Scilab tools for PDE’s : Application to time-reversal

Bruno Pinçon and Karim Ramdani

Corida team INRIA and Institut Elie Cartan
University Henri Poincaré
Nancy, France
[email protected]
[email protected]

Outlines
1. introduction
2. sparse linear algebra tools
3. finite element tools
4. time-reversal in a 2D acoustic waveguide
5. conclusion

- Scilab tools for PDE’S - 2004 CCA/ISIC/CACSD - Scilab/Scicos: an Open Source Environnement for CACSD - 1
 Introduction

While Scilab is rich in tools to deal with ordinary differential equations

(ode’s) it is less practical for partial differential equations (pde’s)
numerical experiments.

some good points:

• almost complete sparse matrix algebra support but:
- the sparse solver (Sparse 1.3) is outperformed by more modern ones
(umfpack, superlu,...);
- some sparse operations (extraction/insertion) are slow.
• a routine for solving 1D pde’s system (bvode) but nothing for 2D.

our goals:
• design a pde’s toolbox for 2D (space) simulations such that:
- it may be easy to use;
- it may be adaptable;
- it may be well integrated in scilab.
• improve some scilab sparse matrix operations (this is necessary to deal
with high discretisations).

- Scilab tools for PDE’S - 2004 CCA/ISIC/CACSD - Scilab/Scicos: an Open Source Environnement for CACSD - 2
sparse matrix algebra 1: an interface onto a good sparse solver

We have developped a Scilab interface onto the umfpack v4-x sparse linear solver
(Tim Davis). Freely available (SCISPT toolbox) at the url:

http://www.iecn.u-nancy.fr/∼pincon/scilab/scilab.html.

x = umfpack(A,"\",b)
LUp = umf lufact(A)
main features:
x = umf lusolve(LUp, b [,A])
umf ludel([LUp])

A small benchmark (done on a Latitude D400 Dell laptop, umfpack using an ATLAS BLAS library)

matrix name order nnz t1 t2 t1 /t2

bcsstk24 3562 159910 22.59 0.26 87
bcsstk38 8032 355460 160.31 0.75 160 t1 current sparse solver
ex14 3251 65875 87.61 0.25 350 t2 umfpack solver
epb2 25268 175027 571.60 1.07 534
utm5940 5940 83842 170.43 0.29 594

Clearly these results show that it is time for Scilab to change its sparse
solver !

- Scilab tools for PDE’S - 2004 CCA/ISIC/CACSD - Scilab/Scicos: an Open Source Environnement for CACSD - 3
sparse matrix algebra 2: improvement of some sparse operations

insertion of a full matrix in a sparse matrix: A(ii, jj) = B

• When B is small comparing to A it is interesting to try in place insertion:
n = 3000;
A = sprand(n,n,6/n);
// get the indices of nnz elements
old insertion way 66 s
[ij,v,mn]=spget(A);
in place insertion 0.48 s
timer();
for k = 1:size(ij,1), A(ij(k,1),ij(k,2)) = 1; end
t = timer()

• But when in place insertion is not possible (or not interesting),

the insertion algorithm used by Scilab is too slow. We have
written a new insertion code based on (scilab sparse format being row oriented):
- fast copy of sparse rows;
- carrefully assembling of a row built from a sparse and a full row.
a small benchmark:
old insertion code 9.2 s
A = sprand(10000,10000,0.001);
new insertion code 0.1 s
B = rand(500,500);
Matlab insertion code 350 s
timer(); A(101:600,201:700) = B; t=timer();
Matlab in place insertion code 0.42 s

- Scilab tools for PDE’S - 2004 CCA/ISIC/CACSD - Scilab/Scicos: an Open Source Environnement for CACSD - 4
sparse matrix extraction: B = A(ii, jj)

• When a row has many elements it is interesting to use a dichtomic search onto
the column indices in place of a linear search (feature added in the scilab core);

• When the indexing domain is very large (for instance A is 20000 × 20000 and
ii = 1 : 10000, jj = 10001 : 20000) and A have a low element density (that is A
is a normal sparse matrix !) then the usual algorithm don’t perform well. But it is
possible to invert the search that is to do a loop on the sparse matrix elements and
test if they are in the extraction region. We have written a new insertion code based
on this idea.

