strong> 06/14/2007 17 / 30
SE-1D: Space Discretization<br />
Regular Element<br />
I j<br />
Reference Element<br />
!<br />
x j!1/2<br />
x j+1/2<br />
! 1<br />
GLL Grid<br />
+1<br />
How do the neighboring elements talk?<br />
Nodal - Averaging the values on the element boundaries<br />
Time Integration Scheme<br />
Leapfrog Scheme u n+1 = u n−1 − 2∆tF (u n )<br />
The solution within each element is interpolated with a high-order<br />
polynomial using nodal basis set.<br />
Vani Cheruvu (ASP Research Review) Adjoint <strong>Sensitivity</strong> <strong>Analysis</strong> 06/14/2007 18 / 30
SFV in 1D<br />
For a given order of accuracy k, each SV is divided into k control<br />
volumes (CV)<br />
The ith CV of jth SV is : C j,i = [x j,i−1/2 , x j,i+1/2 ] and with cell width<br />
∆x j,i = x j,i+1/2 − x j,i−1/2<br />
x<br />
CV<br />
j,i!1/2<br />
x j,i+1/2<br />
x j!1/2<br />
SV<br />
I j<br />
x j+1/2<br />
Discontinuous Galerkin: Flux evaluation - Riemann Solver<br />
Spectral Finite Volume: Numerical flux at element boundaries and<br />
analytical flux at the cell boundaries<br />
Vani Cheruvu (ASP Research Review) Adjoint <strong>Sensitivity</strong> <strong>Analysis</strong> 06/14/2007 19 / 30
SFV-1D: Flux term<br />
(<br />
_<br />
) ( + ) (_<br />
) ( + )<br />
L R L R<br />
x j!1/2<br />
I j<br />
x j+1/2<br />
Flux function F (U) is not uniquely defined at x j±1/2<br />
F (U) is replaced by a numerical flux function ˆF (U), dependent on<br />
the left and right limits of the discontinuous function U<br />
Typical flux <strong>for</strong>mulae: Godunov, Lax-Friedrichs, Roe, HLLC, etc<br />
Vani Cheruvu (ASP Research Review) Adjoint <strong>Sensitivity</strong> <strong>Analysis</strong> 06/14/2007 20 / 30
2D Grid<br />
X<br />
X<br />
X<br />
X<br />
X<br />
X<br />
CV<br />
X<br />
X<br />
X<br />
X<br />
X<br />
X<br />
Given SV is partitioned into 4 × 4 CVs using 5 × 5 GLL nodes<br />
3 GL nodes are used <strong>for</strong> the evaluation of boundary integral on each<br />
side of a CV<br />
Ū ij are centered at each CV<br />
Vani Cheruvu (ASP Research Review) Adjoint <strong>Sensitivity</strong> <strong>Analysis</strong> 06/14/2007 21 / 30
<strong>Sensitivity</strong> <strong>Equations</strong><br />
Sensitivities required <strong>for</strong> subcritical flow control are given by<br />
δJ<br />
δu ′ (0,t) and<br />
δJ<br />
δη(L,t)<br />
Sensitivities required <strong>for</strong> supercritical flow control are given by<br />
δJ<br />
δu ′ (0,t) and<br />
δJ<br />
δη(0,t)<br />
To derive expressions <strong>for</strong> sensitivities in terms of the adjoint variables<br />
Initial condition <strong>for</strong> the basic problem<br />
η(x, 0) = η 0 (x), u ′ (x, 0) = u ′ 0 (x); δη(x, 0) = 0, δu′ (x, 0) = 0<br />
Initial conditions in the adjoint problem<br />
φ(x, T ) = 0 = ψ(x, T )<br />
Selection of boundary conditions <strong>for</strong> the adjoint problem is motivated<br />
by the boundary conditions of the basic problem<br />
Expressions are then derived <strong>for</strong> sensitivities to these boundary values<br />
Vani Cheruvu (ASP Research Review) Adjoint <strong>Sensitivity</strong> <strong>Analysis</strong> 06/14/2007 22 / 30
Nonreflective Outflow Boundary<br />
Adjoint boundary conditions:<br />
ψ(0, t) = 0 and<br />
φ(L, t) + 2 u′ (L, t)<br />
ψ(L, t) = 0<br />
η(L, t)<br />
Sensitivities <strong>for</strong> the subcritical flow then are:<br />
δJ<br />
δu ′ (0, t p )<br />
δJ<br />
δη(L, t p )<br />
= φ(0, t p )<br />
[ (u ′ )<br />
(L, t p ) 2<br />
=<br />
− gη(L, t p )]<br />
ψ(L, t p )<br />
η(L, t p )<br />
Vani Cheruvu (ASP Research Review) Adjoint <strong>Sensitivity</strong> <strong>Analysis</strong> 06/14/2007 23 / 30
