144
1 Higher Order Modelling for Computational Electromagnetics Roberto D. Graglia Dipartimento di Elettronica Politecnico di Torino – Italy e-mail: [email protected]

DL_PERU

Embed Size (px)

DESCRIPTION

DL_PERU

Citation preview

  • 1Higher Order Modelling forComputational Electromagnetics

    Roberto D. GragliaDipartimento di ElettronicaPolitecnico di Torino Italy

    e-mail: [email protected]

  • WHERE IS TORINO (TURIN)?

    In the heart of EuropeThe taste of industryThe industry of tasteThe shape of technologyResources for development

  • TORINO

    Turin, the Capital of the Alps, is wellknown as the home of the Shroud of Turin,the headquarters of automobilemanufacturers Fiat, Lancia and Alfa Romeo(Torino is the Automobile Capital ofItaly), and as host of the 2006 WinterOlympics. It is one of Italy's main industrialcenters, being part of the famous industrialtriangle, along with Milan and Genoa.

    Torino was founded by the Romans in the first century BC, it was the capital of the Duchyof Savoy from 1563, then of the Kingdom of Sardinia ruled by the Royal House of Savoyand finally the first capital of unified Italy (1861).

    The city currently hosts some of Italy's best universities, colleges, academies, such as thePolytechnic University of Turin. Prestigious and important museums, such as theEgyptian Museum (the 2nd in the world after the Cairo one) and the Mole Antonelliana arealso found in the city. Turin's several monuments and sights make it one of the world's top250 tourist destinations, and the tenth most visited city in Italy in 2008.

  • Francesco Vercelli (22/10/1883 - 24/11/1952) Geofisico piemontese, laureato a Torino nel 1908 in fisica e nel 1909 in matematica, si occup principalmente del mare e delle altre acque

    Amedeo Avogadro (09/08/1776 - 09/07/1856) Scienziato piemontese che studi le propriet dei gas con l'enunciazione del principio che porta il suo nome

    Gustavo Colonnetti (08/11/1886 - 20/03/1968) Il fondatore dellIstituto di Metrologia del CNR di Torino

    Bernardino Drovetti (1776 - 1852) Archeologo e diplomatico piemontese "padre" del Museo Egizio di Torino

    Cesare Lombroso (06/11/1835 - 19/10/1909) Psichiatra, antropologo e criminologo del XIX secolo

    Giuseppe Luigi Lagrange (25/01/1736 - 10/04/1813) Matematico torinese famoso in tutta lEuropa di fine Settecento

    Giovanni Antonio Amedeo Plana (06/11/1781 - 20/01/1864) Astronomo, fu il vero fondatore dell'Osservatorio Astronomico di Torino. La sua fama specialmente legata agli scritti su la "Teoria della Luna"

    Galileo Ferraris (30/10/1847 - 07/02/1897) Elettrotecnico, studi i motori a correnti alternate e fond a Torino la prima scuola italiana per ingegneri elettrotecnici

    Giovanni Schiaparelli (14/03/1835 - 04/07/1910) Uno dei pi prolifici astronomi italiani del XIX secolo, famoso per le sue osservazioni della superficie di Marte

    Primo Levi (31/07/1919 - 11/04/1987) Scrittore ebreo piemontese celebre per aver narrato lorrore dei campi di concentramento, chimico di professione, la sua opera un interessante esempio di connubio fra letteratura e scienza

    Rita Levi Montalcini (22/04/1909 30/12/2012) Scienziata neurobiologa piemontese vincitrice nel 1986 del premio Nobel per la medicina

  • The Politecnico has 32.000 students studying on 78 courses (18 held in English, 24Doctorates). In the academic year 2012/2013 the Politecnico had around 4.900 students inthe first year; in 2012 over 5.300 students graduated with a Master of Science or aBachelor's Degree. Each year, between lectures, laboratories and practical exercises thereare 170.000 hours of teaching. There is a staff of over 900 lecturers and researchers, andaround 800 administration staff. The income in the 2013 forecast balance is 280 millionEuros (in 1990 the figure was 52 million). The Ministero dell'Istruzione Universit eRicerca or M.I.U.R. (Ministry for Universities Education and Research) contributes aroundthe 46% of the income.

  • 6Higher Order Modelling forComputational Electromagnetics

    Roberto D. GragliaDipartimento di Elettronica e Telecomunicazioni

    Politecnico di Torino Italy

    e-mail: [email protected]

  • Considered equations & problems

    We deal with Electromagnetic problems modelled byMaxwells equations.

    The techniques described have applications to manypractical problems, e.g.: EMC & EMP; Shielding radiationfrom printed circuits; Microwave hazards; Electromagneticradiation from and penetration into vehicles, aircrafts,ships, missiles; Antennas near ground; Design offrequency selective surfaces for reflector antennas andradomes; Radar scattering; Propagation in opticalcomponents & systems; etc.

    7

  • 8

  • Numerical methods (for linear problems)- in 3 slides -

    9

    Introduce a set of testing functions {wm}, and a suitable inner product

    For a linear, inhomogeneous equation:

  • 10

    Then obtain a matrix equation

    for a square, non-singular matrixwith

    The solution

    or approximate depending on the choice of {fn} and {wm}

    if wn=fn Galerkins method

    may be exact

  • How to choose {fn}, {wm}? The set {fn} (linearly independent functions) has to

    approximate f reasonably well; The set {fn} could be chosen in order to satisfy the

    BCs (FEM applications) ; The set {wm} (linearly independent functions) has to

    be chosen so that depends on relatively independent properties of g;

    The choice depends on the desired accuracy of the solution;

    The ease of evaluation of matrix elements; The realization of a well-conditioned matrix [m,n].