Here is a small benchmark:

n = 20000; old extraction code 2.33 s
A = sprand(n, n, 6/n); new extraction code 0.04 s
timer(); C = A(1:n/2,n/2+1:n); t=timer(); Matlab extraction code 0.04 s

- Scilab tools for PDE’S - 2004 CCA/ISIC/CACSD - Scilab/Scicos: an Open Source Environnement for CACSD - 5
 finite elements tools 1: boundary description and domain
discretisation

Some remarks:
• For numerical experiments with pde’s control problems it is important to be
able to locate some regions (2D subdomain or 1D curve or even points) where a
specific action (like a feedback) will be applied;

• From a first simulation it may be useful to be able to refine the mesh where
needed.

We have chosen to use:

• triangle (J.R. Shewchuk) for meshing purpose, triangle is:
- fast;
- may be called through a C interface;
- can refine a previous mesh;
- may manage points / edges / triangles (regional) attributes;
- have many options.

• Scilab for the description (tlist) and the discretisation of the boundaries (and
also “internal boundaries”)

- Scilab tools for PDE’S - 2004 CCA/ISIC/CACSD - Scilab/Scicos: an Open Source Environnement for CACSD - 6
 boundary description and domain discretisation: how it works
• A boundary is described from elementary border pieces:
S = Seg(P1,P2, bdy_marker)
A = Arc(C, theta1, theta2, orientation, bdy_marker)
Ci = Circle(C, r, orientation, bdy_marker)
U = Ufunc(t1, t2, x_expr, y_expr, bdy_marker)
Sp = Spline(Pts, bdy_marker)
CSp= CSpline(Pts, bdy_marker)

• In fact the boundary is formed from one exterior closed contour and eventually with
one or several interior closed contour(s). The function MakeContour take several (or
only one) element border pieces and build a closed contour.
CC = MakeContour(S1, S2, S3, S4)

• The interior constrained curves (which are not part of the boundary) must be
described only from elementary border pieces.

• The boundary (with the eventually interior constrained curves) is discretised with:
[PtB, EdB, MarkEdB, ne] = BoundaryDiscretisation(h, CCext, CCint1,.., Ebp1, Ebp2, ...)

• From the boundary edges (and also internal edges), the triangulation is got with:
[P, T, markT, E, T2T] = triangulate(PtB, EdB [, ne, Region]);

- Scilab tools for PDE’S - 2004 CCA/ISIC/CACSD - Scilab/Scicos: an Open Source Environnement for CACSD - 7
boundary description and domain discretisation: an example

Γ1 P3
P2









 
Γ1 R0












 



  "internal curve"
Γ−2
P1 Γ4



 

 

  
 
 


 

 

  

    
"control region"
Γ−1


  

 
Γ2 
 
 

R 1   Γ 3

P4 Γ4
Γ1

P5 P6
Γ1

- Scilab tools for PDE’S - 2004 CCA/ISIC/CACSD - Scilab/Scicos: an Open Source Environnement for CACSD - 8
 the scilab script
P1 = [0;1]; P2 = [0;2]; P3 = [2;2]; P4 = -P1; P5 = -P2; P6 = [2;-2];
C = [0;0]; C1 = [1.3;1]; C2 = [1.3;-1]; h = 0.15; r = 0.4;

S1 = Seg(P3,P2,1);
S2 = Seg(P2,P1,1);
A1 = Arc(C,1, -%pi/2, %pi/2, -1, 2);
S3 = Seg(P4,P5,1);
S4 = Seg(P5,P6,1);
T = Ufunc(-%pi/2,%pi/2, "x=2+0.5*cos(t)", "y=2*sin(t)", 3);
Ce = MakeContour(S1,S2,A1,S3,S4,T);
Ci1 = MakeContour(Circle(C1,r,-1,4));
Ci2 = MakeContour(Circle(C2,r,-1,4));

// internals constrained curves:

