Reduced basis method for parametrized optimal control problems … · 2017. 5. 10. · Gianluigi...
Transcript of Reduced basis method for parametrized optimal control problems … · 2017. 5. 10. · Gianluigi...
Introduction Saddle-point formulation RB method Results
Reduced basis method for
parametrized optimal control problems
governed by PDEs
Gianluigi Rozzajoint work with Federico Negri, Alfio Quarteroni, Andrea Manzoni
MATHICSE - CMCS Modelling and Scientific Computing
Ecole Polytechnique Federale de Lausanne
J.-L. Lions and E. Magenes Memorial Days Workshop
University Pierre et Marie Curie (UPMC Paris VI) Laboratoire Jacques-Louis Lions
Paris, France, 14-15 December 2011.
Acknowledgements: A.T. Patera (MIT)
Sponsors: European Research Council - Mathcard Project, Progetto Roberto Rocca Politecnico di Milano-MIT, Swiss National Science Foundation
Introduction Saddle-point formulation RB method Results
Outline
1. Introduction
Motivation and ingredients
2. Parametrized linear-quadratic Optimal Control Problems
Abstract formulation
Saddle-point formulation
3. Reduced Basis method for Optimal Control Problems
Reduced Basis (RB) methodology
Approximation, stability, a posteriori error bounds
4. Preliminary numerical results
Optimal control of parametrized advection-diffusion problems
Optimal control of parametrized Stokes problems
An (idealized) application for blood flows
Introduction Saddle-point formulation RB method Results
Reduction strategies for Parametrized Optimal Control Problems
The characterization of a system in
terms of physical quantities and/or
geometrical configuration usually depends
on a set of parameters
the system response will be
parameter dependent as well,
and so will be the optimal control
Parametrized optimal control problems: the prediction of optimal control inputs and
the optimization of given output of interests is required for each different
value of the parameters.
The computational effort may be unacceptably high and, often, unaffordable when
performing the optimization process for many different parameter values (many-query
context)
for a given new configuration, we want to compute the solution in a rapid way
(real-time context)
Goal: to achieve the accuracy and reliability of a high fidelity approximationbut at greatly reduced cost of a low order model
Introduction Saddle-point formulation RB method Results
Reduction strategies for Parametrized Optimal Control Problems
The characterization of a system in
terms of physical quantities and/or
geometrical configuration usually depends
on a set of parameters
the system response will be
parameter dependent as well,
and so will be the optimal control
Parametrized optimal control problems: the prediction of optimal control inputs and
the optimization of given output of interests is required for each different
value of the parameters.
The computational effort may be unacceptably high and, often, unaffordable when
performing the optimization process for many different parameter values (many-query
context)
for a given new configuration, we want to compute the solution in a rapid way
(real-time context)
Goal: to achieve the accuracy and reliability of a high fidelity approximationbut at greatly reduced cost of a low order model
Introduction Saddle-point formulation RB method Results
Previous works and improvements on parametrized optimal control
early works by Ito and Ravindran [1998 and 2001]: optimal control of Navier-Stokes
equations
the RB method has been applied to parametrized linear-quadratic advection
diffusion optimal control problems in different contexts:
[Quarteroni, R., Quaini, 07], [Tonn, Urban, Volkwein, 11], [Grepl, Karcher, 11]
considering low-dimensional control variable.
we aim at developing a reduced framework that enables to handle with general
control functions, i.e. infinite dimensional distributed and/or boundary control
functions.
an efficient and rigorous a posteriori error estimation, necessary both for
constructing the reduced order model and measuring its accuracy, is still partially
missing for a large class of optimal control problems. Previuous preliminary works by
[Dede, 2010] [Tonn, Urban, Volkwein, 11] [Grepl, Karcher, 11]
Introduction Saddle-point formulation RB method Results
Optimal control problems [Lions, 1971]
In general, an optimal control problem (OCP) consists of:
a control function u, which can be
seen as an input for the system,
a controlled system, i.e. an
input-output process: E(y , u) = 0,
beign y the state variable
an objective functional to be
minimized: J (y , u)
STATE
PROBLEM
STATE
PROBLEM
µ
Output J (y, u)
Output
yd (µ) (data)
J (y, u; µ)
y(u)
Optimization:
update control u
u
Introduction Saddle-point formulation RB method Results
Optimal control problems [Lions, 1971]
In general, an optimal control problem (OCP) consists of:
a control function u, which can be
seen as an input for the system,
a controlled system, i.e. an
input-output process: E(y , u) = 0,
beign y the state variable
an objective functional to be
minimized: J (y , u)
STATE
PROBLEM
STATE
PROBLEM
µ
Output J (y, u)
Output
yd (µ) (data)
J (y, u; µ)
y(u)
Optimization:
update control u
u
find the optimal control u∗ and the state y(u∗) such that the cost functional
J (y , u) is minimized subject to E(y , u) = 0(OCP)
Introduction Saddle-point formulation RB method Results
Optimal control problems [Lions, 1971]
In general, an optimal control problem (OCP) consists of:
a control function u, which can be
seen as an input for the system,
a controlled system, i.e. an
input-output process: E(y , u) = 0,
beign y the state variable
an objective functional to be
minimized: J (y , u)
STATE
PROBLEM
STATE
PROBLEM
µ
Output J (y, u)
Output
yd (µ) (data)
J (y, u; µ)
y(u)
Optimization:
update control u
u
find the optimal control u∗ and the state y(u∗) such that the cost functional
J (y , u) is minimized subject to E(y , u) = 0(OCP)
We restrict attention to:
quadratic cost functionals, e.g. J(y , u) = 12‖y − yd‖2 + α
2‖u‖2
stationary, linear, elliptic state equations: in particular (scalar coercive)
advection-diffusion equations and then (vectorial noncoercive) Stokes equations
Introduction Saddle-point formulation RB method Results
Parametrized optimal control problems
A parametrized optimal control problem (OCPµ) consists of:
a control function u(µ), which can
be seen as an input for the system,
a controlled system, i.e. an
input-output process:
E(y(µ), u(µ);µ) = 0,
an objective functional to be
minimized: J (y(µ), u(µ);µ)
STATE
PROBLEM
STATE
PROBLEM
µ
Output J (y, u)
Output
yd (µ) (data)
J (y, u; µ)y(u(µ))
Optimization:
update control u
u(µ)
given µ ∈ D, find the optimal control u∗(µ) and the state y∗(µ) such that the
cost functional J (y(µ), u(µ);µ) is minimized subject to E(y(µ), u(µ);µ) = 0(OCPµ)
where µ ∈ D ⊂ Rp denotes a p-vector whose components can represent:
coefficients in boundary conditions
geometrical configurations
physical parametrization
data (observation)
Introduction Saddle-point formulation RB method Results
Saddle-point formulation
for optimal control problems
Introduction Saddle-point formulation RB method Results
The abstract optimization problem
Functional setting: Y state space, U control space, Q(≡ Y ) adjoint space,
Z observation (data) space s.t. Y ⊂ Z,
In the following: y , z ∈ Y u, v ∈ U, p, q ∈ Q.
Introduction Saddle-point formulation RB method Results
The abstract optimization problem: quadratic cost functional
Functional setting: Y state space, U control space, Q(≡ Y ) adjoint space,
Z observation (data) space s.t. Y ⊂ Z,
In the following: y , z ∈ Y u, v ∈ U, p, q ∈ Q.
we consider the following quadratic cost functional to be minimized
J(y , u;µ) =1
2m(y − yd (µ), y − yd (µ);µ) +
α
2n(u, u;µ)
where yd (µ) ∈ Z is a given function, α > 0 regularization parameter.
