Computational Electromagnetics: Software Development and High Frequency Modelling of Surface Currents on Perfect Conductors

SANDY SEFI

Doctoral Thesis Stockholm, Sweden 2005

TRITA-NA-0541 ISSN 0348-2952 KTH School of Computer Science and Communication ISRN KTH/NA/R-05/41-SE SE-100 44 Stockholm ISBN 91-7178-203-6 SWEDEN Akademisk avhandling som med tillstnd av Kungl Tekniska hgskolan framlgges till oentlig granskning fr avlggande av teknologie doktorsexamen i numerisk analys och datalogi fredagen den 27:e januari 2006 klockan 10.00 i sal E3, Lindstedtsvgen 3, Kungl Tekniska hgskolan, Stockholm. Sandy Se, December 2005 Tryck: Universitetsservice US AB

iii
Abstract In high frequency computational electromagnetics, rigorous numerical methods become unrealistic tools due to computational demand increasing with the frequency. Instead approximations to the solutions of the Maxwell equations can be employed to evaluate the electromagnetic elds. In this thesis, we present the implementations of three high frequency approximate methods. The rst two, namely the Geometrical Theory of Diraction (GTD) and the Physical Optics (PO), are commonly used approximations. The third is a new invention that will be referred to as the Surface Current Extraction-Extrapolation (SCEE). Specically, the GTD solver is a exible and modular software package which uses Non-Uniform Rational B-spline (NURBS) surfaces to model complex geometries. The PO solver is based on a triangular description of the surfaces and includes fast shadowing by ray tracing as well as contribution from edges to the scattered elds. GTD ray tracing was combined with the PO solver by a well thought-out software architecture. Both implementations are now part of the GEMS software suite, the General ElectroMagnetic Solvers, which incorporates state-of-the-art numerical methods. During validations, both GTD and PO techniques turned out not to be accurate enough to meet the industrial standards, thus creating the need for a new fast approximate method providing better control of the approximations. In the SCEE approach, we construct high frequency approximate surface currents extrapolated from rigourous Method of Moments (MoM) models at lower frequency. To do so, the low frequency currents are projected onto special basis vectors dened on the surface relative to the direction of the incident magnetic eld. In such conguration, we observe that each component displays systematic spatial patterns evolving over frequency in close correlation with the incident magnetic eld, thus allowing us to formulate a frequency model for each component. This new approach is fast, provides good control of the error and represents a platform for future development of high frequency approximations. As an application, we have used these tools to analyse the radar detectability of a new marine distress signaling device. The device, called "Rescue-Wing", works as an inatable radar reector designed to provide a strong radar echo useful for detection and positioning during rescue operations of persons missing at sea. Keywords: High frequency approximations, Maxwells equations, Method of Moments, Physical Optics, Geometrical Theory of Diraction, Extraction Extrapolation.

Preface
This thesis is focused on the development of high frequency techniques in computational electromagnetics. It is composed of two parts: rst, an introduction, giving background to the subject, followed by a list of papers, each describing relevant results. Paper 1: Sandy Se, Fredrik Bergholm, Modeling and Extrapolating Highfrequency Electromagnetic Currents on Conducting Obstacles. Proceedings ICNAAM 2005 International Conference of Numerical Analysis and Applied Mathematics, Rhodes, Greece, September 2005. Paper 2: Sandy Se, Fredrik Bergholm, Extrapolation and Modelling of Method of Moments Currents on a PEC surface. Technical report, TRITA-NA0539, Royal Institute of Technology presented at the MMWP05 Conference on Mathematical Modelling of Wave Phenomena, Vxj, Sweden, August 2005. Paper 3: Sandy Se, Jesper Oppelstrup, Physical Optics And NURBS for RCS Calculations. Proceedings EMB04 Conference on Computational Electromagnetics - Methods and Applications, Gteborg, Sweden, October 2004. Paper 4: Tomas Melin, Sandy Se, The Rescue Wing: Design of a Marine Distress Signaling Device. Proceedings OCEANS 2005 Conference on Ocean Science sponsored by the Marine Technology Society (MTS) and the IEEE Oceanic Engineering Society, Washington DC, United States, September 2005. Paper 5: Stefan Hagdahl, Fredrik Bergholm, Sandy Se, Modular Approach to GTD in the Context of Solving Large Hybrid Problems. Proceedings AP2000 Millennium Conference on Antennas & Propagation, Davos, Switzerland, April 2000. Paper 6: Sandy Se, Architecture and Geometrical Algorithms in MIRA, a Ray-based Electromagnetic Wave Simulator. Proceedings EMB01 Conference on Computational Electromagnetics - Methods and Applications, Uppsala, Sweden, November 2001.

Acknowledgements
This oeuvre has been carried out from August 2000 to September 2005 at the Department of Numerical Analysis and Computer Science (NADA), at the Royal Institute of Technology (KTH), in Stockholm, Sweden. During these years, it has been a true privilege to study at NADA. For this, I am grateful to my supervisor Prof. Jesper Oppelstrup who oered me such an opportunity. I thank him for his time, his constant encouragement and for always enlighting the positive side of a situation. I am also grateful to my coauthor in half of the papers, Fredrik Bergholm, for providing invaluable assistance and ultimately made this thesis a reality. His involvement has considerably enriched this work and I deeply appreciate his willingness to help. I also thank Stefan Hagdahl for his very early eorts as a contributor as well as Tomas Melin for a pleasant and rewarding collaboration and Martin Nilsson for his comments on the early draft. Lastly, a very special thank you to my ance Eva for her encouragement and great support. Financial support has been provided by NADA, KTH, PSCI and the National Aeronautical Research Program (NFFP) within the General ElectroMagnetic Solvers (GEMS) and the Signature Modeling and Reduction Tools (SMART) projects.

vii

Vita
Sandy Se was born on March 11, 1974 in Grenoble, France where he graduated in 1997 from the Joseph Fourier University (UJF) with a Degree in Applied Mathematics and Computer Science. He spent the last ve months of his graduate diploma working with Prof. Patrick Chenin in the LMC Laboratory at the Institute for Applied Mathematics of Grenoble (IMAG) on the development of a computer program for 3D geometrical modelling. After graduation, he was awarded a scholarship from the French Government to pursue a Master of Science Degree (DESS) in Ingnierie Mathmatiques at the same university. In May 1998, he left the country to complete a training period at NADA, KTH, in Sweden and in December the same year he enlisted for a sixteen months French national service as Cooprant du Service National en Sude, for which he obtained a scholarship from NADA to work on a survey of numerical methods in Computational Electromagnetics. At the same time, he begun to work in the GEMS project as well as on the Diplme de Recherche Technologique (DRT) which he defended in autumn 2000 at UJF. The same year he was granted a Ph.D. position at NADA focused on the development of high frequency methods under the supervision of Prof. Jesper Oppelstrup. In June 2003, he obtained the Licentiate Degree in Numerical Analysis and Computer Science at KTH concluding his work on ray tracing in Electromagnetics.

ix

Contents
Contents 1 Introduction 1.1 Outline and main results . . . . . . . . . . . . . . . . . . . . . . . . . 1.2 Background: the GEMS project . . . . . . . . . . . . . . . . . . . . . 2 Governing Field Equations 2.1 The Maxwell Equations . . . . . 2.2 Boundary conditions on a Perfect 2.3 Plane wave solution in free space 2.4 The vector Helmholtz Equation . 2.5 Greens function . . . . . . . . . 2.6 The eld integral equations . . . . . . . . . . Conductor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xi 1 2 4 7 8 10 12 12 14 15 19 19 20 23 25 26 31 31 31 32 32 33 34 35

3 Scattering Analysis Methods 3.1 Geometrical Theory of Diraction . . . . . . 3.2 Physical Optics . . . . . . . . . . . . . . . . . 3.3 Method of Moments . . . . . . . . . . . . . . 3.4 Hybrid Methods . . . . . . . . . . . . . . . . 3.5 Modelling the behavior of the surface currents

4 Summary of the Papers 4.1 Paper 1: Modelling and Extrapolation, an Extended Abstract 4.2 Paper 2: Modelling and Extrapolation, Continued . . . . . . . 4.3 Paper 3: Physical Optics and NURBS . . . . . . . . . . . . . 4.4 Paper 4: The Rescue Wing, an Engineering Application . . . 4.5 Paper 5: MIRA, a Modular Approach to GTD . . . . . . . . . 4.6 Paper 6: Architecture and Geometrical Algorithms in MIRA . Bibliography

xi

Chapter 1

Introduction
Computational electromagnetics (CEM) may be dened as the branch of electromagnetics that involves the use of computers to simulate electric and magnetic sources as well as the elds these sources produce in specied environments. In this work we address the CEM problem of predicting, as realistically as possible, the electromagnetic behavior of a three-dimensional conducting structure subjected to an incident wave. Such models can be applied to a large variety of industrial applications. Computation of mobile phone coverage, design and placement of transmitter or receiver antennas and stealth aircraft technologies, are a few examples. In the latter application, CEM simulations are conducted to assist the analysis of the radar signature of an aircraft, either during the design or at manufacturing stages. Ultimately they aim to replace very expensive and time consuming measurements. In all these diverse applications the physics is the same: an incident eld Einc or current source1 induces a current Js on the surface of the conductor, which, in turn, radiates back a scattered eld response Escat propagating throughout space.

A8132

Figure 1.1: Field scattered by an aircraft. In the monostatic conguration we look for reections in the radar direction.

1 For

example in the case of transmitting antenna.

CHAPTER 1. INTRODUCTION

Mesh

CEM Solver for Maxwell Equations

scat

CAD

E inc

RCS

Prototype

Measurements
Optimization

Figure 1.2: Design chain leading to the radar signature.

