Optimal Shape Design of Wave MakerIntroduction Robin Goix 1 Introduction 2 Modelling of wave...
Transcript of Optimal Shape Design of Wave MakerIntroduction Robin Goix 1 Introduction 2 Modelling of wave...
Optimal Shape Design of Wave Maker
Robin Goix
June 23, 2014
Introduction Robin Goix
WavegardenSource: Youtube.(2014, Jan. 30) WAVE GARDEN avec SURF REPORT
Introduction Robin Goix
Wavegarden’s Technology: an underwater object pulled in pool
Wavegarden - Wake of the objectSource: c©2014 Instant Sport S.L.
Underwater moving Object The entire apparatus
Introduction Robin Goix
Wavemaker project main goals:
Develop the computational framework to run simulations of wavesgenerated by an underwater moving object
Optimize the shape of that moving object to obtain the best waveaccording to criteria (high, shape, energy...)
Introduction Robin Goix
1 Introduction
2 Modelling of wave generationBoussinesq system derived by PeregrineWeak formulation and time discretizationAssessment of the code: the solitary wave solutionResults
3 Optimization of the shapeControl of the optimization problemCoupling with a transport equationFirst resultsImprovement possibilities and other solutions
Modelling of wave generation Robin Goix
Boussinesq system derived by Peregrine
Seabed profil:h(x, y, t) = D(x, y) + εζ(x, y, t)
Free surface: η(x, y, t)
Fluid horizontal velocity: u(x, y, z, t)
Scaling: ε =a0
h0, σ =
h0
λ0
ut +∇η + ε(u · ∇)u− σ2 h
2∇(∇ · (hut)) + σ2 h
2
6∇(∇ · ut)− σ2 h
2∇ζtt = 0
ηt + ζt +∇ · [(h+ εη)u] = 0
Modelling of wave generation Robin Goix
Boussinesq system derived by Peregrine
Seabed profil:h(x, y, t) = D(x, y) + εζ(x, y, t)
Free surface: η(x, y, t)
Fluid horizontal velocity: u(x, y, z, t)
Scaling: ε =a0
h0, σ =
h0
λ0
ut +∇η + ε(u · ∇)u− σ2 h
2∇(∇ · (hut)) + σ2 h
2
6∇(∇ · ut)− σ2 h
2∇ζtt = 0
ηt + ζt +∇ · [(h+ εη)u] = 0
Modelling of wave generation Robin Goix
Weak formulation and time discretization
ut +∇η + ε(u · ∇)u− σ2 h
2∇(∇ · (hut)) + σ2 h
2
6∇(∇ · ut)− σ2 h
2∇ζtt = 0
ηt + ζt +∇ · [(h+ εη)u] = 0u · n = 0 on ∂Ω
Let (v, ξ) ∈ V×H1(Ω), with V = w ∈ H1(Ω)2,w · n = 0 on ∂Ω,integrating and using divergence theorem, we get:
0 =
∫Ωut · v dx−
∫Ωη (∇ · v) dx + ε
∫Ω(u · ∇)u · v dx +
σ2
2
∫Ω(∇ · (hut)) (∇ · (hv)) dx
−σ2
6
∫Ω(∇ · ut) (∇ · (h2v)) dx +
σ2
2
∫Ωζtt (∇ · (hv)) dx
0 =
∫Ωηt ξ dx +
∫Ωζt ξ dx−
∫Ω∇ξ · [(h+ εη)u] dx
Modelling of wave generation Robin Goix
Weak formulation and time discretization
ut +∇η + ε(u · ∇)u− σ2 h
2∇(∇ · (hut)) + σ2 h
2
6∇(∇ · ut)− σ2 h
2∇ζtt = 0
ηt + ζt +∇ · [(h+ εη)u] = 0u · n = 0 on ∂Ω
Let (v, ξ) ∈ V×H1(Ω), with V = w ∈ H1(Ω)2,w · n = 0 on ∂Ω,integrating and using divergence theorem, we get:
0 =
∫Ωut · v dx−
∫Ωη (∇ · v) dx + ε
∫Ω(u · ∇)u · v dx +
σ2
2
∫Ω(∇ · (hut)) (∇ · (hv)) dx