We make the following assumptions: for any µ ∈ D
m(·, ·;µ) : Z × Z → R is a symmetric, continuous and positive bilinear form
n(·, ·;µ) : U × U → R is a symmetric, continuous and coercive bilinear form
Introduction Saddle-point formulation RB method Results
The abstract optimization problem: linear constraint
Functional setting: Y state space, U control space, Q(≡ Y ) adjoint space,
Z observation (data) space s.t. Y ⊂ Z,
In the following: y , z ∈ Y u, v ∈ U, p, q ∈ Q.
the linear constraint is expressed by the following state equation (elliptic PDE in
weak form)
a(y , q;µ) = c(u, q;µ) + 〈G(µ), q〉 ∀q ∈ Q
where G(µ) ∈ Q ′ for all µ ∈ D.
We make the following assumptions: for any µ ∈ D
a(·, ·;µ) : Y × Q → R is a continuous and (weakly) coercive bilinear form
c(·, ·;µ) : U × Q → R is a continuous bilinear form
Holding these assumptions, given u ∈ U, the state equation is uniquely solvable
for any µ ∈ D.
Introduction Saddle-point formulation RB method Results
The abstract optimization problem
Parametrized optimal control problem
minimize J(y , u;µ) =1
2m(y − yd (µ), y − yd (µ);µ) +
α
2n(u, u;µ)
s.t. a(y , q;µ) = c(u, q;µ) + 〈G(µ), q〉 ∀q ∈ Q.
In order to develop the Reduced Basis (RB) method
we firstly recast the problem in the framework of saddle-point problem
[Gunzburger, Bochev, 2004]
we then apply the well-known Brezzi theory [Brezzi, Fortin, 1991]
→ existence, uniqueness, stability and optimality conditions
This way we can exploit the analogies with the already developed theory of RB
method for Stokes-type problems [R., Veroy, 2007; R., Huynh, Manzoni 2010]
Introduction Saddle-point formulation RB method Results
The abstract optimization problem
Parametrized optimal control problem
minimize J(y , u;µ) =1
2m(y − yd (µ), y − yd (µ);µ) +
α
2n(u, u;µ)
s.t. a(y , q;µ) = c(u, q;µ) + 〈G(µ), q〉 ∀q ∈ Q.
In order to develop the Reduced Basis (RB) method
we firstly recast the problem in the framework of saddle-point problem
[Gunzburger, Bochev, 2004]
we then apply the well-known Brezzi theory [Brezzi, Fortin, 1991]
→ existence, uniqueness, stability and optimality conditions
This way we can exploit the analogies with the already developed theory of RB
method for Stokes-type problems [R., Veroy, 2007; R., Huynh, Manzoni 2010]
Introduction Saddle-point formulation RB method Results
Saddle-point formulation
Let X ≡ Y × U the state and control space
Q the adjoint (Lagrange multiplier) space as before
and let x = (y , u) ∈ X , w = (z , v) ∈ X , with ‖x‖2X = ‖y‖2
Y + ‖u‖2U
Let A(x ,w ;µ) = m(y , z ;µ) + αn(u, v ;µ)
〈F (µ),w〉 = m(yd (µ), z ;µ)
t(µ) =1
2m(yd (µ), yd (µ);µ)
the cost functional can be expressed as
J(x ;µ) = 12A(x , x ;µ) − 〈F (µ), x〉 + t(µ) = J (x ;µ) + t(µ)
Let B(w , q;µ) = a(z , q;µ)− c(v , q;µ),
the constraint equation can be expressed as
B(x , q;µ) = 〈G(µ), q〉 ∀q ∈ Q
Introduction Saddle-point formulation RB method Results
Saddle-point formulation
Let X ≡ Y × U the state and control space
Q the adjoint (Lagrange multiplier) space as before
and let x = (y , u) ∈ X , w = (z , v) ∈ X , with ‖x‖2X = ‖y‖2
Y + ‖u‖2U
Let A(x ,w ;µ) = m(y , z ;µ) + αn(u, v ;µ)
〈F (µ),w〉 = m(yd (µ), z ;µ)
t(µ) =1
2m(yd (µ), yd (µ);µ)
the cost functional can be expressed as
J(x ;µ) = 12A(x , x ;µ) − 〈F (µ), x〉 + t(µ) = J (x ;µ) + t(µ)
Let B(w , q;µ) = a(z , q;µ)− c(v , q;µ),
the constraint equation can be expressed as
B(x , q;µ) = 〈G(µ), q〉 ∀q ∈ Q
Introduction Saddle-point formulation RB method Results
Saddle-point formulation
Let X ≡ Y × U the state and control space
Q the adjoint (Lagrange multiplier) space as before
and let x = (y , u) ∈ X , w = (z , v) ∈ X , with ‖x‖2X = ‖y‖2
Y + ‖u‖2U
Let A(x ,w ;µ) = m(y , z ;µ) + αn(u, v ;µ)
〈F (µ),w〉 = m(yd (µ), z ;µ)
t(µ) =1
2m(yd (µ), yd (µ);µ)
the cost functional can be expressed as
J(x ;µ) = 12A(x , x ;µ) − 〈F (µ), x〉 + t(µ) = J (x ;µ) + t(µ)
Let B(w , q;µ) = a(z , q;µ)− c(v , q;µ),
the constraint equation can be expressed as
B(x , q;µ) = 〈G(µ), q〉 ∀q ∈ Q
Introduction Saddle-point formulation RB method Results
Saddle-point formulation: applying Brezzi theory
The optimal control problem can be recast in the form: given µ ∈ D,
min J (x ;µ) =1
2A(x , x ;µ)− 〈F (µ), x〉, subject to
B(x , q;µ) = 〈G(µ), q〉 ∀q ∈ Q.
Let X0 = w ∈ X : B(w , q;µ) = 0 ∀q ∈ Q ⊂ X , the assumptions we have made
imply that (e.g. [Gunzburger, Bochev, 2004]):
the bilinear form A(·, ·;µ) is continuous over X × X and coercive over X0
the bilinear form B(·, ·;µ) is continuous over X × Q and satisfies the following
inf-sup condition
∃β0 > 0 : β(µ) = infq∈Q
supw∈X
B(w , q;µ)
‖w‖X‖q‖Q≥ β0, ∀µ ∈ D
the bilinear form A(·, ·;µ) is symmetric and non-negative over X
Introduction Saddle-point formulation RB method Results
Saddle-point formulation: applying Brezzi theory
The optimal control problem can be recast in the form: given µ ∈ D,
min J (x ;µ) =1
2A(x , x ;µ)− 〈F (µ), x〉, subject to
B(x , q;µ) = 〈G(µ), q〉 ∀q ∈ Q.
Let X0 = w ∈ X : B(w , q;µ) = 0 ∀q ∈ Q ⊂ X , the assumptions we have made
imply that (e.g. [Gunzburger, Bochev, 2004]):
the bilinear form A(·, ·;µ) is continuous over X × X and coercive over X0
the bilinear form B(·, ·;µ) is continuous over X × Q and satisfies the following
inf-sup condition
∃β0 > 0 : β(µ) = infq∈Q
supw∈X
B(w , q;µ)
‖w‖X‖q‖Q≥ β0, ∀µ ∈ D
the bilinear form A(·, ·;µ) is symmetric and non-negative over X
Introduction Saddle-point formulation RB method Results
Saddle-point formulation: applying Brezzi theory
the optimal control problem
minx∈X
J (x ;µ) subject to B(x , q;µ) = 〈G(µ), q〉 ∀q ∈ Q.
has a unique solution x = (y , u) ∈ X for any µ ∈ D
that solution can be determined by solving the optimality systemA(x(µ),w ;µ) + B(w , p(µ);µ) = 〈F (µ),w〉 ∀w ∈ X ,
B(x(µ), q;µ) = 〈G(µ), q〉 ∀q ∈ Q,
Compact form
given µ ∈ D, find U(µ) ∈ X s.t:
B(U(µ),W;µ) = F(W;µ) ∀W ∈ X .