Ti = Ufunc(-%pi/2,%pi/2, "x=1.8+0.5*cos(t)", "y=1.9*sin(t)", -1);
P7 = [1.2;0.4]; P8 = [1.9;0.4];
P10 = [1.2;-0.4]; P9 = [1.9;-0.4];
S5 = Seg(P7,P8,-2); S6 = Seg(P8,P9,-2);
S7 = Seg(P9,P10,-2); S8 = Seg(P10,P7,-2);

[PtB, EdB, MarkEdB, ne] = BoundaryDiscretisation(h,Ce,Ci1,Ci2,Ti,S5,S6,S7,S8);

Region = [1.1 1.3 ;... // x coordinate of a point in each region
0 0 ;... // y coordinate of a point in each region
0 1 ;... // region markers identifiers
%nan %nan]; // wanted max size of the triangles in each region
[P, T, markT, E, T2T] = triangulate(PtB, EdB, ne, Region);

- Scilab tools for PDE’S - 2004 CCA/ISIC/CACSD - Scilab/Scicos: an Open Source Environnement for CACSD - 9
 the resulting mesh

a mesh

boundary 1
2 boundary 2
boundary 3
boundary 4
internal boundary 1
internal boundary 2
region 0
region 1
number of triangles 1066
number of vertices 594
min triangle quality 0.830
0
mean triangle quality 0.943
max triangle quality 1.000

-2

0 1.2498288 2.4996576

- Scilab tools for PDE’S - 2004 CCA/ISIC/CACSD - Scilab/Scicos: an Open Source Environnement for CACSD - 10
 finite elements tools 2: mesh refinement

To refine the mesh one have to give an area constraint for each triangle:
[Pn, Tn, markTn, EdBn, MarkEdBn, En, T2Tn] = refine_mesh(P, T, markT, area_cstr, EdB, MarkEdB);

boundary 1
boundary 2
boundary 3
boundary 4
internal boundary 1
internal boundary 2
region 0
region 1
number of triangles 3068
number of vertices 1655
min triangle quality 0.797
mean triangle quality 0.947
max triangle quality 1.000

- Scilab tools for PDE’S - 2004 CCA/ISIC/CACSD - Scilab/Scicos: an Open Source Environnement for CACSD - 11
 finite elements tools 3: f.e.m. matrices and vector
Considering
 a problem of the form:
 −∇ · (K(x, y) × ∇u) + V (x, y) · ∇u + c(x, y)u = f (x, y) in Ω
(K(x, y) × ∇u).n = g(x, y) on ΓN
u = h(x, y) on ΓD


discretized with the classic f.e.m.. Each matrix can be computed with:
sp = SparsePattern(E, np)
...................... with matrix type:
cM = build_values_per_triangle(P, T, MarkT, dimM, ... "stiffness"
mkr1, fM1, mkr2, fM2, ...) "tensor_stiffness"
M = assemb_fem_matrix( matrix_type, P, T, sp, cM); "mass"
...................... "advection"
A = M + others matrices

build_values_per_triangle is quite flexible, each “function” fMx may be a scalar

or a scilab function or even a scilab list with a function and its parameters.
the right hand side vector is built with:
coefb = build_values_per_triangle(P, T, MarkT, 1, mkr1, f1, mkr2, f2, ...)
b = assemb_fem_rhs(P, T, coefb);

and the boundary conditions are introduced by:

b = put_neumann_conditions(b, P, EdB, MarkEdB, mkbn1, g1, mkbn2, g2, ...);
[J,valJ] = dirichlet_cond(P, EdB, MarkEdB, mkbd1, h1, mkbd2, h2, ...);
[A,b] = put_dirichlet_conditions(J, valJ, A, b);

- Scilab tools for PDE’S - 2004 CCA/ISIC/CACSD - Scilab/Scicos: an Open Source Environnement for CACSD - 12
 an application: time-reversal in a 2D acoustic waveguide

time reversal mirror (TRM)

a time-reversal mirror can:

• emit
• measure
• record
diffracted field • achieve specific treatments of
acoustic waves.
scattering
obstacles
(These mirrors have been developped in the team of
M. Fink at LOA (ESPCI, France))
incident field

The main idea is to use the reversibility of the wave equation in a non dissipative
medium to back-propagate signals to the sources that emitted them.