    11

  • 12

    Representations of Fields in Terms of Potentials (Frequency-Domain)

    Solutions of Maxwells equations can be expressed in terms of potentials:

  • 13

    Representations of Fields in Terms of Potentials (Frequency-Domain)

    Solutions of Maxwells equations can be expressed in terms of potentials:

  • To invert the differential operator is to solve the problem to do this one has to know (find) the Green function G of the given problem;

    If G is known, then the solution is simply obtained by integration;

    Unfortunately, for all problems of practical interest G is unknown numerical approaches to differential/integral problems are required because of the ignorance of the Green function;

    The numerical problem could be well- or bad-posed; this depends on the properties of the operator;

    14

  • In general, in integral formulations, use of the proper (unknown) G is avoided by resorting to general principles (e.g., equivalence principle), or to integral theorems (e.g., Green theorems);

    Integral formulations guarantee the BCs and all the physical properties of the solutions, as a matter of fact, in many cases, the proper integral equations are obtained only at the moment of imposing the BCs (in the finite region);

    This is why IE are widely used in EM (since radiation conditions / far-field conditions are automatically satisfied);

    Because of the Green kernel, the matrix approximation of the problem is given by a dense matrix.

    15

  • FEM MOMDifferential formulation Integral formulation

    Non linear problems can be dealt withApplications to linear problems are well known

    BCs have to be enforced BCs automatically satisfied

    Infinite domains: there is some problems Infinite domains: no problem

    Sparse matrices Dense matrices (problems in dealing with very large problems)

    The integrals to be evaluated are simpleThe singularities of the kernel render numerical integration quite difficult

    Complex geometry are easily dealt with

    Higher order models have been introduced more recently

    16

  • Higher order modelling

    Required:

    To better represent the geometry of the problem. To reduce the L2 (least squares norm) of the

    solution error on regions of interest.

    - It usually improves the convergence of the results and/or reduces the size of the numerical matrices.

  • Higher order modelling

    18

    Regarding the geometry of the problem;

    need to approximate the curvature of the geometry;need to approximate small details in multiscale problems.

    use subsectional bases thereby subdividing the geometry intosimple (curved) elements of small size.

  • Use subsectional bases thereby subdividing the geometryinto simple (curved) elements of small size.

    19

    Pablo Picasso: "Retrato de Ambroise Vollard (1910).

  • Regarding the geometry of the problem

    20

    Contribution from specular points (in the optical limit).Points of reflection on the body at which the angle of incidence isequal to the angle of reflection relative to the observation point.

    e.g., the RCS of a metallic sphere of radius a is =a2

  • 21

    In the high-frequency range, the PEC-sphere models shown here poorly model the commonly used sphere-benchmark (see below)

    NASA Communications SatelliteEcho Project (1960-1969)

  • Regarding the geometry of the problem

    22

    Travelling waves (contribution from points with high curvature).

    e.g., estimation of the near-nose-oncross-sections of long, thin bodies.

  • Regarding the geometry of the problem

    23

    Creeping waves

    of importance for the analysis in the shadow region(the importance of the creeping waves depends on the dimension in wavelength of the body).

  • Higher order modelling

    Regarding the expansion/testing functions used in the numerical application.

    reduce the L2 (least squares norm) of the solution error on regions of interest;improve convergence of the results;reduce the size of the numerical matrices.

    use vector functions of high polynomial order.

    24

  • How to get curved (i.e., distorted) elements.

    25Source: Zinkiewiczs book

  • How to get curved / distorted elements.

    Any curved cell is obtained by mapping a parent cell into the object domain.

    The parent cell for triangular patches is a rectilinear triangle. In object space, the 3 edges are the zero-coordinate lines 1, 2, 3=0.

    i is the area coordinate i = Ai /AT Dependency relations: 1 + 2 + 3 = 1

  • How to get curved / distorted elements.

    The rectilinear triangle on the right-hand side is r = 1 r1+ 2 r2+ 3 r3

    Where 1, 2, 3 are 3 linear shape functions, each associated to a different corner node of the cell.

  • Interpolatory polynomialsfor shape functions.

    Use Lagrange interpolation polynomials written interms of interpolatory polynomials of Silvester.

  • Interpolatory polynomials of Silvester

    Use polynomials of degree i in , where is in the interval[0, 1].

    The parameter p indicates the number of uniform subintervals into which the interval is divided.

    The polynomial is unity at =i/p and has zeros at =0, 1/p, 2/p, , (i-1)/p

    01

    1)(!1

    ),(1

    0

    i

    pikpipR

    i

    ki

  • Silvester polynomials: example for p=1

    There are 2 polynomials: one of 0th and one of 1st degree. Please draw the polynomials on the interval [0, 1].

    01

    1)(!1

    ),(

    ),1(;1),1(

    1

    0

    10

    i

    pikpipR

    RR

    i

    ki

  • Silvester polynomials: example for p=1

    There are 2 polynomials: one of 0th and one of 1st degree. Please draw the polynomials on the interval [0, 1].

    ),1(;1),1( 10 RR

  • Silvester polynomials: example for p=2

    Three polynomials: one of 0th, one of 1st and one of 2nd degree. Please draw these polynomials on the interval [0, 1].

    01

    1)(!

    1),(

    21)12(2),2(;

    12),2(;1),2(

    1

    0

    210

    i

    pikpipR

    RRR

    i

    ki

  • Silvester polynomials: example for p=2

    Please draw these 3 polynomials on the interval [0, 1].