X = X × Q, U = (x , p), W = (w , q)
B(U,W;µ) = A(x ,w ;µ) +B(w , p;µ) +B(x , q;µ)
F(W;µ) = 〈F (µ),w〉+ 〈G(µ), q〉
at this point we may apply the Galerkin-FE approximation
Introduction Saddle-point formulation RB method Results
Saddle-point formulation: applying Brezzi theory
the optimal control problem
minx∈X
J (x ;µ) subject to B(x , q;µ) = 〈G(µ), q〉 ∀q ∈ Q.
has a unique solution x = (y , u) ∈ X for any µ ∈ D
that solution can be determined by solving the optimality systemA(x(µ),w ;µ) + B(w , p(µ);µ) = 〈F (µ),w〉 ∀w ∈ X ,
B(x(µ), q;µ) = 〈G(µ), q〉 ∀q ∈ Q,
Compact form
given µ ∈ D, find U(µ) ∈ X s.t:
B(U(µ),W;µ) = F(W;µ) ∀W ∈ X .
X = X × Q, U = (x , p), W = (w , q)
B(U,W;µ) = A(x ,w ;µ) +B(w , p;µ) +B(x , q;µ)
F(W;µ) = 〈F (µ),w〉+ 〈G(µ), q〉
at this point we may apply the Galerkin-FE approximation
Introduction Saddle-point formulation RB method Results
Saddle-point formulation: applying Brezzi theory
the optimal control problem
minx∈X
J (x ;µ) subject to B(x , q;µ) = 〈G(µ), q〉 ∀q ∈ Q.
has a unique solution x = (y , u) ∈ X for any µ ∈ D
that solution can be determined by solving the optimality systemA(x(µ),w ;µ) + B(w , p(µ);µ) = 〈F (µ),w〉 ∀w ∈ X ,
B(x(µ), q;µ) = 〈G(µ), q〉 ∀q ∈ Q,
Compact form
given µ ∈ D, find U(µ) ∈ X s.t:
B(U(µ),W;µ) = F(W;µ) ∀W ∈ X .
X = X × Q, U = (x , p), W = (w , q)
B(U,W;µ) = A(x ,w ;µ) +B(w , p;µ) +B(x , q;µ)
F(W;µ) = 〈F (µ),w〉+ 〈G(µ), q〉
at this point we may apply the Galerkin-FE approximation
Introduction Saddle-point formulation RB method Results
Reduced Basis methodfor parametrized optimal control problems
Introduction Saddle-point formulation RB method Results
RB method: Construction
Pb(µ; U(µ))
µ-OCP, optimality system
U(µ) ∈ X : B(U(µ),W;µ) = F(W) ∀W ∈ X
PbN (µ; UN (µ))
Truth approximation (FEM)
UN (µ) ∈ XN : B(UN (µ),W;µ) = F(W) ∀W ∈ XN
Truth Hypothesis: UN (µ) “indistinguishable” from U(µ).
RB Motivation: µ→ UN (µ), JN (µ) too expensive and slow in
many-query and real-time contexts.
Sampling (Greedy)
Space Construction
(Hierarchical Lagrange basis)
OFFLINE
SN = µi , i = 1, . . . ,N
XN = spanUN (µi ), i = 1, . . . ,N
dim(XN ) = N N = dim(XN )
PbN (µ; UN (µ))
Galerkin projection
ONLINE
Reduced Basis (RB) approximation
UN (µ) ∈ XN : B(UN (µ),W;µ) = F(W) ∀W ∈ XN
[R., Huynh, Patera 2008, Talk by A. Manzoni]
Introduction Saddle-point formulation RB method Results
RB method: Construction
Pb(µ; U(µ))
µ-OCP, optimality system
U(µ) ∈ X : B(U(µ),W;µ) = F(W) ∀W ∈ X
PbN (µ; UN (µ))
Truth approximation (FEM)
UN (µ) ∈ XN : B(UN (µ),W;µ) = F(W) ∀W ∈ XN
Truth Hypothesis: UN (µ) “indistinguishable” from U(µ).
RB Motivation: µ→ UN (µ), JN (µ) too expensive and slow in
many-query and real-time contexts.
Sampling (Greedy)
Space Construction
(Hierarchical Lagrange basis)
OFFLINE
SN = µi , i = 1, . . . ,N
XN = spanUN (µi ), i = 1, . . . ,N
dim(XN ) = N N = dim(XN )
PbN (µ; UN (µ))
Galerkin projection
ONLINE
Reduced Basis (RB) approximation
UN (µ) ∈ XN : B(UN (µ),W;µ) = F(W) ∀W ∈ XN
[R., Huynh, Patera 2008, Talk by A. Manzoni]
Introduction Saddle-point formulation RB method Results
RB method: Construction
Pb(µ; U(µ))
µ-OCP, optimality system
U(µ) ∈ X : B(U(µ),W;µ) = F(W) ∀W ∈ X
PbN (µ; UN (µ))
Truth approximation (FEM)
UN (µ) ∈ XN : B(UN (µ),W;µ) = F(W) ∀W ∈ XN
Truth Hypothesis: UN (µ) “indistinguishable” from U(µ).
RB Motivation: µ→ UN (µ), JN (µ) too expensive and slow in
many-query and real-time contexts.
Sampling (Greedy)
Space Construction
(Hierarchical Lagrange basis)
OFFLINE
SN = µi , i = 1, . . . ,N
XN = spanUN (µi ), i = 1, . . . ,N
dim(XN ) = N N = dim(XN )
PbN (µ; UN (µ))
Galerkin projection
ONLINE
Reduced Basis (RB) approximation
UN (µ) ∈ XN : B(UN (µ),W;µ) = F(W) ∀W ∈ XN
[R., Huynh, Patera 2008, Talk by A. Manzoni]
Introduction Saddle-point formulation RB method Results
RB method: Construction
Pb(µ; U(µ))
µ-OCP, optimality system
U(µ) ∈ X : B(U(µ),W;µ) = F(W) ∀W ∈ X
PbN (µ; UN (µ))
Truth approximation (FEM)
UN (µ) ∈ XN : B(UN (µ),W;µ) = F(W) ∀W ∈ XN
Truth Hypothesis: UN (µ) “indistinguishable” from U(µ).
RB Motivation: µ→ UN (µ), JN (µ) too expensive and slow in
many-query and real-time contexts.
Sampling (Greedy)
Space Construction
(Hierarchical Lagrange basis)
OFFLINE
SN = µi , i = 1, . . . ,N
XN = spanUN (µi ), i = 1, . . . ,N
dim(XN ) = N N = dim(XN )
PbN (µ; UN (µ))
Galerkin projection
ONLINE
Reduced Basis (RB) approximation
UN (µ) ∈ XN : B(UN (µ),W;µ) = F(W) ∀W ∈ XN
[R., Huynh, Patera 2008, Talk by A. Manzoni]
Introduction Saddle-point formulation RB method Results
RB method: smooth parametric dependency
Guess: MN is low dimensional - smooth
dependence on µ (Lipschitz continuity)
(Adaptive) sampling procedure for parameter
exploration
reduced basis made of optimal solutions of the
FE “truth” problem
Evaluation procedure: (optimal) Galerkin
projection
[Noor, Peters, Porshing, Lee, ...]