This dynamic interaction can be modeled in several ways. Our methods work in the frequency domain where all the excitations and their responses are sinusoidal waves of the same frequency. This allows removal of the time dependence in the variables. For radar applications, we focus on the high frequency range, a regime where the wavelengths of the electromagnetic waves are much shorter than the typical dimension of the conductor. We also assume that all the scatterers are perfect electric conductors (PEC). Physically this means that the elds do not penetrate the surfaces of the conductor resulting in a null electric eld inside the conductor. The excitations are all plane waves and we look in particular at the monostatic response, a typical conguration, exemplied in Figure 1.1, where sources and receivers are located at the same position. Figure 1.2 illustrates the typical design loop applied to the radar cross section (RCS) optimization. The RCS characterizes the eective area illuminated by the incident plane wave which is scattered back to the radar. A popular description of the concept of RCS can be found in Paper 4. RCS is measured in square meters, often normalized and displayed using a logarithmic scale dened in decibel square-meters (dBsm), see Figure 1.3. The reference 0dB corresponds to the return from a sphere of 1m2 cross section.

1.1

Outline and main results

The starting point for modelling the dynamics of the elds is provided by the fundamental Maxwell equations, to be discussed in detail in Chapter 2. Their solutions in the frequency domain give rise to a great variety of numerical methods. Three of them have been studied and implemented in this work. We classify them into two main groups of frequency methods, the ray-based and the current-based methods, see Chapter 3. High frequency approximations which employ rays for the computation of the elds will be referred to as ray-based. More precise denitions of rays and their

1.1. OUTLINE AND MAIN RESULTS

Figure 1.3: Typical RCS values in m2 and in dBsm for a range of dierent objects.

propagation will be given later. These methods require known asymptotic high frequency solutions such as Geometrical Optics and its generalization, the Geometrical Theory of Diraction (GTD, UTD) [Keller 62, Pathak 84, McNamara 89]. Ray-based methods are fast and have the advantage to provide a good insight in the physics of the elds. Paper 5 introduces the implementation of our GTD solver called MIRA and presents some illustrations for the rays. Paper 6 further describes the software architecture of MIRA and shows how to design a exible architecture which can handle a complex CAD geometry [Farin 88]. The idea is to avoid simplications of the geometry of the scatterers and to use, for computations, the same CAD geometry as used for manufacturing the prototype which is required for measurements. When solving the Maxwell equations for the currents on the conductors, the method will be referred to as current-based. A classic example is the crude currentbased approximation called Physical Optics (PO). Paper 3 presents the implementation of our PO solver. The aim was to assess eciency and error analysis by focusing on ecient numerical evaluation of the PO integral, which will be dened in Chapter 3. An adaptive triangular subdivision scheme for the surface integral is also provided. In Paper 4, we use this tool to assist the design of a new passive radar reector signaling device to be used in case of distress at sea. This represents

CHAPTER 1. INTRODUCTION

a good illustration of the potential of these methods to solve real life engineering problems. Both GTD and PO can provide very fast results but are approximations that do not incorporate a complete modelling of the physics and thus, typically degrade in accuracy. The last contribution is this work, introduced as an extended abstract in Paper 1 and further developed in Paper 2, draws the basis of a new approximate current-based approach. After a careful examination of the related work in the introduction, Paper 2 will lead us to the study of the extraction-extrapolation methods. These methods are based on a simple yet powerful idea: to numerically model (extrapolate) the behavior of the surface current vector elds at high frequency, using information extracted from lower frequency solutions. Our extraction-extrapolation approach obtains the behavior of the surface current from Method of Moments (MoM) models at lower frequency. The MoM, see [Harrington 68, Fares 99], can been viewed as exact brute force technique providing accurate induced currents, but which do not incorporate any knowledge about the geometrical structure of the vector eld involved. The MoM produces a dense matrix problem, see Chapter 3 for detail, limiting its use to electrically small problems. In such a context, our goal is to provide a solution when the MoM is no longer an option and when conventional high frequency techniques fail. In this niche, our new approach, detailed in Paper 2, is fast, easy to implement and represent a platform for future development of high frequency approximations.

1.2

Background: the GEMS project

All the numerical methods implemented are now part of the industrial software suite called GEMS: General ElectroMagnetic Solvers. GEMS has been developed in cooperation with Uppsala University, Chalmers University, Ericsson Microwave Systems AB and Saab AB (Aerotech Telub) during the time period 1998 - 2005, making the GEMS project the largest Swedish CEM project ever. In the Swedish industry, GEMS is used by Aerotech Telub and by the Swedish Defense Research Agency (FOI) to simulate the radar performance of advanced stealth aircraft designs. An example of such a design, the unmanned aerial vehicle code-named Eikon, is illustrated in Paper 3. The software suite now incorporates a large choice of numerical tools for various CEM applications both in time and frequency domain. In the latter, the methods are ecient for the simulation of single frequency but inecient for the simulation of broadband phenomena. Large electrical size problems can be solved using high frequency approximations. Time domain methods have the exact opposite property, they are ecient over broad bands but can not eciently handle high frequency. In the time domain, GEMS species Finite Elements, Finite Volumes and Finite Dierences, see [Taove 95], as well as hybrid methods between these [Edelvik 02, Andersson 01, Ledfelt 01, Abenius 04] . In the frequency domain GEMS species Method of Moments, Fast Multipole Method, Geometrical Theory of Diraction,

1.2. BACKGROUND: THE GEMS PROJECT

Physical Optics and Surface Current Extraction Extrapolation, as well as hybrid methods between these [Nilsson 04, Edlund 01, Hagdahl 05, Se 03]. The software is implemented for a number of dierent computer platforms, Linux, Sun, IBM, SGI and others, both in serial and parallel versions and mainly written in the Fortran 90 and MPI programming language. The version control system CVS [1] is used to keep track of the changes. We use the netCDF [2] format for output data which allows visualization in Matlab [3] and OpenDX [4]. The geometries are generated by the CAD software CADx [5]. A more detailed description of the GEMS software suite can be found in [Oppelstrup 99].

Chapter 2

Governing Field Equations

In this chapter we describe how the Maxwell equations can be derived from the laws of physics that govern electric and magnetic phenomena, such as Coulombs law and Faradays law. We will see that electromagnetic waves are created by time-varying currents and charges governed by the Maxwell equations. We will demonstrate that it is possible to write down the elds at a frequency > 0 E = iA as solutions to the Maxwell equations in form of a complex sum of a magnetic vector potential A and the gradient of a scalar potential . These two potentials will be found as solutions to vector dierential Helmholtz equations, which will be derived on the surface of the conductors from the Maxwell equations given a sensible choice of boundary conditions. The vector potential A explains mostly the variations in the magnetic eld, whereas the scalar potential can be linked to the variation of charge. Such analogy, when applied to the tangential component of the magnetic eld on the surface - the current, will come in handy in Paper 1 and Paper 2 to explain, to decompose and to predict how the surface currents behave. First, lets introduce the basic concepts necessary for the derivation of the equations. Identities for Zero Divergence and Zero Curl of a Vector Field In the derivation of the equations for the elds, the following identities for the nulls of divergence and curl, as well as Helmholtzs theorem will be useful tools. Identity 1 The curl of the gradient of any scalar eld is identically zero, () 0 It is also true that if the curl of a vector eld B is zero, it can be written as the gradient of a scalar eld , if B 0 , B = . This type of eld is called conservative. This is valid in any system of coordinates. Thus we say that a conservative eld can always be written as the gradient of a scalar eld. 7

CHAPTER 2. GOVERNING FIELD EQUATIONS

Identity 2 The divergence of the curl of any vector eld is identically zero, ( A) 0 It follows that if the divergence of a vector eld B is zero, it can be written as the curl of another vector eld A, if B 0 A, B = A. This type of eld is called solenoidal. Helmholtzs Theorem 2.1 A general vector eld v, which vanishes at innity, can be represented as the sum of two independent vector elds; one that is conservative (zero curl) and another which is solenoidal (zero divergence), v = A In other words, a general vector eld which is zero at innity is completely specied once its divergence and curl are given.

2.1

The Maxwell Equations

This section results from [Stratton 41], to which the reader can return for more details. The Maxwell equations involve linear operations on vectorial quantities E, D, H, B, functions of position (r IR3 ) and time (t IR). The vector functions D and B will later be eliminated from the description of the electromagnetic elds via suitable constitutive relations. For now, we have Faradays law E = B t D = v Gauss law for electric elds H B = J + = 0
D t

Amperes law Gauss law for magnetic elds

where in SI units, E H D B J v

Electric eld strength (or intensity) Magnetic eld strength Electric ux density (or displacement) Magnetic ux density Current density Volume charge density Divergence operator Curl operator

[V olt/m] [Ampere/m] [Coulomb/m2 ] [W eber/m2 ] [Ampere/m2 ] [Coulomb/m3 ]

The rst equation is called Faradays law and expresses how changing magnetic elds produce electric elds. The rst divergence condition is Gausss law and describes how electric charges produce electric elds. The distribution of charges is

2.1. THE MAXWELL EQUATIONS

given by the scalar charge density function v . The next equation is Amperes law as modied by Maxwell. It describes how currents and changing electric elds which act like currents, produce magnetic elds. Finally the second divergence equation expresses the experimental absence of magnetic charges (magnetic monopoles do not exist) resulting in the fact that the magnetic ux is solenoidal. We can simplify the notations by considering time-harmonic elds where the time-varying elds E oscillating at temporal frequency = 2f , are replaced by their corresponding vector phasor E(r, t) = Re[eit E(r)]. This leads to the timeharmonic Maxwell equations for single frequency environments, E = iB, D = H = J iD B = 0

The continuity equations for the current The two divergence conditions (the two Gauss laws) in the Maxwell equations are consequences of the fundamental curl eld equations provided charge is conserved. This is shown by taking the divergence of the two curl equations (Faradays and Amperes law), recalling the null identity for any vector eld A, ( A) 0, and by introducing the physical law that relate the density of current J to the distribution of charge v , J = v t Conservation of charge. (2.1)

It is then enough to study the two curl equations augmented with the continuity equation Eq. 2.1, see [Stratton 41], in the derivations of the theorems. Our motivation is to use as few concepts as possible to describe the elds. In addition, we will see that the two curl equations can be written as one second order equation. The constitutive equations for the medium The Maxwell equations must be augmented by two constitutive laws that relate E and H to D and B respectively, E = D and B = H (2.2)

where we denote the electric permittivity = r 0 and the magnetic permeability = r 0 . These variables depend on the properties of the matter in the domain occupied by the electromagnetic elds. 0 and 0 are constants, 0 = 8.854 1012 Farads/m and 0 = 4 107 Newton/Ampere2 . r and r are dimensionless scalar functions characterizing a medium. Vacuum, free space, air and material with no magnetic or no electric properties have r = 1 and r = 1. As a result, the speed of the wave in such conditions is given by c = 1/ 0 0 . In GEMS, frequency dispersive materials with () and metals with r = 1 which behave nearly like PECs at radio frequency are of main interest. However, ferromagnetic materials with r 1, e.g. magnets, require dierent models.