    21)12(2),2(;

    12),2(;1),2( 210

    RRR

  • 34

    Scalar Lagrangian interpolation on the canonical elements: the segment.

    There are two coordinates: 1, 2 . These are dependent coordinates (1+2 =1)

    ij(1, 2 )=Ri(p, 1)Rj(p, 2 ) with i+j=pis a p-th order Lagrangian polynomialinterpolating points within a segment whosenormalized coordinates (1, 2 ) are (i/p, j/p).

  • 35

    Interpolation on a segment, example for p=1

    There are two interpolatory polynomials of 1st degree. ij(1, 2 )=Ri(p, 1)Rj(p, 2 ), (with i+j=1).

    Please draw the polynomials on the interval 1=[0, 1].

    221102101

    120112110

    ),1(),1(),(),1(),1(),(

    RRRR

  • 36

    Interpolation on a segment, example for p=1

    221102101

    120112110

    ),1(),1(),(),1(),1(),(

    RRRR

  • 37

    Interpolation on a segment, example for p=2

    Example for p=2: there are three interpolatrory polynomials of 2nd degree.

    ij(1, 2 )=Ri(p, 1)Rj(p, 2 ), (with i+j=2). Please draw the polynomials on the interval 1=[0, 1].

    2)12(2),2(),2(),(

    22),2(),2(),(2

    )12(2),2(),2(),(

    2222102102

    2121112111

    1120122120

    RR

    RR

    RR

  • 38

    Interpolation on a segment for p=2

    2)12(2),2(),2(),(

    22),2(),2(),(2

    )12(2),2(),2(),(

    2222102102

    2121112111

    1120122120

    RR

    RR

    RR

  • 39

    Scalar Lagrangian interpolationon a segment.

    For the p-th order Lagrangian polynomial interpolation of a segment there are (p+1) interpolating polynomials of p-th degree.

  • Parameterization of the points on a triangle.

    Recall that the coordinates of a rectilinear triangle may beparameterized as

    where ri is a vertex position vector for vertex i.

    40

    ,332211r rrr

  • Parametrization of the points on a triangle.

    A Lagrange parametrization of order q for a curvilineartriangle can be expressed in terms of interpolatingpolynomials of the Silvester form as:

    where a triple indexing scheme is used to label the positionvector rijk interpolating the point with normalized coordinates(1, 2, 3)=(i/q, j/q, k/q).

    qkjiqRqRqR kjq

    kjiiijk

    ),,(),(),( 32

    0,,1

    rr

  • Geometry description for a triangle.

    Normalized coordinates related by 1 + 2+ 3 =1

    Edge vectors derived from the independent coordinates 1 and 2i=r/ i , i=1,2,

    from which the edge vectors that follow are found:

  • Edge vectors of a triangle.

    1 = - 2, 2 = 1, 3 = 2 - 1

  • Gradient vectors of a triangle.

    The gradient vectors are determined from the edgevectors as:

    i = (n x i ) /J

    where n = 1 x 2/J is the unit vector normal to thetriangle while J= | 1 x 2| is the Jacobian.

  • Edge and gradient vectors of a triangle.

    The above definitions for edge, (height) and gradient vectors apply for triangles on curved surfaces with an extended interpretation

    Triangle tangent to a curvilinear triangle at a point. The curvilinear and (rectilinear) tangent triangles have the same element coordinates, Jacobian, edge vectors, and height vectors at the point of tangency.

  • Why to use curved (i.e., distorted) elements?

    46

    the same model (that is, the same database) used by structural /mechanical engineers can be used this is a big plus!!!

    better approximations of boundaries are obtainedwith very small error in volume/surface/line models (no rescaling)

    usually a smaller number of unknowns is needed whenever higher-order expansion functions are used shorter computation time

  • Why to use curved (i.e., distorted) elements?

    47

    mixed approaches (FEM + MoM) are facilitated because the expansion functions of the two methods can be chosen within the same set.

    there is the possibility to construct singular functions able to model singular current or field behaviors (though applications of this are rather new).

    But, is there any problem?

  • But, is there any problem?

    48

    In FEM applications you now have to use quadrature to evaluate the matrix entries, whereas for non-distorted cells there are often closed form results possible.

    For MoM applications people are already using quadrature, but the treatment of the Green's function singularity might be different with curved cells.

    The main cost of curved cells is(1) the additional data structure necessary in the model,(2) the additional computation arising from the quadrature.

  • Vector functions

    We first consider the lowest order functions.

    Then we move on and consider higher-order vector functions.

    49

  • The family of generalized triangle basis functions (for current representation)

    50

    Source: Don Wiltons notes

  • Div-Conf functions on a triangular element

    51

    Divergence-conforming functions of the Nedelec typemaintain only normal continuity across elementBoundaries (they do not prescribe tangential continuity).

  • Div-Conf functions on a triangular element

    52

    They eliminates spurious solution of the EFIE operatorwhile discarding the highest order degrees of freedomassociated with the nullspace of the divergence operator inthe EFIE operator.

  • 53

    Divergence-conforming bases on triangles.

    .1

    ,1

    ,1

    12213

    31132

    23321

    J

    J

    J

    Zeroth-order bases. Three vector basis

    functions of first order. They have constant

    normal and linear tangential (CN/LT) components at element edges.

  • 54

    Curl-conforming bases on triangles.

    .,,

    :

    12213

    31132

    23321

    n Zeroth-order bases Three vector basis functions

    of first order. They have constant

    tangential and linear normal (CT/LN) components at element edges(prove this by use of eq. (48) of 1997 paper).

  • 55

    Completeness of zeroth order triangular bases(we consider only the curl-conforming ones).