MN
M
XN
M = U(µ) ∈ X : µ ∈ D
MN = UN (µ) ∈ XN : µ ∈ D
XN = spanUN (µi ), i = 1, . . . ,N
How to be rigorous, rapid and reliable?
depends on the sampling procedure for parameter exploration (greedy algorithm)
exploits an Offline/Online stratagem based on the affinity assumption:
B(U,W;µ) =
Qb∑q=1
Θqa (µ)Bq(U,W), F(W;µ) =
Qf∑q=1
Θqf (µ)Fq(W)
relies on a posteriori error analysis
Introduction Saddle-point formulation RB method Results
RB method: smooth parametric dependency
Guess: MN is low dimensional - smooth
dependence on µ (Lipschitz continuity)
(Adaptive) sampling procedure for parameter
exploration
reduced basis made of optimal solutions of the
FE “truth” problem
Evaluation procedure: (optimal) Galerkin
projection
[Noor, Peters, Porshing, Lee, ...]
MN
XN
M = U(µ) ∈ X : µ ∈ D
MN = UN (µ) ∈ XN : µ ∈ D
UN (µ1)
UN (µN )
XN = spanUN (µi ), i = 1, . . . ,N
How to be rigorous, rapid and reliable?
depends on the sampling procedure for parameter exploration (greedy algorithm)
exploits an Offline/Online stratagem based on the affinity assumption:
B(U,W;µ) =
Qb∑q=1
Θqa (µ)Bq(U,W), F(W;µ) =
Qf∑q=1
Θqf (µ)Fq(W)
relies on a posteriori error analysis
Introduction Saddle-point formulation RB method Results
RB method: smooth parametric dependency
Guess: MN is low dimensional - smooth
dependence on µ (Lipschitz continuity)
(Adaptive) sampling procedure for parameter
exploration
reduced basis made of optimal solutions of the
FE “truth” problem
Evaluation procedure: (optimal) Galerkin
projection
[Noor, Peters, Porshing, Lee, ...]
MN
XN
M = U(µ) ∈ X : µ ∈ D
MN = UN (µ) ∈ XN : µ ∈ D
UN (µ1)
UN (µN )
XN = spanUN (µi ), i = 1, . . . ,N
How to be rigorous, rapid and reliable?
depends on the sampling procedure for parameter exploration (greedy algorithm)
exploits an Offline/Online stratagem based on the affinity assumption:
B(U,W;µ) =
Qb∑q=1
Θqa (µ)Bq(U,W), F(W;µ) =
Qf∑q=1
Θqf (µ)Fq(W)
relies on a posteriori error analysis
Introduction Saddle-point formulation RB method Results
Reduced Basis Method: approximation stability
Reduced Basis (RB) approximation: given µ ∈ D, find (xN (µ), pN (µ)) ∈ XN × QN :A(xN (µ),w ;µ) + B(w , pN (µ);µ) = 〈F (µ),w〉 ∀w ∈ XN
B(xN (µ), q;µ) = 〈G(µ), q〉 ∀q ∈ QN
(∗)
How to define the reduced basis spaces?
we have to provide a spaces pair XN ,QN that
guarantee the fulfillment of an equivalent parametrized Brezzi inf-sup condition
βN (µ) = infq∈QN
supw∈XN
B(w , q;µ)
‖w‖X ‖q‖Q≥ β0, ∀µ ∈ D.
For the state and adjoint variables: aggregated spaces
YN ≡ QN = span
yN (µn), pN (µn)N
n=1
For the control variable:
UN = span
uN (µn)N
n=1
Let XN = YN × UN , we can prove that
βN (µ) ≥ αN (µ) > 0 being αN (µ) the coercivity constant associated to the FE
approximation of the PDE operator
Brezzi theorem =⇒ for any µ ∈ D, the RB approximation (∗) has a unique solution
depending continuously on the data
Introduction Saddle-point formulation RB method Results
Reduced Basis Method: approximation stability
Reduced Basis (RB) approximation: given µ ∈ D, find (xN (µ), pN (µ)) ∈ XN × QN :A(xN (µ),w ;µ) + B(w , pN (µ);µ) = 〈F (µ),w〉 ∀w ∈ XN
B(xN (µ), q;µ) = 〈G(µ), q〉 ∀q ∈ QN
(∗)
How to define the reduced basis spaces? we have to provide a spaces pair XN ,QN that
guarantee the fulfillment of an equivalent parametrized Brezzi inf-sup condition
βN (µ) = infq∈QN
supw∈XN
B(w , q;µ)
‖w‖X ‖q‖Q≥ β0, ∀µ ∈ D.
For the state and adjoint variables: aggregated spaces
YN ≡ QN = span
yN (µn), pN (µn)N
n=1
For the control variable:
UN = span
uN (µn)N
n=1
Let XN = YN × UN , we can prove that
βN (µ) ≥ αN (µ) > 0 being αN (µ) the coercivity constant associated to the FE
approximation of the PDE operator
Brezzi theorem =⇒ for any µ ∈ D, the RB approximation (∗) has a unique solution
depending continuously on the data
Introduction Saddle-point formulation RB method Results
Reduced Basis Method: approximation stability
Reduced Basis (RB) approximation: given µ ∈ D, find (xN (µ), pN (µ)) ∈ XN × QN :A(xN (µ),w ;µ) + B(w , pN (µ);µ) = 〈F (µ),w〉 ∀w ∈ XN
B(xN (µ), q;µ) = 〈G(µ), q〉 ∀q ∈ QN
(∗)
How to define the reduced basis spaces? we have to provide a spaces pair XN ,QN that
guarantee the fulfillment of an equivalent parametrized Brezzi inf-sup condition
βN (µ) = infq∈QN
supw∈XN
B(w , q;µ)
‖w‖X ‖q‖Q≥ β0, ∀µ ∈ D.
For the state and adjoint variables: aggregated spaces
YN ≡ QN = span
yN (µn), pN (µn)N
n=1
For the control variable:
UN = span
uN (µn)N
n=1
Let XN = YN × UN , we can prove that
βN (µ) ≥ αN (µ) > 0 being αN (µ) the coercivity constant associated to the FE
approximation of the PDE operator
Brezzi theorem =⇒ for any µ ∈ D, the RB approximation (∗) has a unique solution
depending continuously on the data
Introduction Saddle-point formulation RB method Results
RB method: algebraic formulation
We can express the RB state, control and adjoint solutions as
xN (µ) =3N∑j=1
xNj (µ)σj , denoting XN = spanσj
3N
j=1
pN (µ) =2N∑j=1
pNj (µ)τj , denoting QN = spanτj
2N
j=1
recalling the assumption of affine parametric dependence (also for the RHS)
A(x ,w ;µ) =
Qa∑q=1
Θqa (µ)Aq(x ,w), B(w , p;µ) =
Qb∑q=1
Θqb(µ)Bq(w , p)
for a new parameter µ, the RB solution can be written as a combination of basis functions
with weights given by the following reduced basis linear system:
3N∑j=1
Qa∑q=1
Θqa (µ)Aq
ij xNj (µ) +2N∑l=1
Qb∑q=1
Θqb(µ)Bq
li pNl (µ) =
Qf∑q=1
Θqf (µ)F q
i , 1 ≤ i ≤ 3N
3N∑j=1
Qb∑q=1
Θqb(µ)Bq
lj xNj (µ) =
Qg∑q=1
Θqg (µ)G q
l , 1 ≤ l ≤ 2N
size: 5N × 5N
Introduction Saddle-point formulation RB method Results
RB method: Offline/Online decomposition
3N∑j=1
Qa∑q=1
Θqa (µ)Aq
ij xNj (µ) +2N∑l=1
Qb∑q=1
Θqb(µ)Bq
li pNl (µ) =
Qf∑q=1
Θqf (µ)F q
i , 1 ≤ i ≤ 3N,
3N∑j=1
Qb∑q=1
Θqb(µ)Bq
lj xNj (µ) =
Qg∑q=1
Θqg (µ)G q
l , 1 ≤ l ≤ 2N,
Offline pre-processing: compute and store the basis function σii and τjj
store the matrices Aq and Bq given by
(AN )qij = Aq(σj , σi ), (BN )q
li = Bq(σi , τl ), 1 ≤ i , j ≤ 3N, 1 ≤ l ≤ 2N,
as well as the RHS F qi , G q
l .