10

CHAPTER 2. GOVERNING FIELD EQUATIONS

E H k

Figure 2.1: Scattering conguration of a plane wave inducing surface currents on a perfect conductor .

Before any attempt to write a solution, boundary conditions for both electric and magnetic elds on the bounding surface should be specied to form a complete boundary value problem. In particular we are interested in the situation illustrated in Figure 2.1, where a plane wave travelling in a source-free medium encounters a PEC scatterer.

2.2

Boundary conditions on a Perfect Conductor

When the space near a set of charges contains electric material, the electric eld no longer has the same form as in vacuum. The interactions between the charges and the electric eld obey the boundary conditions of the Maxwell equations. Together with the constitutive relations, Eq. 2.2, the Maxwell equations, in presence of a

2.2. BOUNDARY CONDITIONS ON A PERFECT CONDUCTOR prefect conductor , can now be restated as a complete system of equations, in R3 E = iH E = in R3 H = J iE in R3 H = 0 in R3 nE nE nH nH = 0 = s on on

11

= Js = 0

on on

where n is the normal to the surface S = of the PEC. The expression n E = s represents the component of E along the normal n, where s is the surface charge density. The tangential component of E is given by n E = 0. In other words, the tangential E-eld must go to zero on the surface of a conductor. The normal component of the magnetic eld is given by n H = 0 and the surface currents are the tangential component of the magnetic eld n H = Js . Other boundary conditions, for instance for dielectric media, are beyond the scope of this thesis. At large distance, it is assumed that only outgoing waves are present and that the elds are bounded when r . The energy carried by the eld propagates along the Poynting vector k, k= 1 EH 0 (2.3)

which has the dimension of a power density, watts per square meter. This represents the direction of the GTD rays described in Paper 5. We now introduce the wavenumber = |k| which satises the following relation in terms of frequency and wavelength , = 2 = . c (2.4)

The second order Maxwell equation The two curl equations for the electric and magnetic elds E = iH H = J iE

are rst order in spatial derivatives. They can be combined into one second order equation as ( E) 2 E = iJ (2.5)

thus removing H. Eq. 2.5 will be used later for the derivation of Helmholtzs equations.

12

CHAPTER 2. GOVERNING FIELD EQUATIONS

2.3

Plane wave solution in free space

In a source-free medium, the divergence of the electric eld is zero, E = 0. Using ( E) = 2 E + ( E) in R3 , the second order Maxwell system Eq. 2.5 in a source-free medium, ( E) 2 E = 0, simplies to 2 E + 2 E = 0 in R3 . (2.8) (2.7) (2.6)

Eq. 2.8 is established as Helmholtzs equation (homogeneous, source-free) and is also called the wave equation. A possible solution to Helmholtzs equation is a plane wave. E(r) = E0 eikr (2.9)

The constant term E0 must be orthogonal to k. Hence the concept of polarization sets the actual direction of E. The associated magnetic eld for a plane wave is obtained from Faradays law H(r) = where Z is the wave impedance Z = Zvacuum ; Zvacuum = 0 = 377 0 (2.11) 1 k E0 (r) Z (2.10)

2.4

The vector Helmholtz Equation

As mentioned in section 2.1, the two curl equations suce together with the continuity equation. In the previous section we have seen that the two curl equations can be replaced by one second order curl equation in E. In this section we demonstrate that the electric eld E is completely determined by one magnetic vector potential A together with the gradient of the scalar potential . We show that these two quantities are derived from the second order Maxwell equation and, when they fulll the Lorenz conditions, they satisfy two inhomogeneous Helmholtz equations. First lets decompose E in the following complex expression, E = iA (2.12)

2.4. THE VECTOR HELMHOLTZ EQUATION

13

From the null identity of the curl, such that E iA = , that we restate E = iA .

Proof: We start with E = iB. Since the magnetic ux is solenoidal B = 0. Then A such that B = A, thus E = i A, that we restate (E iA) = 0.

Such decomposition already appears in [Stratton 41]. However, the authors do not give any physical interpretation, using it only as a convenient formulation to derive equations. Note that in Paper 1 and Paper 2, a similar decomposition has been applied to the surface current vector eld. Lets now express the quantities A and as functions of the surface currents. To do so, we start from the second order Maxwell system Eq. 2.5 ( E) 2 E = iJ. We insert the electric eld dened in Eq. 2.12 into Eq. 2.5 to obtain (iA ) 2 (iA ) = iJ. Because of the null identity () 0, this reduces to Imposing the condition, = i A 2 (Lorenz Gauge), (2.15) i ( A) 2 iA + 2 = iJ (2.13)

(2.14)

so that the gradient 2 = i( A), we obtain

i ( A) i( A) 2 iA = iJ, ( A) ( A) = 2 A.

(2.16)

which allows us to simplify the expression using the vector triple product

Thus A satises the dierential equation of the vector potential, 2 A + 2 A = J a Helmholtz equation with source term J. Taking the divergence and applying the continuity equation J = i with 2 / 2 = , 2 + 2 = (2.19) we see that also satises a Helmholtz equation but with source term . In the next sections, the solutions A and of Eq.2.17 and Eq.2.19 respectively, will be expressed as integral equations using Greens function. 2 ( A) + 2 A = J (2.18) (2.17)

14

CHAPTER 2. GOVERNING FIELD EQUATIONS

2.5

Greens function

To get the solution of Helmholtzs equation with a source term, we introduce an auxiliary function called Greens function. First, let r (r) = (r r ) denote the Dirac delta function such that for any continuous function f we have
+

f (r)r (r)dr =
rr

f (r)r (r)dr = f (r ).

(2.20)

2 G + 2 G is zero everywhere, except possibly at r = r . For r = r we can express the solution G(r, r ) of Eq. 2.21 as an outward spherical wave of origin r , ei|rr | G(r, r ) = 4|r r |

This has to be understood as the result of a limit-process, where the source is distributed over a nite domain which shrinkes until, in the limit, the entire source is applied at the point r = r . Greens function represents the inverse of the dierential Helmholtz operator [2 + 2 ] which has the Dirac delta function as source term. 2 G(r, r ) + 2 G(r, r ) = (r r ) (2.21)

(2.22)

Greens method: Let be a function with bounded support, then the solution (r) to the general Helmholtz equation 2 (r) + 2 (r) = (r); is given by (r) =

0 as r

(2.23)

G(r, r )(r )d

(2.24)

Consequently, to get the solution to our problem, we multiply G(r, r ) by the source function and integrate over the region of interest. Far eld approximation: If we introduce the unit vector = r/|r| pointing r from the origin of the coordinates to the observation point, the distance |r r | becomes |r r | = r2 + r2 2r r = |r| 1 + [r2 /|r|2 2 r /|r|] r |r r | |r|(1 + [r2 /|r|2 2 r /|r|]/2) r Through the above Taylor expansion, the term in r2 becomes negligable, |r r | |r| r r (2.27) (2.25)

which can be approximated with respect to r2 and in the limit as r using (2.26)

2.6. THE FIELD INTEGRAL EQUATIONS Greens function in the far eld can now be approximated as G(r, r ) ei|r| ir e r 4|r|

15

(2.28)

This approximation will be used later to derive the PO integral studied in Paper 3. The criterion for far eld is that the distance R = |r r | between source point and observation point, is much larger than both the size of the scattering object and the wavelength .

2.6

The eld integral equations

E = iA

In previous sections we have seen that the eld E can be decomposed into where A satises Then the electric eld E can be computed from the surface currents by using Greens method applied to A and . We nd the solution for the magnetic vector potential A as A(r) = J(r )G(r, r )dS (2.29) 2 + 2 =
S

and satises

2 A + 2 A = J

and, since =

1 i

J , the scalar potential as = 1 i

S

( J(r ))G(r, r )dS

(2.30)

where S denotes the surface of the PEC with surface currents J and denotes the vector derivative with respect to r . With w = Z and w = Z 1 , we obtain the integral formulation for the electric eld, iZ ( J(r ))G(r, r )dS , (2.31) E = iZ J(r )G(r, r )dS S S and the integral formulation for the magnetic eld, H = iZ 1
S

J(r )G(r, r )dS +

iZ 1

( J(r ))G(r, r )dS . ei|rr | dS , 4|r r |

(2.32)

Greens function Eq. 2.22 substituted into Eq. 2.31 produce E = iZ

S

J(r )

which only depends on the surface currents J. For further readings on the subject see [Bendali 99]. Now it remains to nd an integral equation for the unknown surface currents J.

ei|rr | iZ dS | 4|r r

( J(r ))

(2.33)

16

CHAPTER 2. GOVERNING FIELD EQUATIONS We denote the tangential compon(2.34)

Electric Field Integral Equation (EFIE) ent of the total eld on the surface, Etan ,

Etan = E (n E)n = n E.