    .,,

    12213

    31132

    23321

    The basis set is incomplete to first order since 6 degrees of freedom are required to model linear variations in two independent vector components on a surface.

  • 56

    Completeness of zeroth order triangular bases(we consider only the curl-conforming ones).

    .,,

    12213

    31132

    23321

    To render the bases first-order complete one must include the curl-free combinations:

    .,,

    1221

    22

    11

  • 57

    Completeness of zeroth order triangular bases(we consider only the curl-conforming ones).

    .,,

    12213

    31132

    23321

    Completeness to zeroth-order is proved by noticing that the following linear combinations are able to represent two independent basis vectors on a 2D element (verify this by yourself and express 3).

    .,

    223

    132

  • 58

    Completeness of zeroth order triangular bases(we consider only the curl-conforming ones).

    .,,

    12213

    31132

    23321

    The Nedelec conditions also require completeness of the curl to the same order as the bases. Completeness of the curl to zeroth-order follows from (verify this by use of (50, 53) of the 1997 paper).

    3,2,1,2 nJ

  • Higher order vector bases

    59

  • Higher order bases

    One can construct higher order bases complete to order p by forming the product of zeroth-order bases with complete polynomial factors of order p.

    The set of polynomials factors used may take one of several different forms chosen for convenience.

    60

    pkji:alhierarchicptsr:oryinterpolat

    jipsr :ousinhomogeneptsr:shomogeneou

    sj

    ri

    tsr

    0,),)))

    ,0,

    ,

    321

    321

    321

    ,,(H(p,R(p,R(p,R

    ijk

    tsr

  • Interpolatory vector functions

    For interpolatory vector functions the key idea is to use shifted polynomials of Silvester to move interpolation points away from two of the edges-those along with the tangential (for curl-conf.) and the normal (for div.-conf.) components of the 0th-order basis factor vanish.

    61

    Interpolation nodes for curl- or divergence-conforming bases on triangular elements. Only nodes in basis subset 1i j k or 1i j k for p = 3 are shown.

  • 62

    Interpolation nodes for curl- or divergence-conforming bases on quadrilateral elements. Only nodes in basis subset 3ik; j or 3ik; j for p = 2 are shown.

  • Interpolatory vector functions

    The polynomials with interpolating nodes as shown in figure are of global order p=3, and have the form:

    63

    Interpolation nodes for curl- or divergence-conforming bases on triangular elements. Only nodes in basis subset 1i j k or 1i j k for p = 3 are shown.

    21,,2,1,;,,1,0

    ),2(),2(),2( 321

    pkjipkjpipRpRpR kji

    with

  • 64

    Volumetric Elements

  • 65

    Volumetric Elements

  • 66

    Volumetric Elements

  • 67

    Volumetric Elements

  • 68

    0 1 2 3 4 5 6 7 8 9 100

    0.5

    1

    1.5

    2

    2.5

    3

    k0*a

    k

    z

    /

    k

    0

    m=1

    2

    3

    4

    5

    6 7

    3

    8

    2

    1

    1

    0

    2

    3

    4

    5

    6

    0

    Example of results for surface elements: INHOMOGENEOUSLY FILLED WAVEGUIDES

    a=2b, h=0.1br =10

    a

    bhr

    0

    analytical

    +, * FEM

  • 69

    102 103 10410-5

    10-4

    10-3

    10-2

    10-1

    100

    MATRIX DIMENSIONS

    R

    E

    L

    A

    T

    I

    V

    E

    E

    R

    R

    O

    R

    INHOM. FILLED WG - RELATIVE ERROR

    r0

    P=1

    P=2P=3

  • 70

    0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 20

    5

    10

    15

    Normal field component at the air-dielectric interface.

    P=0~ 1800 UNKNOWNS

    E

    y

    a

    /

    E

    y

    d

    x

    P=1

    P=2 P=3a

    bhr

    0

    P N0 16411 18732 18973 1681

    a =2b, h =0.2 br =10, k0a =7

    Fundamental mode

  • 71

    Image waveguide

    0 2 4 6 820

    25

    30

    35

    40

    F

    R

    E

    Q

    U

    E

    N

    Z

    A

    G

    H

    z

    Kz mm-1

    a

    ba`

    b`

    a =1.3 mm, b=1.6 mm, a` =0.55 mm, b`=0.82 mm

    rx =170, ry = ry 85

    [1] J.I.Askne, E.L. Kolberg, L. Pettersson,``Propagation in a waveguide partially filled with anisotropic dielectric material``, IEEE Trans. MTT, vol.30, n5, pp.795-799, Maggio 1982.

    24 triangoli19 nodi

    291 triangoli166 nodi

    - - p=0, (37 incognite), mesh densa- p=3, (325 incognite), mesh lasca

    espansione modale [1]

    modo 1

    modo 2

    3

    4

  • 72

    Hierarchical Vector Basis Functions for Meshes withHexahedra, Tetrahedra, and Triangular Prism Cells

    Roberto D. Graglia*, and Andrew F. Peterson

    * Politecnico di Torino ITALY Georgia Institute of Technology USA

    e-mail: [email protected] ; [email protected]

  • 73

    NEW VECTOR BASESHierarchical curl / divergence-conforming vector bases ofthe Nedelec kind for 2D and 3D cells that:

    A.Allow one to mesh a structure with differently shaped cells2D Triangles & Quadrilaterals3D Tetrahedrons & Bricks & Triangular Prisms

    B.Yield well-conditioned system matrices, even in case ofhigh order bases

  • Why Nedelec spaces?

    The Nedelec curl-conforming spaces eliminate the degrees of freedom associated with the gradient of a higher degree polynomial.