Operation count: depends on N, Qa, Qb, Qf , Qg and N
Online evaluations: evaluate coefficients Θq∗(µ), assemble RB matrices and vectors
AN (µ) =∑Qa
q=1 Θqa (µ)Aq
N , BN (µ) =∑Qb
q=1 Θqb(µ)Bq
N
FN (µ) =∑Qf
q=1 Θqf (µ)F q
N , GN (µ) =∑Qg
q=1 Θqf (µ)G q
N ;
and solve
(AN (µ) BT
N (µ)
BN (µ) 0
)︸ ︷︷ ︸
KN (µ)
(xN (µ)
pN (µ)
)︸ ︷︷ ︸
UN (µ)
=
(FN (µ)
GN (µ)
)
Operation count: O((5N)3 + (Qa + Qb)N2 + (Qf + Qg )N)
independent of N , N N
Introduction Saddle-point formulation RB method Results
RB method: Offline/Online decomposition
3N∑j=1
Qa∑q=1
Θqa (µ)Aq
ij xNj (µ) +2N∑l=1
Qb∑q=1
Θqb(µ)Bq
li pNl (µ) =
Qf∑q=1
Θqf (µ)F q
i , 1 ≤ i ≤ 3N,
3N∑j=1
Qb∑q=1
Θqb(µ)Bq
lj xNj (µ) =
Qg∑q=1
Θqg (µ)G q
l , 1 ≤ l ≤ 2N,
Offline pre-processing: compute and store the basis function σii and τjj
store the matrices Aq and Bq given by
(AN )qij = Aq(σj , σi ), (BN )q
li = Bq(σi , τl ), 1 ≤ i , j ≤ 3N, 1 ≤ l ≤ 2N,
as well as the RHS F qi , G q
l . Operation count: depends on N, Qa, Qb, Qf , Qg and N
Online evaluations: evaluate coefficients Θq∗(µ), assemble RB matrices and vectors
AN (µ) =∑Qa
q=1 Θqa (µ)Aq
N , BN (µ) =∑Qb
q=1 Θqb(µ)Bq
N
FN (µ) =∑Qf
q=1 Θqf (µ)F q
N , GN (µ) =∑Qg
q=1 Θqf (µ)G q
N ;
and solve
(AN (µ) BT
N (µ)
BN (µ) 0
)︸ ︷︷ ︸
KN (µ)
(xN (µ)
pN (µ)
)︸ ︷︷ ︸
UN (µ)
=
(FN (µ)
GN (µ)
)
Operation count: O((5N)3 + (Qa + Qb)N2 + (Qf + Qg )N)
independent of N , N N
Introduction Saddle-point formulation RB method Results
RB method: Offline/Online decomposition
3N∑j=1
Qa∑q=1
Θqa (µ)Aq
ij xNj (µ) +2N∑l=1
Qb∑q=1
Θqb(µ)Bq
li pNl (µ) =
Qf∑q=1
Θqf (µ)F q
i , 1 ≤ i ≤ 3N,
3N∑j=1
Qb∑q=1
Θqb(µ)Bq
lj xNj (µ) =
Qg∑q=1
Θqg (µ)G q
l , 1 ≤ l ≤ 2N,
Offline pre-processing: compute and store the basis function σii and τjj
store the matrices Aq and Bq given by
(AN )qij = Aq(σj , σi ), (BN )q
li = Bq(σi , τl ), 1 ≤ i , j ≤ 3N, 1 ≤ l ≤ 2N,
as well as the RHS F qi , G q
l . Operation count: depends on N, Qa, Qb, Qf , Qg and N
Online evaluations: evaluate coefficients Θq∗(µ), assemble RB matrices and vectors
AN (µ) =∑Qa
q=1 Θqa (µ)Aq
N , BN (µ) =∑Qb
q=1 Θqb(µ)Bq
N
FN (µ) =∑Qf
q=1 Θqf (µ)F q
N , GN (µ) =∑Qg
q=1 Θqf (µ)G q
N ;
and solve
(AN (µ) BT
N (µ)
BN (µ) 0
)︸ ︷︷ ︸
KN (µ)
(xN (µ)
pN (µ)
)︸ ︷︷ ︸
UN (µ)
=
(FN (µ)
GN (µ)
)
Operation count: O((5N)3 + (Qa + Qb)N2 + (Qf + Qg )N)
independent of N , N N
Introduction Saddle-point formulation RB method Results
RB Method: a posteriori error estimation
Goal: provide rigorous, sharp and inexpensive estimators for the error on the solution
variables and for the error on the cost functional
General noncoercive (Babuska) framework (as Stokes problems, e.g. [Rozza et al., 2010]):
Recall the compact (weakly coercive) formulation: given µ ∈ D,
find U(µ) ∈ X s.t: B(U(µ),W;µ) = F(W;µ) ∀W ∈ X .
Note that ∃ β0 > 0 : β(µ) = infW∈X
supU∈X
B(U,W;µ)
‖U‖X ‖W‖X≥ β0, ∀µ ∈ D.
the FE and RB approximations satisfy an equivalent inf-sup condition
define the residual r(W;µ) = F(W;µ)− B(UN ,W;µ) and apply Babuska theorem
A posteriori error estimation on the solution variables
(‖xN (µ)− xN (µ)‖2
X + ‖pN (µ)− pN (µ)‖2Q
)1/2 ≤‖r(·;µ)‖X ′βLB(µ)
:= ∆N (µ)
0 < βLB(µ) ≤ βN (µ) is a constructible lower bound given by the
successive constraint method (SCM) [R., Huynh, Manzoni 2010]
thanks to the affinity assumption, we can provide the standard
Offline/Online stratagem for the efficient computation of ‖r(·;µ)‖X ′
Introduction Saddle-point formulation RB method Results
RB Method: a posteriori error estimation
Goal: provide rigorous, sharp and inexpensive estimators for the error on the solution
variables and for the error on the cost functional
General noncoercive (Babuska) framework (as Stokes problems, e.g. [Rozza et al., 2010]):
Recall the compact (weakly coercive) formulation: given µ ∈ D,
find U(µ) ∈ X s.t: B(U(µ),W;µ) = F(W;µ) ∀W ∈ X .
Note that ∃ β0 > 0 : β(µ) = infW∈X
supU∈X
B(U,W;µ)
‖U‖X ‖W‖X≥ β0, ∀µ ∈ D.
the FE and RB approximations satisfy an equivalent inf-sup condition
define the residual r(W;µ) = F(W;µ)− B(UN ,W;µ) and apply Babuska theorem
A posteriori error estimation on the solution variables
(‖xN (µ)− xN (µ)‖2
X + ‖pN (µ)− pN (µ)‖2Q
)1/2 ≤‖r(·;µ)‖X ′βLB(µ)
:= ∆N (µ)
0 < βLB(µ) ≤ βN (µ) is a constructible lower bound given by the
successive constraint method (SCM) [R., Huynh, Manzoni 2010]
thanks to the affinity assumption, we can provide the standard
Offline/Online stratagem for the efficient computation of ‖r(·;µ)‖X ′
Introduction Saddle-point formulation RB method Results
RB Method: a posteriori error estimation
Goal: provide rigorous, sharp and inexpensive estimators for the error on the solution
variables and for the error on the cost functional
General noncoercive (Babuska) framework (as Stokes problems, e.g. [Rozza et al., 2010]):
Recall the compact (weakly coercive) formulation: given µ ∈ D,
find U(µ) ∈ X s.t: B(U(µ),W;µ) = F(W;µ) ∀W ∈ X .