Cubed-Sphere Geometry<br />
Sadourny (1972)<br />
HOMME High-Order Method Modeling Environment<br />
Z<br />
P 1<br />
Y<br />
P5<br />
(Top)<br />
X<br />
P4<br />
P1 P2 P3<br />
Figure: Cube inscribed in a<br />
Sphere<br />
P6<br />
(Bottom)<br />
Figure: Faces of the Cube<br />
Vani Cheruvu (ASP Research Review) Adjoint <strong>Sensitivity</strong> <strong>Analysis</strong> 06/14/2007 24 / 30
Neighboring elements that share same cube edge<br />
Face ! A<br />
Face ! B<br />
y<br />
x<br />
x<br />
y<br />
Figure: (a)<br />
Figure: (b)<br />
To compute the flux on the edge of the cubed-sphere, both the left and<br />
right components F − n and F + n are required.<br />
A l<br />
"<br />
F +<br />
l<br />
G +<br />
l<br />
#<br />
»<br />
F<br />
+<br />
= s<br />
G s<br />
+<br />
– »<br />
, A −1 F<br />
+<br />
s<br />
n G s<br />
+<br />
– »<br />
F<br />
+<br />
= n<br />
G n<br />
+<br />
–<br />
⇒<br />
A −1<br />
n A l<br />
"<br />
F +<br />
l<br />
G +<br />
l<br />
#<br />
.<br />
Vani Cheruvu (ASP Research Review) Adjoint <strong>Sensitivity</strong> <strong>Analysis</strong> 06/14/2007 25 / 30
Spectral Finite Volume Results<br />
0.8<br />
0.7<br />
0.6<br />
0.5<br />
0.4<br />
0.3<br />
0.2<br />
0.1<br />
0<br />
!1 !0.8 !0.6 !0.4 !0.2 0 0.2 0.4 0.6 0.8 1<br />
Vani Cheruvu (ASP Research Review) Adjoint <strong>Sensitivity</strong> <strong>Analysis</strong> 06/14/2007 26 / 30
Gibbs Phenomenon<br />
(sin(x) + 1 3 sin(3x) + 1 5 sin(5x) + . . . )<br />
f (x) = 4 π<br />
1.5<br />
S1<br />
1<br />
S3<br />
0.5<br />
0<br />
!0.5<br />
S2<br />
!1<br />
!1.5<br />
!4 !3 !2 !1 0 1 2 3 4<br />
Vani Cheruvu (ASP Research Review) Adjoint <strong>Sensitivity</strong> <strong>Analysis</strong> 06/14/2007 27 / 30
DG-1D : Numerical Examples<br />
U t + F x (U) = 0, Ω ≡ [−1, 1], periodic domain. For linear advection (square wave) F (U) = U without limiter (left panel) and<br />
with limiter (right panel)<br />
1<br />
1<br />
0.5<br />
0.5<br />
0<br />
0<br />
!1 !0.5 0 0.5 1<br />
!1 !0.5 0 0.5 1<br />
Vani Cheruvu (ASP Research Review) Adjoint <strong>Sensitivity</strong> <strong>Analysis</strong> 06/14/2007 28 / 30
DG-1D : Numerical Examples<br />
U t + F x (U) = 0, Ω ≡ [−1, 1], periodic domain. For linear advection (sine wave) F (U) = U without limiter (left panel) and<br />
with limiter (right panel)<br />
1<br />
1<br />
0.8<br />
0.8<br />
0.6<br />
0.6<br />
0.4<br />
0.4<br />
0.2<br />
0.2<br />
0<br />
0<br />
!0.2<br />
!0.2<br />
!0.4<br />
!0.4<br />
!0.6<br />
!0.6<br />
!0.8<br />
!0.8<br />
!1<br />
!1 !0.8 !0.6 !0.4 !0.2 0 0.2 0.4 0.6 0.8 1<br />
!1<br />
!1 !0.8 !0.6 !0.4 !0.2 0 0.2 0.4 0.6 0.8 1<br />
Figure: Without Limiter<br />
Figure: With Limiter<br />
Vani Cheruvu (ASP Research Review) Adjoint <strong>Sensitivity</strong> <strong>Analysis</strong> 06/14/2007 29 / 30
Summary<br />
Adjoint sensitivity method <strong>for</strong> SWE using high-order spectral element<br />
method<br />
◮ Fast and accurate tool <strong>for</strong> computing the sensitivities<br />
◮ Efficient in determining how a specific condition is dependent on many<br />
different parameters which influence it<br />
◮ Adjoint sensitivity method is more efficient than Direct sensitivity<br />
method<br />
Wavelets can be used <strong>for</strong> limiting the discontinuous Galerkin solutions<br />
“Numerical <strong>Analysis</strong> is service to science and engineering” -<br />
Gene Golub<br />
Acknowledgements: Maura Hagan, Al Kellie, Joe Tribbia<br />
Vani Cheruvu (ASP Research Review) Adjoint <strong>Sensitivity</strong> <strong>Analysis</strong> 06/14/2007 30 / 30