−σ2
6
∫Ω(∇ · ut) (∇ · (h2v)) dx +
σ2
2
∫Ωζtt (∇ · (hv)) dx
0 =
∫Ωηt ξ dx +
∫Ωζt ξ dx−
∫Ω∇ξ · [(h+ εη)u] dx
Modelling of wave generation Robin Goix
Weak formulation and time discretization
ut +∇η + ε(u · ∇)u− σ2 h
2∇(∇ · (hut)) + σ2 h
2
6∇(∇ · ut)− σ2 h
2∇ζtt = 0
ηt + ζt +∇ · [(h+ εη)u] = 0u · n = 0 on ∂Ω
Let (v, ξ) ∈ V×H1(Ω), with V = w ∈ H1(Ω)2,w · n = 0 on ∂Ω,integrating and using divergence theorem, we get:
0 =
∫Ωut · v dx−
∫Ωη (∇ · v) dx + ε
∫Ω(u · ∇)u · v dx +
σ2
2
∫Ω(∇ · (hut)) (∇ · (hv)) dx
−σ2
6
∫Ω(∇ · ut) (∇ · (h2v)) dx +
σ2
2
∫Ωζtt (∇ · (hv)) dx
0 =
∫Ωηt ξ dx +
∫Ωζt ξ dx−
∫Ω∇ξ · [(h+ εη)u] dx
Modelling of wave generation Robin Goix
With the following schemes:
fn+θ = (1− θ)fn + θfn+1
fn+1t =
fn+1 − fn
dt, fn+1
tt =fn+1 − 2fn + fn−1
dt2
for each time-step, knowing un and ηn, our problem is therefore:Find (un+1, ηn+1) ∈ V×H1(Ω),∀(v, ξ) ∈ V×H1(Ω):
0 =
∫Ωun+1t · v dx−
∫Ωηn+θ (∇ · v) dx + ε
∫Ω(un+θ · ∇)un+θ · v dx
+σ2
2
∫Ω(∇ · (hn+θun+1
t )) (∇ · (hn+θv)) dx−σ2
6
∫Ω(∇ · un+1
t ) (∇ · ((hn+θ)2v)) dx
+σ2
2
∫Ωζn+1tt (∇ · (hn+θv)) dx
0 =
∫Ωηn+1t ξ dx +
∫Ωζn+1t ξ dx−
∫Ω∇ξ · [(hn+θ + εηn+θ)un+θ] dx
+ stabilization
Modelling of wave generation Robin Goix
Assessment of the code: the solitary wave problem
Modelling of wave generation Robin Goix
Assessment of the code: the solitary wave problem
Fixed simulation time and mesh
Different time-stepping methods
With or without added stabilization
Error L2 and L∞
Modelling of wave generation Robin Goix
Results: some simulations
Optimization of the shape Robin Goix
Control of the optimization problem
ut +∇η + ε(u · ∇)u− σ2 h
2∇(∇ · (hut)) + σ2 h
2
6∇(∇ · ut)− σ2 h
2∇ζtt = 0
ηt + ζt +∇ · [(h+ εη)u] = 0h(x, y, t) = D(x, y) + εζ(x, y, t)
A time-dependent parameter
But Dolfin-Adjoint does not make it easy to optimize such aparameter:
m = InitialConditionParameter(u)
m = ScalarParameter(s)
However it is completely determined by an initial condition and atrajectory:
Optimization of the shape Robin Goix
Control of the optimization problem
ut +∇η + ε(u · ∇)u− σ2 h
2∇(∇ · (hut)) + σ2 h
2
6∇(∇ · ut)− σ2 h
2∇ζtt = 0
ηt + ζt +∇ · [(h+ εη)u] = 0h(x, y, t) = D(x, y) + εζ(x, y, t)
A time-dependent parameter
But Dolfin-Adjoint does not make it easy to optimize such aparameter:
m = InitialConditionParameter(u)
m = ScalarParameter(s)
However it is completely determined by an initial condition and atrajectory:
Optimization of the shape Robin Goix
Control of the optimization problem
ut +∇η + ε(u · ∇)u− σ2 h
2∇(∇ · (hut)) + σ2 h
2
6∇(∇ · ut)− σ2 h
2∇ζtt = 0
ηt + ζt +∇ · [(h+ εη)u] = 0h(x, y, t) = D(x, y) + εζ(x, y, t)
A time-dependent parameter
But Dolfin-Adjoint does not make it easy to optimize such aparameter:
m = InitialConditionParameter(u)