Note that ∃ β0 > 0 : β(µ) = infW∈X
supU∈X
B(U,W;µ)
‖U‖X ‖W‖X≥ β0, ∀µ ∈ D.
the FE and RB approximations satisfy an equivalent inf-sup condition
define the residual r(W;µ) = F(W;µ)− B(UN ,W;µ) and apply Babuska theorem
A posteriori error estimation on the solution variables
(‖xN (µ)− xN (µ)‖2
X + ‖pN (µ)− pN (µ)‖2Q
)1/2 ≤‖r(·;µ)‖X ′βLB(µ)
:= ∆N (µ)
0 < βLB(µ) ≤ βN (µ) is a constructible lower bound given by the
successive constraint method (SCM) [R., Huynh, Manzoni 2010]
thanks to the affinity assumption, we can provide the standard
Offline/Online stratagem for the efficient computation of ‖r(·;µ)‖X ′
Introduction Saddle-point formulation RB method Results
RB Method: a posteriori error estimation
Goal: provide rigorous, sharp and inexpensive estimators for the error on the solution
variables and for the error on the cost functional
Define the error on the cost functional as
JN (µ)− JN (µ) = J(yN (µ), uN (µ);µ)− J(yN (µ), uN (µ);µ)
with standard arguments (e.g. [Becker et al., 2000; Dede, 2010]) we can easily show that
JN (µ)− JN (µ) =1
2r(UN (µ)− UN (µ);µ)
A posteriori error estimation on the cost functional
|JN (µ)− JN (µ)| ≤1
2‖r(·;µ)‖X ′‖UN (µ)− UN (µ)‖X ≤
1
2
‖r(·;µ)‖2X ′
βLB(µ):= ∆J
N (µ).
The error estimator ∆JN (µ) does not need any additional ingredients than those already
discussed:
- the efficient computation of ‖r(·;µ)‖X ′
- the calculation of the lower bound βLB(µ)
Introduction Saddle-point formulation RB method Results
RB Method: a posteriori error estimation
Goal: provide rigorous, sharp and inexpensive estimators for the error on the solution
variables and for the error on the cost functional
Define the error on the cost functional as
JN (µ)− JN (µ) = J(yN (µ), uN (µ);µ)− J(yN (µ), uN (µ);µ)
with standard arguments (e.g. [Becker et al., 2000; Dede, 2010]) we can easily show that
JN (µ)− JN (µ) =1
2r(UN (µ)− UN (µ);µ)
A posteriori error estimation on the cost functional
|JN (µ)− JN (µ)| ≤1
2‖r(·;µ)‖X ′‖UN (µ)− UN (µ)‖X ≤
1
2
‖r(·;µ)‖2X ′
βLB(µ):= ∆J
N (µ).
The error estimator ∆JN (µ) does not need any additional ingredients than those already
discussed:
- the efficient computation of ‖r(·;µ)‖X ′
- the calculation of the lower bound βLB(µ)
Introduction Saddle-point formulation RB method Results
RB Method: a posteriori error estimation
Goal: provide rigorous, sharp and inexpensive estimators for the error on the solution
variables and for the error on the cost functional
Define the error on the cost functional as
JN (µ)− JN (µ) = J(yN (µ), uN (µ);µ)− J(yN (µ), uN (µ);µ)
with standard arguments (e.g. [Becker et al., 2000; Dede, 2010]) we can easily show that
JN (µ)− JN (µ) =1
2r(UN (µ)− UN (µ);µ)
A posteriori error estimation on the cost functional
|JN (µ)− JN (µ)| ≤1
2‖r(·;µ)‖X ′‖UN (µ)− UN (µ)‖X ≤
1
2
‖r(·;µ)‖2X ′
βLB(µ):= ∆J
N (µ).
The error estimator ∆JN (µ) does not need any additional ingredients than those already
discussed:
- the efficient computation of ‖r(·;µ)‖X ′
- the calculation of the lower bound βLB(µ)
Introduction Saddle-point formulation RB method Results
RB Method: the “complete game”
FE kernel
affine decomposition
assembly KqN , Fq
N
successive constraint
method (SCM)
βLB(µ)
basis selection
by greedy algorithm
assembly KqN, Fq
N
a posteriori
error estimation
∆N (µ)
assembly
KN (µ), FN (µ)
solution of the RB-OCPµ
KN (µ)UN (µ) = FN (µ)Θq∗(µ)µ ∈ D
certification
solution UN (µ),∆N (µ)
functional JN (µ),∆JN (µ)
offline
online
Offline stage involves precomputation of FE structures required for the RB space
construction and the certified error estimates.
Online stage has complexity only depending on N and allows resolution of the Optimal
Control Problem for any µ ∈ D with a certified error bound.
Introduction Saddle-point formulation RB method Results
Numerical results - IOptimal control of parametrized advection-diffusion equations
Introduction Saddle-point formulation RB method Results
Boundary control for a Graetz convection-diffusion problem
Ω1o Ω2
o(µ)
(1 + µ2, 0)
(1 + µ2, 1)(1, 1)(0, 1)
(0, 0) (1, 0)
ΓoN
ΓoC
ΓoCΓo
D
ΓoD
ΓoD
Ωo
Ωo the desired function is parameter dependent:
yd (µ) = µ3χΩo
parameter domain: D = [6, 20]× [1, 3]× [0.5, 3]
We consider the following optimal control problem:
minimize J(yo (µ), uo (µ);µ) =1
2‖yo (µ)− yd (µ)‖2
L2(Ωo )+α
2‖uo (µ)‖2
L2(ΓoC
)
subject to
−1
µ1∆yo (µ) + xo2(1− xo2)
∂yo (µ)
∂xo1= 0 in Ωo (µ)
yo (µ) = 1 on ΓoD
1
µ1∇yo (µ) · n = uo (µ) on Γo
C (µ)
1
µ1∇yo (µ) · n = 0 on Γo
N (µ),
- µ1 Peclet number, µ2 heigth of the channel
- the flow has an imposed temperature at the inlet and a parabolic convection field
- observation in the thermal layer Ωo
Introduction Saddle-point formulation RB method Results
Boundary control for a Graetz convection-diffusion problem
Ω1o Ω2
o(µ)
(1 + µ2, 0)
(1 + µ2, 1)(1, 1)(0, 1)
(0, 0) (1, 0)
ΓoN
ΓoC
ΓoCΓo
D
ΓoD
ΓoD
Ωo
Ωo
min J(yo (µ), uo (µ); µ) =1
2‖yo (µ)− yd (µ)‖2 +
α
2‖uo (µ)‖2
s.t.
−1
µ1∆yo (µ) + xo2(1− xo2)
∂yo (µ)
∂xo1= 0 in Ωo (µ)
1
µ1∇yo (µ) · n = uo (µ) on Γo
C (µ)
+ BCs
Functional spaces: Yo = H1ΓD
(Ωo ) Uo = L2(ΓoC ) Qo ≡ Yo
observation space: Zo = L2(Ωo )
I the problem is mapped to a reference domain Ω = Ωo (µref) with µref = (·, 1, ·)
I we obtain an affine decomposition with Qa = 1, Qb = 5, Qf = 1, Qg = 4
I we fixed α = 0.07 and we used P1 finite elements for the FE approximation of the
state, adjoint and control variables
Introduction Saddle-point formulation RB method Results
Boundary control for a Graetz convection-diffusion problem
1 1.4 1.8 2.2 2.6 330.1
0.2
0.3
0.4
0.5
0.6
comparison of Brezzi inf-sup constant βN (µ) and
coercivity constant of the state equation αN (µ)
as a function of µ2
Number of FE dof N 8915
Number of parameters P 3
Number of RB functions N 39
Dimension of RB linear system 39 · 5Affine operator components Q 6
Linear system dimension reduction 50:1
FE evaluation tFE (s) 14.5
RB evaluation tonlineRB (s) 0.1
RB evaluation tofflineRB (s) 3970
state yN adjoint yN
1 1.4 1.8 2.2 2.6 3
0
0.5
1
1.5
optimal control uN on ΓC
Representative solution for µ = (12, 2, 2.5).