If imposing the boundary condition that the total tangential electric eld vanishes on the surface, see Section 2.2, written as Etan = Einc + Escat = 0 tan tan (2.35)

where we dene the incident electric eld E inc as the eld that would exist in the absence of the scatterer. If assuming that the scattered eld E scat also can be obtained from the solutions to Maxwells equations, see [Hodges 97], and written as Escat (r) = iA(r) (r), (2.36) then we obtain the famous EFIE (Electric Field Integral Equation) for a PEC as [iA(r) (r)]tan = Einc . tan iZ 2 (2.37)

We plug the expression for A, Eq. 2.29, and , Eq. 2.30, into Eq. 2.37 and get [iZ
S

J(r )G(r, r )dS

( J(r ))G(r, r )dS ]tan = Einc . tan

(2.38)

In the next chapter we will see how this integral equation can be solved numerically using the Method of Moments (MoM) with the surface currents as unknown. Once the surface currents are determined the scattered eld can be found as Escat = n tan [iZJ(r )G(r, r ) +
S

1 J(r ) G(r, r )]dS . iZ 1

(2.39)

Magnetic Field Integral Equation (MFIE) Using the same procedure applied to the magnetic eld with the boundary condition of the magnetic eld n H = J, see Section 2.2, the MFIE (Magnetic Field Integral Equation) is obtained, J(r) (2.40) n J(r ) G(r, r )dS = n Hinc (r). 2 S where is called the principal value integral with regions where the source and eld points coincide have been excluded. The EFIE and the MFIE can be used either combined or separatly to solve for the currents, except for frequencies corresponding to interior body resonaces. As it will be shown in the next chapter, if one ignores the eld expressed by the integral term in the MFIE, one obtains the Physical Optics assumption, that is the current is given by twice the tangential component of the incident magnetic eld J(r) = 2n Hinc , without the need to solve an integral equation.

2.6. THE FIELD INTEGRAL EQUATIONS

17

Far eld In the far eld, Greens function is approximated by Eq. 2.28 and plugged into Eq. 2.33 where the divergence can be placed outside the integral using the denition of the Lorenz Gauge combined with Eq.2.17 and Eq.2.19 E = iZ[1 which can be written as E = iZ[1 ei|r| 1 ] 2 4|r|
r J(r ) eir dS .

1 ] 2

J(r )

r ei(|r|.r ) dS , 4|r|

(2.41)

(2.42)

It is now suitable to simplify the notation by introducing the vector eld dened by i2 Z r J(r ) eir dS . (2.43) K(r) = 4 S Then the electric eld is written as E = [1 1 ei|r| ]( K(r)). 2 |r| (2.44)

1 Under approximation in the far zone, the operator 2 in Eq. 2.44 is approximated using the fact that components that vanish faster than 1/|r| are negligible, see [Rumsey 54], such that the dominating contribution to the scattered electric eld becomes ei|r| Escat = [K(r) (K(r) )]. r r (2.45) |r|

Using the triple vector identity, we nally get Escat = ei|r| [ (K(r) )] r r |r| (2.46)

which is called the far eld radiation integral. This expression is a function of the direction r to the observation point and of the surface currents only and the expression [ (K(r) )] represents the far eld amplitude of the wave. Finally, r r the RCS is then computed as = 4 scat 2 |E | 2 (2.47)

The magnetic scattered eld is found by using Faradays law E = iZH by neglecting the components that vanish faster than 1/|r| in the evaluation of the curl operator, ei|r| [ (K(r) )]. r r r (2.48) Hscat Z 1 |r|

Chapter 3

Scattering Analysis Methods

3.1 Geometrical Theory of Diraction

Exact solutions to the Maxwell equations are known for canonical scattering geometries such as an innite cylinder, sphere or cone. For complicated geometries, approximate methods can be derived from these analytical solutions [Kouyoumjian 65]. In the limit of vanishing wavelength ( 0), a widespread approximate method is the Geometrical Theory of Diraction (GTD) [Keller 62]. It is based on the fact that diraction phenomena exhibit local properties at high frequency. As a consequence, the scattered eld does not depend on the interactions from all points on the surface of the conductor, but rather on the contributions from few points located in the neighborhood of some special positions called diraction points. Under such assumptions, the complex scattered eld is analytically approximated with a superposition of known solutions from simple canonical scattering geometries. The total scatter solution is expanded, see [McNamara 89], as

Escat (r) = ei(r)

n=0

(i)n En (r)

(3.1)

where (r) is the optical distance from the source point r and the amplitude En (r) is independent of the wavelength . Furthermore, GTD postulates that energy propagates along direct, reected and diracted rays. This enables the use of the ray tracing algorithms introduced in Paper 5 and further detailed in the authors Licentiate thesis [Se 03] and in [Catedra 98]. These algorithms draw (trace) the various dominant propagation paths between sources and receivers interacting with the surface of the conductor, describing the following asymptotic phenomena: Direct eld Reected eld Diracted eld 19

20

CHAPTER 3. SCATTERING ANALYSIS METHODS Multiple elds, e.g. double reected, diracted-reected, ...

The ray trajectories are determined independently and the computational cost depends on the geometry of the obstacle instead of the electrical size. Despite these nice features, the GTD has a few drawbacks which may sometimes reduce the usefulness of the results. First, there may be many higher order multiple ray interactions which involved combinations of diractions, increasing the complexity of the ray tracing task. This has been addressed in Paper 6. Second, the accuracy of the calculated eld is relatively low since the theory will only yield the leading terms in the asymptotic high frequency solution of the Maxwell equations.

3.2

Physical Optics

The Physical Optics (PO) technique is also a well-known and widely used high frequency approximate technique for the calculation of the electromagnetic eld scattered from a PEC illuminated by an incident electromagnetic eld. It is a very fast technique since it does not require an integral equation to be solved. The idea is to approximate the surface currents as if they were obtained on an innite at plate. Physically it can be interpreted as replacing locally the conductor by a at plate and neglecting the contributions from sharp edges, corners and all mutual interactions such as multiple reections or creeping waves, what remains being the main reection. In order to derive the PO current, we rst need to have a look at the elds which induce it. As previously mentioned in Chapter 2, the total eld E can be decomposed in a sum of incident and scattered eld ETotal = EInc + EScat . (3.2)

The boundary conditions on a PEC impose that the tangential component of ETotal vanishes n ETotal = 0 on , (3.3) which means n EScat = n EInc . (3.4) Consecutively, the tangential component of the electric elds ips on reection, as illustrated in Figure 3.1. Figure 3.1 displays the two generic cases, the well-known TM- and TE-case, characterized by magnetic (or electric) incident elds transverse to the plane of incidence, spanned by k and n. In the TEcase, there is no component normal to the surface, n ETotal = 0, n HTotal = 0 (3.5) which means that there are no charges on the surface (s = 0). The normal component of the incident and scattered magnetic elds cancel each other. Due to

3.2. PHYSICAL OPTICS

21

E H
k

E H E H Etan =0 H

E
k n

H E

TM case: tangent. comp. E flips total tangent H double

TE case: normal comp. H flips total tangent H double tangent. comp. E flips

Figure 3.1: TE, TM: case on a plate.

the dynamics in the Maxwell equations, what is lost in the normal component is gained by the tangential component making the tangential component for the total magnetic eld double. Thus on the surface where Js = n HTotal we get Js = 2n Hinc (3.6)

In the TMcase, on the surface, the total Eeld has only a component normal to the surface, n ETotal = s (3.7) which means that we have a maximal transverse charge. The tangential component of the incident and scattered magnetic elds cancel each other making the tangential component of the total magnetic eld to double, getting once more Eq. 3.6 for the current. On an innite at plate there is no concentration of charge J = 0 in the lit region. On the shadow side of an innite plate, no eld can penetrate resulting in no current. This leads to the following surface currents referred to as the PO currents: 2n Hinc (lit region) (3.8) JP O = 0 (shadow region) The vector eld of the PO currents has zero divergence making it solenoidal. The next step is to obtain the far eld scattered eld induced by JP O . From Eq. 2.46, we directly get Escat = iZeiR [ r 4R
r JP O (r )eir dS )] r

Escat =

iZeiR [ r 4R

JP O (r )ei(k
S

scat kinc )r

dS ] r

E=0

(3.9)

(3.10)

22
1.5

CHAPTER 3. SCATTERING ANALYSIS METHODS

1.5

0.5

0.5

0.5

0.5

1.5 1

0.5

0.5 x

1.5

2.5

1.5 1

0.5

0.5 x

1.5

2.5

Figure 3.2: Direction of PO currents on a sphere viewed from the top.

Figure 3.3: Direction of PO currents on a sphere viewed from the side.

For a nite plate and for even moderately high frequencies JP O provides a reasonably good approximation of the surface currents, at least away from diraction and creeping wave phenomena generated by the sharp edges of the plate or when grazing incidence occurs. In the case of monostatic direction, we have kinc = kscat thus Eq. 3.10 simplies to Escat = iZeiR inc [k 4R [2 Hinc ]ei2k n
inc r

dS kinc ].

(3.11)
1 inc Einc ), 0 Zk

Using the associated magnetic eld for a plane wave Eq. 2.10 (Hinc = we obtain i2eiR inc inc E0 [kinc n]ei2k r dS . Escat = 4R S

(3.12)

This is the PO integral which has been further studied in Paper 3. In particular we observe that the amplitude of the integrand is proportional to cos( , kinc ) which n is a slowly varying function over the surface whereas the phase is an exponential function rapidly varying. We will come back to this point later. For a smoothly curved perfect conducting body, the PO current is an amazingly good approximation of the induced surface currents, away from shadow boundaries and as long as the radius of curvature is larger than the wavelength (high frequency). The physical origin of the PO currents is clear the magnetic eld only. In Figure 3.2 and Figure 3.3 we see how the direction of the PO current on a sphere is everywhere perpendicular to the magnetic eld. The properties of the direction of the PO currents have been further detailed in Paper 1 and Paper 2 leading to the concept of PO streamline. A comparison between the RCS of a sphere obtained with PO, and the analytic solution for the RCS of a sphere, see Mie solution in [Bowman 87], can be seen in Figure 3.5.