    They do this locally, without introducing global constraints that complicate the definition of basis functions, the sparseness of the FEM system, boundary conditions, etc.

    These are the easy degrees of freedom to discard (eliminating all the gradient DoF requires global operations, matrix partitioning, etc.)

  • 75

    Other existing vector bases

    In the open literature, only one group has developed (curl-conforming) bases with similar features; however those basesdo not satisfy Nedelecs constraints:S.Zaglmayr, High order Finite Element Methods forelectromagnetic field computation, Ph. D. Thesis, JohannesKepler Universitt, Linz, Austria, July 2006.

  • 76

    OUR PAPERS ON THIS HIERARCHICAL SUBJECT FOR

    TRIANGULAR & TETRAHEDRAL cells, QUADRILATERAL & BRICK & PRISM cells;

    1. A. F. Peterson, R. D. Graglia, Scale factors and matrix conditioning associated withtriangular-cell hierarchical vector basis functions IEEE Antennas and WirelessPropagation Letters, vol. 9, pages 40-43, 2010.2. R. D. Graglia, A. F. Peterson, and F. P. Andriulli, Curl-conforming hierarchical vectorbases for triangles and tetrahedra, IEEE TAP, vol. 59, no. 3, pp. 950-959, Mar. 2011.3. A. F. Peterson, R. D. Graglia, Evaluation of hierarchical vector basis functions forquadrilateral cells, IEEE Trans. Magn., due to appear May 2011.4. R. D. Graglia, and A. F. Peterson, Hierarchical curl-conforming Nedelec elements forquadrilateral and brick cells, IEEE TAP, accepted for publication, Dec. 2010.5. R. D. Graglia, and A. F. Peterson, Hierarchical curl-conforming Nedelec elements fortriangular-prism cells,'' in preparation, Nov. 2010.6. R. D. Graglia, and A. F. Peterson, Hierarchical divergence-conforming Nedelec elementsfor volumetric cells,'' Mar. 2011.

  • 77

    HIERARCHICAL BASES: The basis of order m is a subset of the basis of order (m+1). In FEM and MoM applications, these bases enable different expansion orders on different elements in the same mesh (p-adaptation)

    Example 2D-polynomial hierarchical bases:

    0__ 1

    1__ x ______ y

    2__ x2 ___ xy ____ y2

    3__ x3 __ x2 y ____ x y2 __ y3

    4__ x4 __ x3 y __ x2 y2 ___ x y3 __ y4

  • Our ideaWe define and work only with hierarchical

    polynomial bases;Then define a redundant complete vector set by

    multiplying the zeroth-order vector functions withthe hierarchical polynomials of the base;

    and then eliminate redundancy to define the(unisolvent) hierarchical vector bases.

    Same scheme used to define interpolatoryvector bases on elements of different shape.

    78

  • Our ideaor what we actually do

    We linearly combine the terms of the existingHIGH-ORDER INTERPOLATORY VECTORbases to formHIGH ORDER HIERARCHICAL VECTORbasesR. D. Graglia, D. R. Wilton, and A. F. Peterson, Higher order interpolatory vectorbases for computational electromagnetics, special issue on Advanced NumericalTechniques in Electromagnetics, IEEE Trans. Antennas Propagat., vol. 45, no. 3,pp. 329342, Mar. 1997.

    79

  • For the curl-conforming bases, the Generating Polynomials are subdivided from the outset into three different groups of edge (E), face (F), and volume-based (V) functions.

    For each family, the number of the generating polynomials and their maximum polynomial order is the same as in the interpolatory family.

    80

  • 81

    The orthogonalization along edges leads toLegendre polynomials

    (Pro) are hierarchical and could define edge-based polynomials;(Pro) are either symmetric or antisymmetric;(Con) are orthogonal on the cell edge but not on the faces attached to that edge.

    Linear combinations of the face-based polynomials are addedto the edge-based ones to make them orthogonal on the facealso.

  • 82

    Linear combinations of the volume-based polynomialsare added to the face-based polynomials to make themorthogonal on the volume also.

  • Curl-conforming basesThus, the definition process for the curl-conformingcase is:

    1) First define the volume-based polynomials;2) Then the face-based ones;3) Finally the edge-based ones.

    83

  • Div-conforming bases The generating polynomials are subdivided from the

    outset into two different groups of face (F), and volume-based (V) functions.

    For each family, the number of the generating polynomials and their maximum polynomial order is the same as in the interpolatory family.

    The definition process for the divergence-conforming case is slightly simpler than the curl-conforming:

    1. First define the volume-based polynomials;2. Then the face-based ones.

    84

  • Div-conforming bases In this case, the zeroth-order functions are not

    associated to the edges but to each cell-face. Thus, there is the need to choose two reference

    parent variables on each cell-face to write the generating orthogonal polynomials in a way to easily ensure the continuity of the normal component of the vector functions across adjacent elements (i.e., by sign adjustment).

    The other face variables are obtained from the dependency relations.

    85

  • Div-conforming bases The reference-variables are easily individuated by the

    pivoting edges of the face. The pivoting edges depart from the face corner-node with

    the lowest global node-number, and each reference-variable vanishes only on one of the two pivoting edges.

    86

  • To obtain and write down these polynomialssymmetry considerations are extensively used.

    This is done to make the tangent (or normal) vector component continuous across adjacent cells by sign adjustment only.

    The zeroth-order vector functions ALSO show some symmetry properties.

    87

  • To obtain and write down these polynomialssymmetry considerations are extensively used.