Introduction Saddle-point formulation RB method Results
Boundary control for a Graetz convection-diffusion problem
1 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6 2.8 30
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
βLB
(µ)
βRB
(µ)
βFE
(µ)
Lower bound βLB(µ) for the Babuska inf-sup constant
β(µ) as a function of the parameter µ1
Number of FE dof N 8915
Number of parameters P 3
Number of RB functions N 39
Dimension of RB linear system 39 · 5Affine operator components Q 6
Linear system dimension reduction 50:1
FE evaluation tFE (s) 14.5
RB evaluation tonlineRB (s) 0.1
RB evaluation tofflineRB (s) 3970
5 10 15 20 25 30 3510
−6
10−4
10−2
100
102
104
106
average error
∆N average
max error
∆N max
5 10 15 20 25 30 3510
−10
10−5
100
105
average error
∆N
J average
Error estimation (•) and true error (•) for the solution (left) and the cost functional (right)
Introduction Saddle-point formulation RB method Results
Numerical results - IIOptimal control of parametrized Stokes problems
Introduction Saddle-point formulation RB method Results
How to extend the method
minimize J(v, p, u;µ) =1
2m(v − vd (µ), v − vd (µ);µ) +
α
2n(u, u;µ) subject to
a(v, ξ;µ) + b(ξ, p;µ) = 〈F (µ), ξ〉+ c(u, ξ;µ) ∀ξ ∈ V ,
b(v, τ ;µ) = 〈G(µ), τ〉 ∀τ ∈ M,
Functional setting: V = [H1(Ω)]2 M = L2(Ω) velocity and pressure spaces
Y = V ×M state space, Q ≡ Y adjoint space, U control space
two nested saddle-point
• outer: optimal control • inner: Stokes constraint
reduced basis functions computed by solving N times the FE approximation (with
stable spaces pair for velocity and pressure variables)
stability of the RB approximation of the Stokes constraint fulfilled by introducing
suitable supremizer operators [R., Veroy, 2007; R., Huynh, Manzoni 2010]
stability of the RB approximation of the whole optimal control problem fulfilled by
defining suitable aggregated spaces for the state and adjoint variables
Introduction Saddle-point formulation RB method Results
Test 1: confirming theoretical results (Couette flow)
Ωo(µ)
(1, 0)
(1, µ1)(0, µ1)
(0, 0)
ΓoN
ΓoD
ΓoD
ΓoD
µ1: height of the pipe
forcing term state equation: fo (µ) = (0,−µ2)
parameter domain: D = [0.5, 2]× [0.5, 1.5]
We consider the following optimal control problem
minimize J(vo (µ), po (µ), uo (µ)) =1
2‖vo1(µ)− xo2‖2
L2(Ωo )+α
2‖uo (µ)‖2
L2(Ωo )
s.t.
− ν∆vo +∇po = fo (µ) + uo in Ωo (µ)
div vo = 0 in Ωo (µ)
vo1 = xo2, vo2 = 0 on ΓoD (µ)
− po no1 + ν∂vo1
∂no1= 0, vo2 = 0 on Γo
N (µ),
Example of uncontrolled flow
velocity on the left
pressure on the right
Introduction Saddle-point formulation RB method Results
Test 1: confirming theoretical results (Couette flow)
Ωo(µ)
(1, 0)
(1, µ1)(0, µ1)
(0, 0)
ΓoN
ΓoD
ΓoD
ΓoD
min J(vo (µ), po (µ), uo (µ)) =1
2‖vo1(µ)− xo2‖2 +
α
2‖uo (µ)‖2
s.t.
− ν∆vo +∇po = fo (µ) + uo in Ωo (µ)
div vo = 0 in Ωo (µ)
vo1 = xo2, vo2 = 0 on ΓoD (µ)
− po no1 + ν∂vo1
∂no1= 0, vo2 = 0 on Γo
N (µ),
Functional spaces: velocity Vo = [H1ΓD
(Ωo )]2 and pressure Mo = L2(Ωo )
state Yo = Vo ×Mo adjoint Qo ≡ Yo control Uo = [L2(Ωo )]2
I the problem is mapped to a reference domain Ω = Ωo (µref) with µref = (1, ·)
I we obtain an affine decomposition with Qa = 1, Qb = 4, Qf = 1, Qg = 5
I we fixed α = 0.008, ν = 0.1 and used P2-P1 Taylor-Hood finite elements for the FE
approximation of the velocity and pressure variables (for the control variable we used P2
finite elements)
Introduction Saddle-point formulation RB method Results
Test 1: confirming theoretical results (Couette flow)
0.5 1 1.5 2
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
comparison of Brezzi inf-sup constant βN (µ) and
Babuska in-sup constant αN (µ) of the Stokes
operator.
Number of FE dof N 17439
Number of parameters P 2
Number of RB functions N 15
Dimension of RB linear system 225
Affine operator components Q 5
Linear system dimension reduction 80:1
FE evaluation tFE (s) 16.1
RB evaluation tonlineRB (s) 0.1
RB evaluation tofflineRB (s) 7820
Representative solutions for µ = (1.7, 1.5):
(state) pressure on the left, (state) velocity in the middle, control on the right.
Introduction Saddle-point formulation RB method Results
Test 1: confirming theoretical results (Couette flow)
0,5 1 1,5 20
0.005
0.01
0.015
0.02
0.025
0.03
0.035
βLB
(µ)
βFE
(µ)
Lower bound βLB(µ) for the Babuska inf-sup constant
β(µ) as a function of the parameter µ1
Number of FE dof N 17439
Number of parameters P 2
Number of RB functions N 15
Dimension of RB linear system 225
Affine operator components Q 5
Linear system dimension reduction 80:1
FE evaluation tFE (s) 16.1
RB evaluation tonlineRB (s) 0.1
RB evaluation tofflineRB (s) 7820
2 4 6 8 10 12 14
10−6
10−4
10−2
100
102
average error
∆N
average
2 4 6 8 10 12 14
10−10
10−5
100
average error
∆N
J average
Error
estimation (•) and true error (•) for the solution (left) and the cost functional (right)
Introduction Saddle-point formulation RB method Results
An (idealized) application in haemodynamics: a data assimilation problem
we consider an inverse boundary problem in hemodynamics, inspired by the recent works
[D’Elia et al., 2011; Perego et al., 2011]
simplified model of an arterial bifurcation
we suppose to have a measured velocity profile on the red section, but not the Neumann
flux on ΓC that will be our control variable
starting from the velocity measures we want to find the control variable in order to
retrieve the velocity and pressure fields in the whole domain.