m = ScalarParameter(s)
However it is completely determined by an initial condition and atrajectory:
Optimization of the shape Robin Goix
Control of the optimization problem
ut +∇η + ε(u · ∇)u− σ2 h
2∇(∇ · (hut)) + σ2 h
2
6∇(∇ · ut)− σ2 h
2∇ζtt = 0
ηt + ζt +∇ · [(h+ εη)u] = 0h(x, y, t) = D(x, y) + εζ(x, y, t)
A time-dependent parameter
But Dolfin-Adjoint does not make it easy to optimize such aparameter:
m = InitialConditionParameter(u)
m = ScalarParameter(s)
However it is completely determined by an initial condition and atrajectory:
Optimization of the shape Robin Goix
Control of the optimization problem
ut +∇η + ε(u · ∇)u− σ2 h
2∇(∇ · (hut)) + σ2 h
2
6∇(∇ · ut)− σ2 h
2∇ζtt = 0
ηt + ζt +∇ · [(h+ εη)u] = 0h(x, y, t) = D(x, y) + εζ(x, y, t)
A time-dependent parameter
But Dolfin-Adjoint does not make it easy to optimize such aparameter:
m = InitialConditionParameter(u)
m = ScalarParameter(s)
However it is completely determined by an initial condition and atrajectory:
Optimization of the shape Robin Goix
Control of the optimization problem
ut +∇η + ε(u · ∇)u− σ2 h
2∇(∇ · (hut)) + σ2 h
2
6∇(∇ · ut)− σ2 h
2∇ζtt = 0
ηt + ζt +∇ · [(h+ εη)u] = 0h(x, y, t) = D(x, y) + εζ(x, y, t)
A time-dependent parameter
But Dolfin-Adjoint does not make it easy to optimize such aparameter:
m = InitialConditionParameter(u)
m = ScalarParameter(s)
However it is completely determined by an initial condition and atrajectory:ζ(x, y, t) = ζ0(x− Ux(t), y − Uy(t))
Optimization of the shape Robin Goix
Control of the optimization problem
ut +∇η + ε(u · ∇)u− σ2 h
2∇(∇ · (hut)) + σ2 h
2
6∇(∇ · ut)− σ2 h
2∇ζtt = 0
ηt + ζt +∇ · [(h+ εη)u] = 0h(x, y, t) = D(x, y) + εζ(x, y, t)
A time-dependent parameter
But Dolfin-Adjoint does not make it easy to optimize such aparameter:
m = InitialConditionParameter(u)
m = ScalarParameter(s)
However it is completely determined by an initial condition and atrajectory:
ζt + εU · ∇ζ = 0
ζ(x, y, 0) = ζ0(x, y)
Optimization of the shape Robin Goix
Coupling with a transport equation
ut +∇η + ε(u · ∇)u− σ2 h
2∇(∇ · (hut)) + σ2 h
2
6∇(∇ · ut)− σ2 h
2∇ζtt = 0
ηt + ζt +∇ · [(h+ εη)u] = 0h(x, y, t) = D(x, y) + εζ(x, y, t)ζt + εU · ∇ζ = 0ζ(x, y, 0) = ζ0(x, y)
We solve each system separately, step by step.
However, hyperbolic equations are difficult to be solved using FE.The following weak formulation:
∀τ ∈ H10(Ω),
∫Ω
τζn+1 − ζn
∆t− ε∫
Ω
∇τ ·Un+θζn+θ = 0
is not enough to transport the object with a good accuracy.
Optimization of the shape Robin Goix
Coupling with a transport equation
ut +∇η + ε(u · ∇)u− σ2 h
2∇(∇ · (hut)) + σ2 h
2
6∇(∇ · ut)− σ2 h
2∇ζtt = 0
ηt + ζt +∇ · [(h+ εη)u] = 0h(x, y, t) = D(x, y) + εζ(x, y, t)ζt + εU · ∇ζ = 0ζ(x, y, 0) = ζ0(x, y)
We solve each system separately, step by step.