3.3. METHOD OF MOMENTS

23

Concerning the implementation, two major diculties need to be treated. First, the detection of the shadow regions for complex structures has to be taken care of. Shadowing algorithms for PO has been introduced in Paper 6 and a fast algorithm based on ray tracing has been proposed in the authors Licentiate thesis [Se 03]. Other known solutions consist in processing the surfaces with a graphical engine. In [Rius 93] and [Asvestas 95], an image of the surfaces on the computer screen is used and the shadowing is processed eciently by the hardware. A second diculty is to solve numerically the PO integral, in particular the treatment of the oscillating phase factor ei2kr . There exist many dierent ways which more or less can be grouped into three main basic concepts: The most classic way is to evaluated the PO integral using quadrature formulas, for example Gauss quadrature. This requires a number of evaluation points, proportional to the wave number squared. In Gordons method [Gordon 94], the surface integral on special geometries can be converted into simpler single integrals. Approximation methods such as stationary phase or linear phase approximation, produce simplied expressions for the integral which can then be solved analytically. In the stationary phase methods [Catedra 95], an approximation to the integral is found at some points where the phase is stationary. The problem then reduces to determine the positions of all the stationary phase points using complex minimization techniques similar to the ones used in GTD. In the linear phase approximation [Ludwig 68, Moreira 94], polynomial approximations to the phase is used, see Paper 3. Note that partial quadratic phase approximations which do not include any mixed terms have been tried in [Crabtree 91].

3.3

Method of Moments

This section explains the theoretical foundations of the very popular and the most accurate method to solve the integral equation for the surface current J, namely the Method of Moments [Harrington 68]. To do so, the starting point is to use the EFIE Eq. 2.38 and to approximate J by the expansion J(r ) =
q

jq fq (r )

(3.13)

where fq (r ) are chosen vector basis functions and jq are scalar coecients to be determined. A popular choice of linear basis functions is the Rao-Wilton-Glisson (RWG) basis functions [Rao 82], which are dened as lq + + 2A+ + , if r Tq q q l (3.14) fq (r) = q , if r Tq 2A q q 0, otherwize

24

CHAPTER 3. SCATTERING ANALYSIS METHODS

where A denote the area of the triangles Tq , lq the length of the qth edge and q q = (rq r) the vector from the free vertices opposite to the qth edge, r , see q Figure 3.4. The basis functions represent the current owing through the qth edge

n+

T+

ge qth ed

+ q

T-

+ rq

rq

Figure 3.4: Rao-Wilton-Glisson (RWG) basis function.

+ from the triangle Tq to Tq . A nice property is that the divergence of the basis function across the edge is a constant fq =

lq , A q

(3.15)

so no charge can be accumulated along the edge. Multiplying the EFIE with a test current fp and integrating results in a set of linear equations [A]{J} = {V } where the elements of the excitation vector {V } are given by Vp =
S

(3.16)

Einc (r) fp (r)dS

(3.17)

and {J} is the column of the unknown coecients jq . In Paper 2, the basis function coecients have been used as direct input for extrapolation to high frequency. The

3.4. HYBRID METHODS matrix [A] is called the impedance matrix whose entries are given by Apq = iZ
S S

25

fp (r)G(r, r ) fq (r )dSdS ,

(3.18)

where

1 G(r, r ) = [1 2 ( )]G(r, r )

(3.19)

3.4

Hybrid Methods

The MoM requires a number of unknowns N inversely proportional to the square of the wavelength 2 which makes the size of the impedance matrix unmanageable at high frequency. Solving the system in Eq.(3.16) with Gaussian elimination requires O(6 ) arithmetic operations and O(4 ) for storage. To overcome this diculty, many types of hybrid approaches have been proposed. These methods can be classied into four groups. First the methods which focus on obtaining an approximate sparser impedance matrix using domain decomposition between MoM and PO or MoM and GTD [Thiele 75, Burnside 75, Bouche 93]. The idea is to split the large smooth geometry parts away from small details modelled by MoM. This leads to the popular MoM-PO hybrid solvers [Rahmat-Samii 91, Jakobus 95, Hodges 97, Taboada 00, Edlund 01]. In our experience, the main weakness of these methods is that of displaying large errors where PO fails [Burnside 87], i.e. in shadow regions or near edge discontinuities. Second the methods which also aim to avoid expensive linear algebra but using instead ecient matrix algorithms such as Multi-grid [Sarkar 02], Wavelet compression [Leviatan 93] or Fast Multipole Method (FMM) [Greengard 87, Rokhlin 92]. FMM which is the most successful, uses block decomposition [Greengard 87] where the far eld elements are regrouped into small low rank blocks [Nilsson 00]. However, the technique still needs to represent the fastest variation of the phase thus requiring the same large number of unknowns as in MoM. In the third group, we have the methods which reduce the total number of unknowns [Taboada 01, Taboada 05]. They typically require dicult alternative basis functions [Djordjevic 04]. More References as well as descriptions of related work can be found in the introduction of Paper 2. Finally, the last group of methods is the Extraction-Extrapolation methods. They model the behavior of the surface currents by looking at currents obtained at low frequency [Mittra 94, Aberegg 95] and then extrapolate them to higher frequencies [Altman 96, Altman 99], The latest development in this direction, see next section for more details, is the Asymptotic Phasefront Extraction (APE) technique [Kwon 00, Kwon 01], which gives a propagating ray-based description analogue to GTD of the surface currents. APE-MoM Phasefronts In [Kwon 01], the goal is to reduce the total number of unknowns in the MoM procedure by approximating the induced surface current

26

CHAPTER 3. SCATTERING ANALYSIS METHODS

over large smooth regions with one or several linear phase currents, J(r ) = Ceik
m

,
m

(3.20)

where C is a complex vector amplitude assumed to be constant and k is the direction of a phasefront travelling on the surface. At each point r on the surface, a local low frequency phase variation is split into several phasefronts of m dierent directions. The phasefront extraction is achieved using a multi-dimensional complex Fourier transform j(ku , kv ) on a rectangular grid of the low frequency complex current as inc j(ku , kv ) = J(r )eik r dS , (3.21)
S

where ku , kv are the components of the projection of kinc on the surface S and J(r ) is obtained from a low frequency MoM solution. The dominating components in the Fourier space are found using a peak-searching procedure for local maximum. Given m extracted phasefront directions, the linear phase currents are then expanded as
Km

Jh (r ) =
Nh

jh fh (r ) +
Nl p

jl fl eikh

m,p

(r rm )

(3.22)

where the subscript l indicates elements evaluated over the low frequency sparse grid and h over the high frequency dense grid away from smooth regions. On the latter grid, which ideally should correspond to small regions, the basis functions from MoM are used as usual. Over the large smooth regions, the low frequency basis functions fl are multiplied by the built-in phase propagation factor to form the high frequency basis functions. The modied basis functions inherit the quick variation in the exponential to speed up the surface integration. The wavenumber is scaled linearly to the right frequency such that km,p = h low m,p k , high l p = 1, ..., Km (3.23)

The high frequency approximate current is then plugged into the MoM formulation with fewer unknowns than the conventional MoM requires. A shortcoming of this technique is that it still requires computationally expensive numerical integrations when solving the matrix system. Thus, one would like to devise a technique which does not require a second MoM system to be built and solved. In Paper 1 and Paper 2, this has been achieved by constructing high frequency currents with behaviors obtained from current-based methods only (MoM and PO), instead of using a propagating ray-based description of the surface currents.

3.5

Modelling the behavior of the surface currents

Lets apply the Physical Optics and the Method of Moments techniques described in the previous sections to the typical scattering problem of a metallic sphere centered

3.5. MODELLING THE BEHAVIOR OF THE SURFACE CURRENTS

Normalized monostatic RCS of a PECsphere, r=1m

27

15 10 5 0 5

[dBsm]

10 15 20 25 30 35 0

PO Mie serie 100 terms (Exact) MoMQMR solver

50

100

150

200

250
f [MHz]

300

350

400

450

500

Figure 3.5: Monostatic RCS of a one meter sphere using MoM and PO, compared to the analytical Mie solution.

at the point (0,0,0), illuminated by a plane wave propagating in (0,0,-1) direction. For a sphere, the numerical methods can be compared to an analytic solution known as the Mie-series. In Figure 3.5, we display the monostatic Radar Cross Section given by the Mieseries along with the solutions obtained with the Physical Optics and the Method of Moments for various frequencies ranging from 1MHz to 500MHz. Three observations can be made: (i) The MoM breaks down due to poor resolution when the wavelength increases. This happens around f=250MHz as expected since the number of unknowns is kept xed. (ii) For low frequency, all three methods give similar results. The sphere is too small, compared to the wavelength, to be seen. (iii) For high frequency, the PO converges to the Mie-series. The radius of curvature becomes much bigger than the wavelength, getting closer to the conguration of a large plate.

28

CHAPTER 3. SCATTERING ANALYSIS METHODS

Surface Currents on a Sphere, ka =0.1 1 1

Surface Currents on a Sphere, ka =1

0.8

0.8

0.6

0.6

0.4

0.4

0.2

0.2

0H

0H

0.2

0.2

0.4

0.4

0.6

0.6

0.8

0.8

0.8

0.6

0.4

0.2

K 0 y

0.2

0.4

0.6

0.8

0.8

0.6

0.4

0.2

K 0 y

0.2

0.4

0.6

0.8

Figure 3.6: Directions of the surface currents obtained using MoM on a 1m sphere viewed from the side, illuminated by a plane wave coming from the top, for k=0.1 and k=1.0. Note that the triangulation is transparent, thus we see currents on the back and on the front currents simultaneously.

A closer look at the real part1 of the surface currents obtained from the MoM at low frequency is provided in Figure 3.6. In agreement with the observation (ii) that the MoM and the PO solutions match, the induced MoM current vectors computed at the frequency k = 0.1 are aligned like the PO current vectors. Increasing the frequency to k = 1.0, we observe that the direction of the MoM current vectors deviate from their PO positions. They tend to point towards the South pole as well as getting smaller in the shadow zone. Our motivation in Paper 2 is to model this deviation as well as to predict its behavior at higher frequency. One clue about the behavior of the deviation is given in Figure 3.7, which displays the components of one particular vector on the sphere as a function of the frequency. We observe that each spatial component displays systematic sinus patterns evolving smoothly over frequency, until the MoM breaks down. Such patterns describe in fact a rotation of the surface current vector with frequency. The rotation of the vector takes place in the tangent plane of the surface and, according to observation (ii), begins at low frequency with a vector aligned to the PO currents. In addition, we know from the structure of the incident magnetic eld, that the PO currents run along level curves, referred to in Paper 1 and Paper 2 as PO streamlines. Figure 3.8 displays the PO streamlines on a sphere for two dierent incident illuminations. Thus, in order to estimate the dierence between the total
1 The

imaginary part can be treated analogously at a quarter of period later.