120 Chapter 5. Reduced Basis Method for Parametrized Stokes Optimal Control problems
−4 −3 −2 −1 0 1 2 3
−2
−1
0
1
2
3
Γin
ΓD
ΓD
ΓC
ΓD
ΓC
ΓD
ΓD
ΓD
given new geometrical configuration (µgeom)and parametrized measurements µmeas
on the red section
−4 −3 −2 −1 0 1 2 3
−2.5
−2
−1.5
−1
−0.5
0
0.5
1
1.5
2
2.5
−0.2
0
0.2
0.4
0.6
−4 −3 −2 −1 0 1 2 3
−2.5
−2
−1.5
−1
−0.5
0
0.5
1
1.5
2
2.5
0
0.2
0.4
0.6
0.8
Online
vd(µmeas)g(µin)
find the unknown Neumann boundarycondition on ΓC and retrieve
the whole velocity and pressure fields
pressure
velocity
Figure 5.6: An (idealized) example of inverse boundary problem in haemodynamics. Given ageometrical configuration and some velocity measurements on some sections of the domain (bothobtainable via medical image and data assimilation devices, e.g. MRI), we want to retrieve the wholepressure and velocity fields in order to detect possible pathologies, e.g. occlusions or flow disturbancein arterial bifurcations.
The state velocity and pressure variables vo, po satisfy the following Stokes problem in theoriginal domain Ωo(µ):
−ν∆vo +∇po = 0 in Ωo(µ),
divvo = 0 in Ωo(µ),
v = 0 on ΓoD(µ),
vo = g(µin) on Γoin(µ),
−pono + ν∂vo∂no
= uo on ΓoC(µ),
(5.5.1)
where uo is the control variable; the inlet velocity profile g(µin) is given by the following
−4 −3 −2 −1 0 1 2 3
−2
−1
0
1
2
3
Γin
ΓD
ΓD
ΓC
ΓD
ΓC
ΓD
ΓD
ΓD
Γoss
(a) Boundary conditions: no-slip conditions onΓD(µ), Poiseuille velocity profile g(µin) on Γin,unknown Neumann flux on the outflow sections.
−4 −3 −2 −1 0 1 2 3
−3
−2
−1
0
1
2
µ3
µ4
µ6
µ5
µ1 µ
2
(b) Parametrization of the original do-main.
Figure 5.7: Original domain Ωo(µ) of the arterial bifurcation.
Introduction Saddle-point formulation RB method Results
An (idealized) application in haemodynamics: a data assimilation problem
−4 −3 −2 −1 0 1 2 3
−2
−1
0
1
2
3
Γin
ΓD
ΓD
ΓC
ΓD
ΓC
ΓD
ΓD
ΓD
Γoss
−4 −3 −2 −1 0 1 2 3
−3
−2
−1
0
1
2
µ3
µ4
µ6
µ5
µ1 µ
2
The state velocity and pressure variables vo , po satisfy the
following Stokes problem in the original domain Ωo (µ):
−ν∆vo +∇po = 0 in Ωo (µ),
div vo = 0 in Ωo (µ),
vo = 0 on ΓoD (µ),
vo = g(µin) on Γoin(µ),
−po no + ν∂vo
∂no= uo on Γo
C (µ),
Then we consider the following (parametrized) cost functional to be minimized
J(vo (µ), po (µ), uo (µ);µ) =1
2
∫Γo
obs
|vo (µ)− vd (µ)|2 dΓo
+α1
2
∫Γo
C
|∇uo (µ)to |2dΓo +α2
2
∫Γo
C
|uo (µ)|2dΓo
Introduction Saddle-point formulation RB method Results
An (idealized) application in haemodynamics: a data assimilation problem
5 10 15 20 25 30 35 40
10−4
10−3
10−2
10−1
100
101
Average computed error between the truth FE solution
and the RB approximation.
Number of FE dof N 40000
Number of parameters P 8
Number of RB functions N 43
Dimension of RB linear system 43 · 13
Affine operator components Q 71
FE evaluation tFE (s) 22
RB evaluation tonlineRB (s) 0.15
µ = (1, π/5, π/6, 1, 1.7, 2.2, 0.8, 1). µ = (1.2, π/6, π/6, 0.8, 2.5, 2.1, 0.3, 1).
Introduction Saddle-point formulation RB method Results
Conclusions
we provided a certified RB method for the solution of parametrized optimal
control problems with high dimensional control variable
reduction to low-dimensionality of the whole control problem and not just of the
state equation
key ingredient: saddle-point formulation of the optimal control problem
full Offline/Online decomposition strategy
...
References
Bibliography I
Barrault, M., Maday, Y., Cuong, N. Nguyen, & Patera, A.T. 2004. An ‘empirical interpolation’ method: application to
efficient reduced-basis discretization of partial differential equations. C. R. Acad. Sci. Paris. Ser. I Math., 339(9), 667
– 672.
Becker, R., Kapp, H., & Rannacher, R. 2000. Adaptive Finite Element Methods for Optimal Control of Partial Differential
Equations: Basic Concept. SIAM J. Control Optim., 39, 113–132.
Brezzi, F., & Fortin, M. 1991. Mixed and Hybrid Finite Elements Methods. New York: Springer-Verlag.
Dede, L. 2010. Reduced Basis Method and A Posteriori Error Estimation for Parametrized Linear-Quadratic Optimal
Control Problems. SIAM J. Sci. Comput., 32, 997–1019.
Dede, L. 2011. Reduced Basis Method and Error Estimation for Parametrized Optimal Control Problems with Control
Constraints. Journal of Scientific Computing, 1–19.
D’Elia, M., Mirabella, L., Passerini, T., Perego, M., Piccinelli, M., Vergara, C., & Veneziani, A. 2011. Applications of
Variational Data Assimilation in Computational Hemodynamics. In: Ambrosi, D., Quarteroni, A., & Rozza, G. (eds),
Modelling of Physiological Flows. MS&A Series, vol. 5. Springer.
Grepl, M., Maday, Y., Nguyen, N., & Patera, A.T. 2007. Efficient reduced-basis treatment of nonaffine and nonlinear partial
differential equations. Esaim Math. Model. Numer. Anal., 41(3), 575–605.
Grepl, M.A., & Karcher, M. 2011. Reduced basis a posteriori error bounds for parametrized linear-quadratic elliptic optimal
control problems. C. R. Math. Acad. Sci. Paris, 349(15-16), 873 – 877.
Gunzburger, M.D., & Bochev, P.B. 2004. Finite element methods for optimization and control problems for the Stokes
equations. Comp. Math. Appl., 48, 1035–1057.
Lions, J.L. 1971. Optimal Control of Systems governed by Partial Differential Equations. Berlin Heidelberg: Springer-Verlag.
References
Bibliography II
Perego, M., Veneziani, A., & Vergara, C. 2011. A variational approach for estimating the compliance of the cardiovascular
tissue: an inverse fluid-structure interaction problem. SIAM J. Sci. Comput., 33(3), 1181–1211.
Quarteroni, A., Rozza, G., & Quaini, A. 2007. Reduced basis methods for optimal control of advection-diffusion problems.
Pages 193–216 of: Advances in Numerical Mathematics.
Rozza, G., & Veroy, K. 2007. On the stability of the reduced basis method for Stokes equations in parametrized domains.
Comput. Methods Appl. Mech. Engrg., 196(7), 1244 – 1260.
Rozza, G., Huynh, D.B.P., & Patera, A.T. 2008. Reduced Basis Approximation and a Posteriori Error Estimation for
Affinely Parametrized Elliptic Coercive Partial Differential Equations. Arch. Comput. Methods Eng., 15, 229–275.
Rozza, G., Huynh, D.B.P., & Manzoni, A. 2010. Reduced basis approximation and a posteriori error estimation for Stokes
flows in parametrized geometries: roles of the inf-sup stability constants. Tech. rept. 22.2010. MATHICSE.
Submitted.
Tonn, T., Urban, K., & Volkwein, S. 2011. Comparison of the reduced-basis and POD a-posteriori estimators for an elliptic
linear-quadratic optimal control problem. Math. Comput. Model. Dyn. Syst., 17(1), 355–369.
For more information see
http://sma.epfl.ch/∼rozza
http://cmcs.epfl.ch/
http://augustine.mit.edu