    The polynomials at issue are functions of the element parent variables; keep in mind that:

    Line element 2 dependent parent variables; Triangular 3 dependent parent variables; Quad. & Tetra. 4 dependent parent variables; Prism 5 dependent parent variables; Brick 6 dependent parent variables.

    88

  • There are two important issuesone has to consider in the process

    of defining a base

    1. Possibly, the orthogonality of the elements inthe base [this requires us to introduceappropriate inner products!!], and

    2. The scale-factor of each element in the base.

    89

  • Example: 3 different bases for vectors on-a-plane.

    Which is the best? How do we get it?

    90

  • We normalize the polynomialsin some clever manner

    (inner-product definition):

    the integral over the edge for edge-based polynomials Ep

    the integral over the face for face-based polynomials Fp

    the integral over the volume for volume-based polynomials Vp

    91

  • ALL the edge-based polynomials are madeorthogonal over their associated edge, face, andvolume;

    ALL the face-based polynomials are madeorthogonal over their associated face andvolume;

    ALL the volume-based polynomials are madeorthogonal over the associated volume.

    92

  • Some of the polynomial bases follow

    93

    Brick div-conforming bases

  • Few polynomial bases follow

    94

    Quadrilateral & brick curl-conforming bases

  • The polynomial bases are normalized

    95

    Quadrilateral & brick curl-conforming bases

  • Some numerical results

    96

  • 97

    Obtained by use of the Triangular family

  • 98

    Obtained by use of the Triangular family

  • 99

    Obtained by use of the Quad/Brick family

  • 100

    The Table shows the matrix condition numbers arising from an 18-cell model of a 2:1 rectangular cavity, constructed from a 6 by 3model with identical, uniform rectangular cells (3 rows of 6columns each). These results are also obtained from the set of 24quadratic-tangential/cubic-normal (QT/CuN) bases from eachfamily.

    Comparison for quadrilateral bases

  • 101

    The Table compares the matrix condition numbers arising from theregular 18-cell model of a 2:1 rectangular cavity, constructed from a 6 by3 model with identical, uniform rectangular cells (3 rows of 6 columnseach).

    Comparison for quadrilateral bases

  • The Table shows the matrix condition numbers arising from a different 18-cellmodel of a 2:1 rectangular cavity, constructed from a 6 by 3 model withinterior nodes irregularly located to produce skewed quadrilateral cells. Thisexample also used the set of 24 quadratic-tangential/cubic-normal (QT/CuN)bases from each family.

    102

    Comparison for quadrilateral bases

  • Since the Jorgensen and Graglia functions of order p=2 outperformed theother families, their performance for order p=3 was also investigated. Forthis comparison, the set of 40 unscaled p=3 bases (using just the originalscale factors) was employed, producing a system of order 612. The Tablepresents the condition numbers, for the two 18-cell models consideredabove.

    103

    Comparison for quadrilateral bases

    In summary, the Graglia and Jorgensen bases perform in a very similarmanner as indicated by their matrix condition numbers. The other basisfamilies produce more ill-conditioned matrices, suggesting that theirlinear independence is not as good.

  • For completeness, this Table reports the eigenvalues obtained for theregular 18-cell mesh, for orders p = 0, 1, 2, and 3, compared to the exactresults.

    104

  • With our construction scheme we have obtained the bases fortriangular & quadrilateral cells.

    With our construction scheme we have obtained the curl-conformingbases for tetrahedral, brick and prism cells.

    With our construction scheme we have obtained the divergence-conforming bases for 3D elements: tetrahedron, brick and prism.

    Our curl-conforming basis family produce well-conditioned matrices;the other basis families produce more ill-conditioned matrices,suggesting that their linear independence is not as good.

    The transitioning strategy for p-refinement is reported elsewhere. Our bases can be used to mesh a structure with differently shaped

    cells.

    105

    CONCLUSION

  • Singular vector bases

    106

  • 107

    Circular vaned waveguide FEM application

    =1/2=0

    elements type and conformity

  • 108

    Circular vaned waveguide: eigenmodes

    Singular behavior

    Numerical precision

    Regular mode

    Singular mode

    =1/2

  • 109

    Modeling capability: very small thicknessDouble-vaned Circular homogeneous waveguide

    =2/3

  • 110

    Modeling capability: Multiple singular verteces and curvilinear singular elements

    =2/3=/2

    =1/2=

    =3/4=2/3

  • 111

    The square PEC-plate problem at normal incidence MoM Application

    The results at left (a, c) were obtained by using the zeroth-order regular base (p = 0) on the dense mesh A. The results at right (b, d) were obtained by using the coarse mesh B and the singular base of order [p = 2, s = 0].

  • 112

    The square PEC-plate problem at normal incidence

  • 113

    (10 1) PEC-strip Normal incidence Ex Singualar bases p = 2, s = 0

  • 114

    The circular PEC-plate at normal incidence

  • 115

    The circular PEC-plate at normal incidence (d = /100)

  • 116

    Normal incidence on a (1 1) square PEC-plate with a hole of radius r = /10 centered at (x = 0.15, y = +0.15);the incident magnetic field is polarized in the y-direction.

  • 117

    Spherical PEC-shell of radius a = /(2 and aperture angle = 120illuminated by a planewave propagating in the positivez direction

  • 118

    FEM Analysis of DielectricLoaded Waveguides with Additive

    Hierarchical Singular Vector Elements

    Roberto D. Graglia*, Andrew F. PetersonLadislau Matekovits*, and Paolo Petrini*

    * Politecnico di Torino ITALY Georgia Institute of Technology USA

    e-mail: [email protected] ; [email protected]

    APS/URSI 2014, Memphis, TN, USA Thursday, July 10, 2014.