In particular time-reversal can be used to focus acoustic waves on scattering

obstacles whose position is unknown.

Here we want to experiment the D.O.R.T. (“Decomposition of the Time-Reversal

Operator”) method in the case of an (infinite) 2D acoustic waveguide.

- Scilab tools for PDE’S - 2004 CCA/ISIC/CACSD - Scilab/Scicos: an Open Source Environnement for CACSD - 13
 the D.O.R.T. method
1/ the TRM first emits an acoustic wave and measures the diffracted field.
2/ the measured field is then reversed and reemitted by the TRM which also
measures the new diffracted field.

The Time-Reversal Operator T is the operator obtained by this 2 steps procedure.

If one works with time-harmonic waves (UT (x, t) = Re(uT (x)e−iωt)) the reversing
operation corresponds to a conjugation, that is, if S is the operator corresponding to
the first step then:
T = S̄S
Under some assumptions (small scatterers, ...), it has been shown experimentally and
mathematically proved (not currently for waveguides but only for free space) that:

• the number of non zero (or significant) eigenvalues of T is exactly the

number of obstacles (eigenvalues modules being related to the size of the obstacles);
• the corresponding eigenvectors generate waves that focus selectively on
the obstacles.

We want to test and (numerically) experiment the case of 2D waveguides with

time-harmonic waves.

- Scilab tools for PDE’S - 2004 CCA/ISIC/CACSD - Scilab/Scicos: an Open Source Environnement for CACSD - 14
 the mathematical model

uD
uI
The scattering matrix XI XD
satisfies: XD = SXI

To obtain S, we solve a family of

boundary value problems of the form: 2
( ∆ + k ) uD = 0
(each b.v.p. corresponding to an T.B.C. T.B.C.
∂ uD ∂ uI
incident field uI associated to ΓA =− ΓB
∂n ∂n
one of the propagative modes of the
waveguide)

A11 A12 A13 others nodes

which yields through a finite element A=

discretisation to a matrix of the form: A21 A22 O nodes of ΓA

A31 0 A33 nodes of ΓB

SPARSE FULL

- Scilab tools for PDE’S - 2004 CCA/ISIC/CACSD - Scilab/Scicos: an Open Source Environnement for CACSD - 15
 some details on the implementation
Final steps:
• T = S̄S;
• then compute the eigenelements of T
• check the expected focusing properties.
Remarks:
2π
1/ the experimental setting imposes: a << λ = << d. Since the Helmholtz
k
fem approximation requires 10 points per wavelength, the mesh size must be very
small;
2/ to compute the matrix S the use of a fast solver is crucial;
3/ our fast insertion algorithm (full inside sparse) is very useful here to take into
account the TBC.
Our numerical test: uses the values, a1 = 0.13, a2 = 0.1, λ = 1, d = 5.2, L = 5.
This gives 11 propagative modes for the wave guide. With h = 0.02, we got 229114
triangles, 115084 vertices, the different timings being:
0.6 s (bdy description) 1.56 s (triangulation)
3.6 s (mesh optim) 3.7 s (matrix build)
1.9 s (TBC) 16.8 s (factorisation)
12 s (computing of S)

- Scilab tools for PDE’S - 2004 CCA/ISIC/CACSD - Scilab/Scicos: an Open Source Environnement for CACSD - 16
 eigenvalues of the time reversal operator

We must found 2 significant eigenvalues ! (is λ3 negligible ?)

eigenvalues function of the mesh size

λ
⊕ ⊕
0.053 ⊕ ⊕
⊕
⊕ ⊕
λ1
⊕
λ2
⊕
λ3
⊕
λ4
⊕
λ5
⊕
λ6
⊕
λ7
⊕
λ8
⊕
⊕ ⊕ λ9
0.021 ⊕ ⊕ ⊕
λ10
⊕ ⊕
λ11
⊕