3.5. MODELLING THE BEHAVIOR OF THE SURFACE CURRENTS

3

29

x 10

Components of current Real(J) on triangle1 real(Jx) real(Jy) real(Jz)

2 Real(Jx), Real(Jy), Real(Jz)

50

100

150

200

250 f [MHz]

300

350

400

450

500

Figure 3.7: Variations of the components of one current vector with respect to frequency computed with the MoM. When the frequency gets too large, the MoM breaks down requiring more elements per wavelength. Also, frequencies corresponding to interior body resonaces can be observed when all the three components the become singular.

and PO currents, it makes sense to decompose the total current into one component along the PO streamline and one component perpendicular to the streamline, as illustrated in Figure 3.9 and further described in Paper 1. More details are given in the next Chapter and in Paper 1 and Paper 2.

30

CHAPTER 3. SCATTERING ANALYSIS METHODS

H 1 K

E
1 0.8

0.5

0.6 0.4

0.2

0
z 0 0.2

0.5

0.4 0.6 0.8 1 1

1 0.5 0 0.5 1 x 1 y 0.5 0 0.5

0.5 0 0.5 1 1 y x 0.5 0

0.5

Figure 3.8: PO streamlines on a 1m sphere, illuminated from the top with the magnetic eld in y-direction (on the left), illuminated from the side with the magnetic eld in z-direction (on the right).

Jcross

Jmom

Jpo

Figure 3.9: Decomposition of the surface currents along a PO streamline.

Chapter 4

Summary of the Papers

4.1 Paper 1: Modelling and Extrapolation, an Extended Abstract

In this paper we introduce the decomposition of the surface currents J, expressed in the triangular-type Rao-Wilton-Glisson (RWG) vector basis functions in the MoM formulation, into parallel J|| and perpendicular J vector components, denoted nxH-parallel currents and cross-currents respectively, relative to the direction of the incident magnetic eld on the surface. These two currents yield systematic spatial patterns evolving over frequency in close relation to the incident magnetic eld, both on the illuminated and the shadow side of the body. From these observations we derive two reference models for the scalar components J|| and J and compare them to MoM for various frequencies. The author of the thesis performed the numerical experiments and presented it at the ICNAAM 2005 International Conference of Numerical Analysis and Applied Mathematics, Rhodes, Greece, September 2005. The theoretical derivation as well as the implementation has been made by both authors in cooperation. The idea of the reference model for the parallel component and the fact that the incident magnetic eld can be used both on the illuminated and the shadow side come from the rst author. The idea for the model of the perpendicular component comes from the second author.

4.2

Paper 2: Modelling and Extrapolation, Continued

In this paper the analysis presented in Paper 1 is extended. The vector eld decomposition is not standard and yields a better understanding of the structure of the surface currents as well as the underlying physics governing their spatial variations, i.e. charge transport and charge accumulation. The oscillating behavior of the current is strongly linked to that of the incident magnetic eld. The reference models for the components J|| and J of Paper 1 are modied by coecients 31

32

CHAPTER 4. SUMMARY OF THE PAPERS

calculated from MoM at low frequency. In the neighborhood of the vertices of the triangles, the MoM current yield unreliable values. To remedy this, corrections and smoothing have been implemented. Trends in the modied coecients have been studied and have allowed us to better extrapolate the current components to higher frequency. This approach oers a way of creating new approximated currents. As a side result, we can observe that the currents on the shadow side of a roundish body can be modelled by these fairly simple models. Their values at the transition region into the shadow zone are also well handled by the modied reference models. Finally, the perpendicular component in our decomposition can been seen as a measure of the error in the PO current. Thus this provides a numerical procedure to automatically control the PO error. The development of the method and the numerical simulations have been done in close cooperation by both authors as a joined eort. The second author was the main responsible for the implementation of the modifying coecients. The author of this thesis was responsible for the literature studies, the PO streamline concept, the study of the TE-TM cases and the MoM simulations and presented this paper at the MMWP05 Conference on Mathematical Modelling of Wave Phenomena, Vxj, Sweden, August 2005. A shorter version of the report will appear in the proceedings of the conference.

4.3

Paper 3: Physical Optics and NURBS

In this paper we present an implementation of the Physical Optics (PO) technique for Radar Cross Section (RCS) application using an adaptive triangular subdivision scheme for the surface integral which arises during the computation of the electromagnetic scattered elds. Our interest was to assess eciency and error analysis. Classically, the PO surface integral is solved using quadrature formulas which require the wavelength to be resolved, i.e. the number of elements becomes proportional to the square of the electrical size. Instead, if we assume linear phase, we can solve the integral analytically on fewer larger elements, whose number grows only linearly with the electrical size. The size of the elements is determined by an adaptive scheme which supplies control of the error during the integration. The solver includes the fast treatment of the shadowing described in Paper 6 and uses the ray tracer on NURBS described in Paper 5. The author of the thesis implemented the PO solver, performed the computations, wrote the paper and presented it at the EMB04 Conference on Computational Electromagnetics, Gteborg, Sweden, October 2004.

4.4

Paper 4: The Rescue Wing, an Engineering Application

Here we present a multi-disciplinary engineering analysis aiming at the design of a new marine distress signaling device. The device, called "Rescue-Wing", works as

4.5. PAPER 5: MIRA, A MODULAR APPROACH TO GTD

33

an inatable radar reector and is intended for personal use in marine environment, assisting in localization of persons missing at sea during rescue operations. The Rescue-Wing is in fact an inatable gas bag shaped like a wing providing aerodynamic lift. Tethered to the life jacket of a person in distress, it is lled with helium to provide aerostatic lift. The Rescue-Wing has been designed to operate as a balloon in calm winds and, in windy conditions, like a kite. When deployed, it will position itself in tethered ight 10-15 meters above the sea surface providing a radar reector target as well as a strong visual cue for detection and positioning. The analysis has been focused on the multidisciplinary design task of combining results from aerodynamics, ight mechanics, structures and electromagnetics into one design. In order to assess the Rescue-Wings ability to reect radar signals, electromagnetics simulations have been conducted to predict its RCS. The computations involved the use of the General Electromagnetic Solvers "GEMS", e.g. the PO technique described in Paper 3 as well as the Method of Moments. We show how these methods have been used to assist in the analysis and in the design decision making process of the radar detectability of the Rescue-Wing, as well as how its RCS compare to the most popular radar reectors used on yachts and sailboats. The rst author was responsible for the aerodynamic part and presented the paper at the OCEANS 2005 Conference, Washington D.C., United States, September 2005. The author of the thesis was responsible for the analysis and the computations of the radar cross section. The introduction is a collaborative eort.

4.5

Paper 5: MIRA, a Modular Approach to GTD

In this paper we present the GTD solver MIRA: Modular Implementation of Ray Tracing for Antenna Applications. The low cost of GTD, compared to the Method of Moments, is due to both the fact that there is no runtime penalty in increasing the frequency and that the ray tracing, which GTD is based on, is a geometrical technique. The complexity is then no longer dependent on the electrical size of the problem but instead on geometrical sub problems which are manageable. For industrial applications the scattering geometries, with which the rays interact, are modelled by trimmed Non-Uniform Rational B-Spline (NURBS) surfaces [Piegl 91, deBoor 78] the most recent standard used to represent complex freeform geometries. In this paper we focus especially on the architecture of MIRA that separates mathematical algorithms from their implementation details and modelling. The part concerning the geometry has been written by the rst author, the ray-tracing part and the pictures of the rays has been done by the author of the thesis, the application part, the introduction and the last picture has been created by the second author.

34

CHAPTER 4. SUMMARY OF THE PAPERS

4.6

Paper 6: Architecture and Geometrical Algorithms in MIRA

Due to the introduction of NURBS presented in Paper 5, the geometrical sub problems tend to be mathematically and numerically cumbersome and introduce complications such as mathematical complexity and several representations of the same curve. We show how proper Object Oriented programming techniques has allowed us to restructure the ray tracer into a exible software package. The keywords here are portability and modularity. Implementation of these concepts was realized with Fortran 90 because of its robustness and portability, although it does not feature Object Orientation naturally. A well thought-out software design allows simple and ecient implementation of geometrical ray tracing algorithms. Our experience is that good software architecture leads to exibility and modularity, essential in the support of future enhancements. As a consequence, the independent modules of MIRA make suitable platforms for hybrid techniques in combination with other methods such as MoM and PO. In a rst innovative hybrid technique, a triangle-based PO solver uses the shadowing information calculated with the ray tracer part of MIRA. The occlusion is performed between triangles and NURBS surfaces rather than between pairs of triangles, thus reducing the complexity. Then the shadowing information is used in an iterative MoM-PO process in order to cover higher frequencies, where the contribution of the shadowing eects, in the hybrid formulation, is believed to be more signicant. This paper was presented at the EMB01 Conference on Computational Electromagnetics, Uppsala, Sweden, November 2001.