  • 119

    Scope of Presentation:

    Treatment of vertex singularities in a 2D triangular-cell mesh wedge angle singularity with known exponents

    High order hierarchical representations scalar (already done and published) vector curl-conf (done and in publication) vector div-conf (future work) substitutive vs. additive

    Examples from cavity/waveguide FEM problems

  • 120

    ReferencesR.D. Graglia, A.F. Peterson, L. Matekovits, Singular, hierarchical scalarbasis functions for triangular cells, IEEE Trans. AP, vol. 61, no. 7, pp. 3674-3692, July 2013.

    R.D. Graglia, A.F. Peterson, L. Matekovits, and P. Petrini, Hierarchicaladditive basis functions for the finite-element treatment of cornersingularities, Special Issue on Finite Elements for Microwave Engineering,Electromagnetics, vol. 34, pp. 171-198, March 2014.

    R.D. Graglia, A.F. Peterson, L. Matekovits, and P. Petrini, Singularhierarchical curl-conforming vector bases for triangular cells, IEEE Trans.AP, due to appear July 2014.

    R.D. Graglia, P. Petrini, A.F. Peterson, and L. Matekovits, Full-waveanalysis of inhomogeneous waveguiding structures containing corners withsingular hierarchical curl-conforming vector bases, IEEE AWPL, April 2014.

  • 121

    Conclusions from previous works(PEC edges or corners)

    Can achieve true high order behavior, even with edges

    However, representation in edge cells requires an additive expansion with multiple integer exponents and multiple fractional powers

    cells often a quarter wavelength or more in dimension

    1 2 3 4 5 6108

    106

    104

    102

    limit of accuracy

    purely polynomial

    p,1,0

    p,1,1

    p,2,2

    p,p,p

    R

    e

    l

    a

    t

    i

    v

    e

    e

    r

    r

    o

    r

    f

    o

    r

    k

    c

    1

    s

    t

    T

    E

    m

    o

    d

    e

  • 122

    PEC Wedge Singularity

    Use the infinite wedge solution to identify the series of fractional exponents needed

    n n(2 )

  • 123

    Dielectric Wedge Singularity

    Use the infinite wedge solution to identify the series of fractional exponents needed

    Unbounded fields region

    Wedge aperture angle (degrees)

    n

    c

    o

    e

    f

    f

    s

    .

    f

    o

    r

    d

    i

    e

    l

    e

    c

    t

    r

    i

    c

    w

    e

    d

    g

    e

    s

    (

    r

    e

    l

    .

    =

    1

    0

    )

    0 45 90 135 180 225 270 315 3600.5

    1

    1.5

    2

    2.5

    3

    3.5

    4

    4.5

    1~

    ~n

    n

    t

    z

    EE

  • 124

    Substitutive vs Additive:

    Substitutive: replace one regular basis function with a singular function Not recommended for high order bases

    Additive: keep original basis functions and add singular functions to that set necessary for high order/accuracy maximum flexibility difficulty: matrix condition numbers

  • 125

    Conclusions from previous works

    As expected, conditioning is a problem with additive functions some orthogonalization essential more complete orthogonality leads to lower condition

    numbers (in MoM applications the CN can often be lowered by

    using a non-Galerkin approach)

  • 126

    Proposed singular bases:

    Singular bases to be added to regular, hierarchical bases without changing the regular bases

    Hierarchical Accommodate a general set of exponents

    may have any number of exponents (may also have singularities at more than one vertex) may be orthogonal to any number of regular

    polynomial bases

  • 127

    The key idea:

    Field component approximation:

    nqn

    znpolyzz aFF

    1

    ddaaFF

    nq

    nnnpolytt

    1

    kn

    bb nn1

    0

  • 128

    Notation: Modified coordinates

    Singularity incorporatedthrough radial functions

    n: index of exponent k: number of polynomials used in R to enforce

    orthogonality to regular basis functions : exponent (for the singular functions we use only

    the non integer ones)

    i1 i11i

    1 i

    Rn (k,,)

  • 129

    Orthogonality

    and to Ri , for all i

  • 130

    Normalization

    1),,(:

    1),,(:

    1),,(:

    1

    0

    2

    1

    0

    2'

    1

    0

    2

    dkRnC

    dkRnB

    dkRnA

    n

    n

    n

    (3 possibilities)

  • 131

    Singular scalar basis functions:

    Combine radial dependence with Jacobi polynomials in to obtain

    where

    fm ( ) (2m 5)(m 3)(m 4)32(m1)(m 2) Pm(2,2) ( )

    j0i1 Rj ()1 4 j1i1 Rj ()1 4 jm Rj () fm2 ( ) 1 2

  • 132

    Singular vector basis functions: hierarchical polynomial basis subsets are those used in

    Graglia, Peterson, Andriulli, Curl conforming hierarchical vector bases for triangles and tetrahedra, IEEE Trans. AP, March 2011

    Idea is to construct vector bases from gradients of the scalar bases; these form gradient bases functions with zero curl

    Then we define also (irrational) non-zero curl vector functions

    Linear combinations with regular bases used to minimize matrix condition numbers

  • 133

    Singular vector basis functions:The (irrational) zero-curl subspace is modelled by:

    81),,(;

    41),,(

    ;8

    1),,(;4

    1),,(

    2

    2

    2

    2

    fkSkS

    fkRkR

    nn

    nn

    kn

    j

    jnjn

    nn bac

    kR n1

    1),,(

    kn

    j

    jnjn

    nn bac

    kS n1

    21),,(

    with:

  • 134

    Singular vector basis functions:The (irrational) non-zero-curl subspace is modelled by:

    kn

    j

    jnjn

    nn bac

    kS n1

    11),,(

    ddSSkT nnn 2),,(

    inn PkSV 12),,( PkTJ

    nV nn 12),,(

    with:

  • 135

    Singular vector basis functions:What is new with respect to what we did for the scalarbases:

    The radial functions are obtained numerically by usinga recursive algorithm that involves the solution of asquare linear system. (The normalizing coefficients are thesquare root of quadratic forms.)