⊕ ⊕ ⊕ ⊕ ⊕ ⊕
0.0044 h
⊕
⊕ ⊕
⊕ ⊕
⊕ ⊕
⊕ ⊕
⊕ ⊕
0
0.02 0.022 0.025 0.03 0.04 0.05

- Scilab tools for PDE’S - 2004 CCA/ISIC/CACSD - Scilab/Scicos: an Open Source Environnement for CACSD - 17
 acoustic fields associated to the 4 first eigenvectors

module of the total field associated with vp1 module of the total field associated with vp2

module of the total field associated with vp3 module of the total field associated with vp4

- Scilab tools for PDE’S - 2004 CCA/ISIC/CACSD - Scilab/Scicos: an Open Source Environnement for CACSD - 18
 Conclusion

Having a (free) interpreted environment (which is not too slow !) to do

pde’s experiment is something useful !

For instance, at the beginning (and in the paper !) we have used 2 TRM but the
focalisation results are not clear . . . and it stays many things to experience with one TRM !

Remarks:

1/ We think to complete progressively our pde’s toolbox and make a first public
release at the end of this year. You will find it at the url:
http://www.iecn.u-nancy.fr/∼pincon/scilab/scilab.html

2/ Our code for sparse insertion and extraction will be soon integrated in Scilab.

3/ We hope that the umfpack solver will replace the scilab ’s one.

- Scilab tools for PDE’S - 2004 CCA/ISIC/CACSD - Scilab/Scicos: an Open Source Environnement for CACSD - 19
 the mathematical model
• the use of time-harmonic waves
uT (x, t) = Re(u(x)e−iωt) simplifies x2
Γ∝
the wave equation into the Helmholtz
equation: ∆uT + k 2uT = 0, k = ω/c Ω∝
d
(uT = uI + uD being the total field); Γo Γo

• we use Neumann boundary conditions on (0,0) Γ∝ x1

the guide boundaries (Γ∞) and on the
obstacles (Γo): ∂uT /∂n = 0;
• our (abstract) mirror will generate an incident field (uI ) which is supposed to be
a superposition of the propagative (toward the obstacles) modes of the wave guide,
that is:
N  p p
X cn(x) = 2/d cos(λp n x), (c0(x) = 1/d)
uI = anIcn(x2)eiβnx1 , with 2 − λ2
n=0
λ n = nπ/d, β n = k n

• the diffracted field uD is the outgoing solution of the problem:


 ∆uD + k 2uD = 0 in Ω∞ \ Ō
∂uD /∂n = −∂uI /∂n on Γo ∪ Γ∞
uD = 0 on ∂Ω∞.


- Scilab tools for PDE’S - 2004 CCA/ISIC/CACSD - Scilab/Scicos: an Open Source Environnement for CACSD - 20
 the mathematical model (continued)

• the fact that uD is outgoing implies that (x = (x1, x2)):

 X Rd
D iβn x1


 a ne cn(x2) for x1 < 0 (aD
n = 0
uD (0, x2)cn(x2)dx2)
n∈
XN
uD (x) = D −iβn (x1 −L)
Rd


 b ne cn(x2) for x1 > L (bD
n = 0
uD (L, x2)cn(x2)dx2)
n∈N

• our TRM “prescribes” the coefficients XI = (a0I , a1I , . . . , aN

I ) of the incident field
and “measures” the coefficients XD = (a0D , a1D , . . . , aD
I ) of the diffracted field. The
scattering matrix S is the complex (N + 1) × (N + 1) matrix such that: XD = SXI
and the time reversal matrix is: T = S̄S.
x2
• to solve numerically the problem we must Γ
reduce it to problem posed on a bounded Ω
ΓA ΓB
domain (Ω). We use on ΓA and Γb d
Γo Γo
transparent boundary conditions. They may
be obtained by using the Dirichlet to Neumann x1
(0,0) Γ L
operator (explicitely known here).

- Scilab tools for PDE’S - 2004 CCA/ISIC/CACSD - Scilab/Scicos: an Open Source Environnement for CACSD - 21