However, hyperbolic equations are difficult to be solved using FE.The following weak formulation:
∀τ ∈ H10(Ω),
∫Ω
τζn+1 − ζn
∆t− ε∫
Ω
∇τ ·Un+θζn+θ = 0
is not enough to transport the object with a good accuracy.
Optimization of the shape Robin Goix
Coupling with a transport equation
ut +∇η + ε(u · ∇)u− σ2 h
2∇(∇ · (hut)) + σ2 h
2
6∇(∇ · ut)− σ2 h
2∇ζtt = 0
ηt + ζt +∇ · [(h+ εη)u] = 0h(x, y, t) = D(x, y) + εζ(x, y, t)ζt + εU · ∇ζ = 0ζ(x, y, 0) = ζ0(x, y)
We solve each system separately, step by step.
However, hyperbolic equations are difficult to be solved using FE.The following weak formulation:
∀τ ∈ H10(Ω),
∫Ω
τζn+1 − ζn
∆t− ε∫
Ω
∇τ ·Un+θζn+θ = 0
is not enough to transport the object with a good accuracy.
Optimization of the shape Robin Goix
Coupling with a transport equation
ut +∇η + ε(u · ∇)u− σ2 h
2∇(∇ · (hut)) + σ2 h
2
6∇(∇ · ut)− σ2 h
2∇ζtt = 0
ηt + ζt +∇ · [(h+ εη)u] = 0h(x, y, t) = D(x, y) + εζ(x, y, t)ζt + εU · ∇ζ = 0ζ(x, y, 0) = ζ0(x, y)
We solve each system separately, step by step.
However, hyperbolic equations are difficult to be solved using FE.The following weak formulation:
∀τ ∈ H10(Ω),
∫Ω
τζn+1 − ζn
∆t− ε∫
Ω
∇τ ·Un+θζn+θ = 0
is not enough to transport the object with a good accuracy.
Optimization of the shape Robin Goix
Coupling with a transport equation : a second-order Taylor-Galerkinapproximation
ζn+1 − ζn
∆t=
(∂ζ
∂t
)n+
∆t
2
(∂2ζ
∂t2
)n+θ
+O(∆t2)
∂ζ
∂t= −εU · ∇ζ,
∂2ζ
∂t2= −ε
∂U
∂t· ∇ζ − εU · ∇
∂ζ
∂t
ζn+1 − ζn
∆t= −εUn · ∇ζn −
ε∆t
2
(∂U
∂t· ∇ζn+θ − εUn+θ · ∇(Un+θ · ∇ζn+θ)
)+O(∆t2)
which lead us to the following weak formulation: ∀τ ∈ H10(Ω),∫
Ωτζn+1 − ζn
∆t− ε∫
Ω∇τ ·Unζn −
ε∆t
2
∫Ω∇τ ·
∂U
∂tζn+θ
+ε2∆t
2
∫Ω
(∇τ ·U)(∇ζn+θ ·Un+θ) = 0
Optimization of the shape Robin Goix
Coupling with a transport equation
Optimization of the shape Robin Goix
First resultsFunctional to minimize:
infζ0∈Uad
J(ζ0) =
∫ω
j(η(T )) dx
,withj(η(T )) = −1
2(|η(T )|2+|∇η(T )|2)
Domain ω on which the wave isassessed
Shape of the object - Uad = ζ0 ∈H1, ζmin ≤ ζ0 ≤ ζmax
Two optimization algorithms used:• Coupling between FEniCS and IPOPT • Dolfin-Adjoint and scipy
Optimization of the shape Robin Goix
First results
Convergence of IPOPT
Optimization of the shape Robin Goix
First results
Convergence of IPOPT
Optimization of the shape Robin Goix
First results
Convergence of IPOPT
Optimization of the shape Robin Goix
First results
Convergence of IPOPT
A bigger and steeperwave;
however moreperturbations in thewake ...
... which means a lossof energy.
Optimization of the shape Robin Goix
Improvement possibilities and other solutions
Find a better functional which consider the energy aspect throughthe pressure on the object
Improve the optimization:
More accurate way to solve the transport equation(Discontinuous-Galerkin?)Code our own gradient algorithm to get rid of dolfin-adjointconstraints
Develop an other model which could simulate wave breaking
Wavegarden - Breaking waveSource: c©2014 Instant Sport S.L.