    The list of the singularity coefficients does not considerthe integer singularity coefficients BUT can be changedaccording to the aperture angle of the wedge.

  • 136

  • 137

    Waveguide problem formulations

    z

    t

    TLz

    z

    t

    TL

    tTctT

    ScS

    ee

    DCCB

    keeA

    nFormulatioFieldTL

    eBkeAnFormulatioFieldT

    BkAnFormulatioScalar

    2

    2

    2

    000

    The mass-matrices are the B-matrices

    For homogeneous waveguides 222 cz kkk

  • 138

    1 2 3 4 5108

    106

    104

    102

    p order of the polynomial subset

    R

    e

    l

    a

    t

    i

    v

    e

    e

    r

    r

    o

    r

    f

    o

    r

    k

    c

    1

    s

    t

    T

    E

    m

    o

    d

    e

    purely polynomial, all formulations

    singular bases

    limit of accuracy

    Scalar formulationTfield formulationTLfield formulation

    1 2 3 4 5

    102

    104

    106

    108

    p order of the polynomial set

    C

    N

    T

    E

    p

    r

    o

    b

    l

    e

    m

    w

    i

    t

    h

    p

    u

    r

    e

    l

    y

    p

    o

    l

    y

    n

    o

    m

    i

    a

    l

    b

    a

    s

    e

    s

    Scalar formulationTfield formulationTLfield formulation

    0 1 20

    0.5

    1

    1.5

    2

    1 2 3 4 5

    102

    104

    106

    108

    p order of the polynomial subset

    C

    N

    T

    E

    p

    r

    o

    b

    l

    e

    m

    w

    i

    t

    h

    s

    i

    n

    g

    u

    l

    a

    r

    b

    a

    s

    e

    s

    Scalar formulationTfield formulationTLfield formulation

    Homogeneous waveguides canbe studied with all formulations

  • 139

    2 2.2 2.4 2.6 2.8 3 3.2 3.4 3.6150

    100

    50

    0

    50

    100

    150

    200

    250

    300

    Frequency GHz

    z

    N

    p

    /

    m

    z

    r

    a

    d

    /

    m

    Propagating modesEvanescent modesComplex ModesBackward Wave

    0 5 10 15 20 25 30

    0

    5

    10

    15

    20

    25

    y=6mm

    24 X 24mm metal box;12 X 12mm rod with rel=37.13;

    0 5 10 15 20 25

    0

    5

    10

    15

    20

    25

    First (regular) mode 0 5 10 15 20 250

    5

    10

    15

    20

    25

    Backward mode

  • 140

    0 6 12 18 240.7

    1.0

    1.5

    2.0

    2.5First mode withpolynomial base

    Mesh 22Mesh 38

    0 6 12 18 240.7

    1.0

    1.5

    2.0

    2.5

    x (along the y=6mm line)

    First mode withsingular base

    Mesh 22Mesh 38

    0 5 10 15 20 25 30

    0

    5

    10

    15

    20

    25

    y=6mm

    0 5 10 15 20 25 30

    0

    5

    10

    15

    20

    25

    y=6mm

    Continuity of Dnormal along the red line

  • 1410 5 10 15 20 25 300

    5

    10

    15

    20

    25

    y=6mm

    0 5 10 15 20 25 30

    0

    5

    10

    15

    20

    25

    y=6mm

    Continuity of Dnormal along the red line

    0 6 12 18 240.7

    1.0

    1.5

    2.0

    2.5Backward mode with polynomial base

    0 6 12 18 24

    1

    1.5

    2

    2.5Backward mode with singular base

    x (along the y=6mm line)

  • 1420 5 10 15 20 25 300

    5

    10

    15

    20

    25

    y=6mm

    0 5 10 15 20 25 30

    0

    5

    10

    15

    20

    25

    y=6mm

    The field is unboundedat the corners

    0 6 12 18 240

    0.2

    0.4

    0.6

    0.8

    1

    1.2 Backward mode with polynomial base

    0 6 12 18 240

    0.2

    0.4

    0.6

    0.8

    1

    1.2

    x (along the y=6mm line)

    Backward modesingular base

    Mesh 22Mesh 38

  • 143

    1 2 3 4 50

    500

    1000

    1500

    2000

    2500

    p order of the polynomial subset

    D

    e

    g

    r

    e

    e

    s

    o

    f

    f

    r

    e

    e

    d

    o

    m

    Polynomial basesSingular bases

    1 2 3 4 5108

    106

    104

    102

    p order of the polynomial subset

    R

    e

    l

    a

    t

    i

    v

    e

    e

    r

    r

    o

    r

    limit of accuracy

    Singular bases

    Polynomial bases

    First modeBackward mode

    1 2 3 4 5

    102

    104

    106

    108

    p order of the polynomial subset

    C

    o

    n

    d

    i

    t

    i

    o

    n

    n

    u

    m

    b

    e

    r

    Polynomial basesSingular bases

  • 144

    Conclusion:

    Additive basis sets offer flexibility and possibility of true high order accuracy when singularities are present

    Matrix conditioning is a major issue more orthogonalization = lower condition numbers