Bibliography
[Stratton 41] J.A. Stratton.Electromagnetic Theory. McGraw-Hill, 1941. [Rumsey 54] V. H. Rumsey. Reaction concept in electromagnetic theory. Physical Rev. (2), 94:1483 1491, 1954. [Keller 62] Joseph B. Keller. Geometrical Theory of Diraction. Journal of the Optical Society of America, Vol. 52(2), pp. 116-130, February 1962. [Kouyoumjian 65] R.G. Kouyoumjian. Asymptotic High-Frequency Methods. Proc, IEEE, Vol.53, No. 8, pp. 864-876, August 1965. [Ludwig 68] A. C. Ludwig, Computation of radiation patterns involving numerical double integration, IEEE Transactions on Antennas and Propagation, pages 767769, Nov. 1968. [Harrington 68] R.F. Harrington. Field Computation by Moment Methods. New York: Macmillan, 1968. [Thiele 75] G.A. Thiele and T.H. Newhouse, A hybrid technique for combining moment methods with geometrical theory of diraction, IEEE Trans. Antennas Propagat., vol. AP-23, pp. 62-69, Jan. 1975. [Burnside 75] W.D. Burnside, C.L. Yu and R.J. Marhefka, A technique to combine the geometrical theory of diraction and moment method, IEEE Trans. Antennas Propagat.,pp. 551-558, July 1975. [deBoor 78] C. deBoor, A Practical Guide to Splines. Applied Mathematical Sciences 27, Springer-Verlag, New York, 1978. [Rao 82] S. M. Rao, D. R. Wilton, and A. W. Glisson. Electromagnetic scattering by surfaces of arbitrary shape. IEEE Transactions on Antennas and Propagations, 30(3):409418, May 1982. [Pathak 84] R.G. Kouyoumjian and P.H. Pathak. A Uniform Geometrical Theory of Diraction for an Edge in a Perfectly Conducting Surface. Proc. IEEE, Vol. 62, pp. 1448-1481, November 1984. 35

36

BIBLIOGRAPHY

[Burnside 87] I. J. Gupta, W. D. Burnside, A Physical Optics Correction for Backscattering from Curved Surfaces, IEEE Transactions on Antennas and Propagation, vol. 35, N0. 5, May 1987. [Greengard 87] L. Greengard and V. Rokhlin, A fast algorithm for particle simulations, Journal of Computational Physics, 73(2):325-348, Dec. 1987. [Bowman 87] J.J. Bowman, T.B.A. Senior, P.L.E. Uslenghi Electromagnetic and Acoustic Scattering by Simple Shapes, HPC, 1987. [Farin 88] Gerald Farin. Curves and surfaces for computer aided geometric design: a practical guide. Academic Press Professional, Inc., San Diego, CA, 1988. [McNamara 89] D.A. McNamara, C.W.I. Pistorius, J.A.G. Malherbe Introduction to the Uniform Geometrical Theory of Diraction, Artech House, London, 1989. [Glassner 89] Andrew S. Glassner. An introduction to ray tracing. Academic Press Ltd., London, UK, 1989 [Piegl 91] Les Piegl. On NURBS: A survey. IEEE Computer Graphics and Application, South Florida, 1991. [Rahmat-Samii 91] C.S. Kim and Y. Rahmat-Samii, Low prole Antenna Study Using the Physical Optics Hybrid Method (POHM), IEEE Antennas and Propagation Society Int. Symp., vol.3, pp. 1350-1353, June 1991. [Crabtree 91] Glenn D. Crabtree, A numerical quadrature technique for physical optics scattering analysis, IEEE Transactions on Microwave Theory and Techniques, 27(5), Sep. 1991. [Rokhlin 92] N. Engheta, W. Murphy, V. Rokhlin and V. Vassiliou, The fast multipole method (fmm) for electromagnetic scattering problems, IEEE Trans. on Antennas and Propagat., vol 40, pp. 634-641, June 1992. [Numer. Recipes 92] Numerical Recipes in Fortran 77. Cambridge University Press, Volume 1, 1986-1992. [Bouche 93] D.P. Bouche, F.A. Molinet and R. Mittra, Asymptotic and Hybrid Techniques for Electromagnetic Scattering, Proceedings of the IEEE, vol.81(12), Dec. 1993. [Rius 93] J.M. Rius, M. Ferrando and L. Jofre. GRECO: Graphical Electromagnetic Computing for RCS prediction in real time. IEEE Antennas and Propagation Magazine, vol. 35, No.2, pp. 7-17, April 1993. [Leviatan 93] B.Z. Steinberg and Y. Leviatan, On the use of wavelet expansions in the method of moments, IEEE Transactions on Antennas and Propagation, vol. 41, pp. 610619, 1993

37 [Mittra 94] Z. G. Figan, R. Mittra, A. Boag and E. Michielssen, A technique for solving scattering problems at high frequencies, Int. URSI Symp., Seattle WA, June 1994. [Moreira 94] F.J.S Moreira and A. Prata. A self-checking predictor-corrector algorithm for ecient evaluation of reector antenna radiation integrals, IEEE Transactions on Antennas and Propagation, 42(2):246254, Feb. 1994. [Gordon 94] William B. Gordon, High Frequency Approximations to the Physical Optics Scattering Integral, IEEE Transactions on Antennas and Propagation, 42(3), Mar. 1994. [Catedra 95] M. Domingo, F. Rivas, J.Perez, R.P. Torres, M.F. Catedra. RANURS: Computation of the RCS of Complex Bodies Modelling Using NURBS Surfaces. IEEE Antennas and Propagation Magazine, vol. 37, No. 6, December 1995. [Aberegg 95] K.R. Aberegg and A.F. Peterson, Application of the Integral Equation-Asymptotic Phase Method to Two-Dimensional Scattering, IEEE Trans. on Antennas and Propagat., vol. 43, pp. 534-537, May 1995. [Jakobus 95] U. Jakobus and F. M. Landstorfer, Improved PO-MM Hybrid Formulation for Scattering from Three-Dimensional Perfectly Conducting Bodies of Arbitrary Shape, IEEE Transactions on Antennas and Propagation vol. 43, pp. 162169, February 1995. [Taove 95] A.Taove. Computational Electro-dynamics: The Finite-Dierence Time-Domain Method, Artec House, 1995. [Asvestas 95] J. S. Asvestas, The Physical-Optics Integral and Computer Graphics, IEEE Trans. on Antennas and Propagat., vol.43(12), Dec. 1995. [Altman 96] Z. Altman, R. Mittra, O. Hashimoto and E. Michielssen, Ecient representation of the induced currents on large scatterers using the generalized pencil of function method, IEEE Trans. on Antennas and Propagat., vol.44(1), Jan. 1996. [Hodges 97] R.E. Hodges and Y. Rahmat-Samii, An iterative current-based hybrid method for complex structures, IEEE Trans. Ant. Propagat. 45 1997. [Jackson 98] J. D. Jackson. Classical Electrodynamics. 3rd ed. New York: Wiley, p. 177, 1998. [Mittra 98] Mittra R., Peterson, A.F. and Ray S.L.. Computational Methods for Electromagnetics. IEEE Press, 1998. [Catedra 98] M.F Catedra, J. Perez, I. Gonzalez, I. Gutierrez, F. Saez de Adana. FASANT (Version S.4) Theoretical Foundations. Signal Theory and Communications Dept., University of Alcala. 1998.

38

BIBLIOGRAPHY

[Altman 99] Z. Altman and R. Mittra,A technique for Extrapolating Numerically Rigourous Solutions of Electromagnetic Scattering Problems to Higher Frequencies and Thier Scaling Properties, IEEE Trans. on Antennas and Propagat., vol.47(4), April 1999. [Fares 99] A. Bendali, M.B. Fares, CESC: CERFACS Electromagnetics Solver Code. Technical Documentation TR/EMC/99/52, Toulouse, France, 1999. [Bendali 99] A. Bendali, The Rumsey reaction principle. Lecture notes 1999. [Oppelstrup 99] R. Buch, J. Long, J. Oppelstrup PSCI Progress Report 1997-1999. PSCI, KTH, pp 70-82, 1999. [Kwon 00] D. Kwon, Ecient Method of Moments formulation for large conducting scattering problems using asymptotic phasefront extraction, Ph.D. dissertation, Dept. Elect. Eng., The Ohio State University, Colombus, OH, 2000. [Taboada 00] F. Obelleiro, J.M. Taboada, J.l. Rodriguez, J.O. Rubinos and A.M. Arias, Hybrid moment-method physical optics formulation for modeling the electromagnetic behavior of on-boaerd antennas, Microwave Opt. Tech. Letter 27,pp. 88-93, 2000. [Nilsson 00] Martin Nilsson, A fast multipole accelerated block quasi minimum residual method for solving scattering from perfectly conducting bodies, Ant. and Propag. Society International Symposium No 4, 2000. [Andersson 01] U. Andersson. Time-Domain Methods for the Maxwell Equations. Ph.D Thesis, KTH, Stockholm, 2001. [Ledfelt 01] G. Ledfelt. Hybrid Time-Domain Methods and Wire Models for Computational Electromagnetics. Ph.D Thesis, KTH, Stockholm, 2001. [Edlund 01] J. Edlund. A Parallel, Iterative Method of Moments and Physical Optics Hybrid Solver for Arbitrary Surfaces. Licentiate Thesis, Uppsala University, 2001. [Kwon 01] D. Kwon, R.J. Burkholder and P. H. Pathak, Ecient Method of Moments Formulation for Large PEC Scattering Problems Using Asymptotic Phasefront Extraction (APE), IEEE Trans. on Antennas and Propagat., vol. 49, No.4, April, 2001. [Taboada 01] J. M. Taboada, F. Obelleiro, and J. L. Rodrgeuz. Improvement of the hybrid moment method-physical optics method through a novel evaluation of the physical optics operator. Microwave and Optical Technology Letters., 30(5):357363, 2001.

39 [Sarkar 02] C. Su, T.K. Sarkar, Adaptive Multiscale Moment Method (AMMM) for Analysis of Scattering from Three-Dimensional Perfectly Conducting Structures, IEEE Transactions on Antennas and Propagation, vol. 50, N0. 4, April 2002. [Edelvik 02] F. Edelvik. Hybrid Solvers for the Maxwell Equations in TimeDomain. Ph.D Thesis, Uppsala University, 2002. [Se 03] S. Se. Ray Tracing Tools for High Frequency Electromagnetics Simulations. Licentiate thesis, KTH, Stockholm, 2003. [Abenius 04] E. Abenius. Time-Domain Inverse Electromagnetic Scattering using FDTD and Gradient-based Minimization. Licentiate Thesis, KTH, Stockholm, 2004. [Djordjevic 04] M. Djordjevic and B.M. Notaros, Higher Order Hybrid Method of Moments-Physical Optics Modeling Technique for Radiation and Scattering from Large Perfectly Conducting Surfaces, IEEE Transactions on Antennas and Propagation, Feb 2004. [Nilsson 04] M. Nilsson. Fast Numerical Techniques for Electromagnetic Problems in Frequency Domain. Ph.D Thesis, Uppsala University, 2004. [Hagdahl 05] S. Hagdahl. Hybrid Methods for Computational Electromagnetics in Frequency Domain. Ph.D Thesis, KTH, Stockholm, 2005. [Taboada 05] J. M. Taboada, F. Obelleiro, J. L. Rodrgeuz, I. Garcia-Tunon and L. Landesa, Incorporation of Linear-Phase Progression in RWG basis functions, Microwave and Optical Technology Letters., 44(2), Jan. 2005. [1] [2] [3] [4] [5] CVS, the Concurrent Versions System: http://www.cvshome.org netCDF, the network Common Data Form: http://www.unidata.ucar.edu/packages/netcdf Matlab, the Matrix laboratory: http://www.mathworks.com OpenDX, the Open Visualisation Data Explorer: http://www.opendx.org CADx: TranscenData http://www.cadx.com

ICNAAM 2005 Extended Abstracts, 1 4

Modeling and Extrapolating High-frequency Electromagnetic Currents on Conducting Obstacles

Sandy Se1 and Fredrik Bergholm1 .
1

Royal Institute of Technology, KTH, SE-10044 Stockholm, Sweden

Key words Numerical Methods, Electromagnetics, Method of Moments, Physical Optics, Surface Currents. We present a current-based approach to high frequency approximate techniques in Computational Electromagnetics (CEM). Our goal is to numerically model the behavior of electromagnetic and surface current vector elds at high frequency, using information extracted from lower frequency solutions.
c 2005 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim

1 Introduction
The Method of Moments [1] (MoM) is the most common current-based method for solving electromagnetic scattering problems. It involves the solution of a large linear system of equations that has to be computed for each frequency with surface currents as unknowns. The MoM can be seen as an asymptotically exact brute force method which does not incorporate any knowledge about the structure of the vector elds. It results in a full dense system of equations, thus limiting its use from low to mid-range frequency. For high frequency, approximate techniques are also available, either crude current based approximations such as Physical Optics, or ray-based techniques [2] which require known asymptotic high frequency solutions such as Geometrical Optics or the Geometrical Theory of Diffraction [3]. In this work we study the possibility of extracting, from lower frequency (MoM), information about the electromagnetic and surface current vector elds that are slowly varying over smooth bodies. Our goal is to extrapolate these surface currents to higher frequencies, thus avoiding the prohibitively high cost of solving for billions of unknowns that would be required by MoM and yet keeping better accuracy than classical high frequency methods. Our method decomposes the low frequency surface currents, expressed as triangular-type Rao-Wilton-Glisson [4] (RWG) vector basis functions in the MoM formulation, into parallel and perpendicular vector components. These components are labeled nxH-parallel currents and cross-currents respectively and are dened relatively to the direction of the incident magnetic eld on the surface. The two currents show systematic spatial patterns evolving over frequency in close correlation with the incident magnetic eld, both on the illuminated and on the shadow side of the body. Such a vector eld decomposition is not standard and yields a better understanding of the structure of the surface currents as well as the underlying physics governing their spatial variations, i.e. charge transport and charge accumulation.

2 Vector eld Decomposition - a background

We start with the most simple current vector eld, Jpo from the Physical Optics (PO) technique. It is an approximation of the induced surface currents1 on a smooth perfectly conducting (PEC) body whose dimensions are large compared to the wavelength . It is induced only by the magnetic eld Hinc : Jpo (r ) Jpo (r )

= 2 n(r ) Hinc (r ) (r in lit region) = (r in shadow region) 0

(1) (2)

Corresponding author: e-mail: sandy@ nada. h. Phone: +46 8790 62 29, Fax: +46 8790 64 57 kt se,

1 We abbreviate current density by current.

c 2005 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim

 where r is the position vector for a given point on the surface and n the outward surface normal. For a high frequency or a fairly high wavenumber k = 2 , this approximation is sufcient for at smooth areas away from shadow boundaries or geometrical discontinuities. Sharp edges or corners leads to erroneous Jpo currents. From the accounts for the Method of Moments, which computes exact currents, it is evident that, on average, Jpo explains large parts of the surface current pattern, almost irrespectively of the chosen spatial frequency k. Even low frequencies with small k, exhibit Jpo like surface currents. It is intuitively clear from Eq. 2 that the source of the approximate Jpo currents, the magnetic eld, creates J = n Hinc on the illuminated side of the body. For an incident planar wave in the propagation direction k = , where the shadow-lines nk are along the equator, we have: z Einc (r ) = x E0 ejkz ejwt Hinc (r ) = (E0 /Z0 ) ejkz ejwt y

(3) (4)

where r = x + y + z, Z0 = 377 ohms is the free space impedance and w is the angular frequency. x y z We note that, in such a conguration, there are other currents present as well, and they are not equally easily explained. One important observation is nevertheless that the Jpo currents have only a limited relation to charge accumulation, which lead us to following denition: Observation 2.1 For plane wave incidence where {Einc , Hinc , k} form a trihedron, the Jpo currents run along planar sections {span{Einc , k} body} that we call streamlines. Concretely, for a sphere in polar coordinates2 , the Jpo streamlines are: (x, y, z) = (cos cos , sin , cos sin ) (5) These streamlines run as parallel sets of curves there is no convergent or divergent pattern of current lines. There is only a periodic reversal of direction of current along these streamlines, due to the periodic Hinc , see Fig. 1(a), making in most places Jpo = 0. This leads us to the following observation: Observation 2.2 Apart from the periodic variation (spatially) along the Jpo streamlines, the part of the surface current eld associated with charge accumulation is the non-Jpo eld. Hence, if decomposing the total surface current J into two vector elds, irrespective of k, it is natural to let Jpo be one component in that eld, J Jpo be the other component. Observation 2.2 then tells us that charge accumulation behavior must primarily be contained in J Jpo = Jc . We now make an assumption. Let J be the component of J that is parallel to the Jpo current streamline (Eq. 5), which we denote by , a parametric curve (u) = (x(u), y(u), z(u)) on the surface. Assumption 2.3 The parallel component J of the surface currents behaves qualitatively like the Jpo currents. For that reason, we decompose the total current J in J and J = J J , where the latter reminds of Jc which we expect is due to other time-varying charge accumulation patterns than those directly associated with current variations along . Denition 2.4 We decompose J as follows: J = (J T)T : nxH-parallel currents. (6) d : Tangent along the streamline . T= du J = J (J T)T : Cross currents perpendicular to the streamline . (7) (8)

Consequently, we have dened a new entity, which we call cross currents, by the above equation. Corresponding scalar quantities for the amplitudes of the components are: J = J sgn(J ) = J T (9) J = J sgn(J ) (10) where the sign (sgn = 1) of sgn(J ) is determined by the direction of the traveling current (Eq. 7) and J = J nc , where nc is outward normal to the curve. Our underlying strategy, from now on, is to model the entities J and J spatially over the body, as well as over frequency k.
2 Non-standard polar angles are used to create plane sections, see Fig. 1(a) for = 0, [0, 2] 2

1 0.8 0.6 0.4 0.2

2 4

x 10

Components of current Real(J) on triangle1 real(Jx) real(Jy) real(Jz)

E H k

Real(Jx), Real(Jy), Real(Jz)

0 0.2 0.4 0.6 0.8 1 0.2 0.1 0

1 0.1 0.2 1 0.5 0.5 0

50

100

150

200

250 f [MHz]

300

350

400

450

500

(a) Incident magnetic eld Hinc along the = 0 streamline .

(b) Evolution over frequency of the component of the MoM current at a point of the streamlines.

Fig. 1 Evaluation of MoM currents on a sphere 1 meter in radius.

3 Modeling of J and J
For general roundish bodies, we suggest that cross currents follow the uctuations of Hinc . Let Y = E0 /Z0 . We model J by : J = 2Y cos(k cos sin ) on , both on illuminated and shadow side and we expect a-priori that an approximate model of J is: 2Y cos(k cos sin ) + C0 (12) d where C0 is a modelisation constant. The logic is that cross currents J are due to changes of charge and J accumulation of charge takes place at turning points of J , where d is high. J = (11)

4 Preliminary Results and Discussion

To get a better intuition of the structure of vector elds represented by the MoM currents, we have studied their evolution at one point on the surface of a sphere of radius a = 1 meter. The result is displayed in Fig. 1(b), where we observe how the current rotates slowly with frequency by reversing its components. In addition we see that, at higher frequency (f > 350MHz in this case), the MoM requires more elements to resolve the wavelength and without what, the MoM starts to predict erroneous values. Fig. 1(a) illustrates the behavior of the magnetic eld on the sphere along the = 0 streamline. The directions of the eld on the surface oscillate following the exponential term in Eq. 4. The model of our new approximate currents has been implemented and the results along a Jpo current streamline are displayed in Fig. 3, and compared to MoM. We note here that the MoM result contains discontinuities due to the coarse triangulation used. Smoothing in the data is required to emphasize the pattern in MoM, especially when looking at the cross current. We use the simple model dened in section 3 and get good agreement for the parallel component J at both low and higher frequency. The nxH-parallel component, like the Jpo current, displays error which amplies with k close to the shadow-lines. The perpendicular component J is more difcult to model, as seen in Fig. 3(a) for low frequency ka = 10 and in Fig. 3(b) for higher frequency. To this end, we intend to use extraction from low frequency MoM in order to improve the amplitude and phase in our model, in particular around the shadow-lines.
3

x 10

Parallel Current Component: k=10 cut Phi=0.3927 6 From MoM Modeled

x 10

Parallel Current Component: k=19 cut Phi=0.3927 From MoM Modeled

4 2 2 |J| |J| 0 2 2 4 6 0 50 100 150 200 250 0

50

100

150

200

250

(a) Case ka=10

(b) Case ka=19

Fig. 2 Simulation results for the parallel component J compared to MoM.

1.5

x 10

Perpendicular Current Component: k=10 cut Phi=0.3927 2 From MoM Modeled 1.5

x 10

Perpendicular Current Component: k=19 cut Phi=0.3927 From MoM Modeled

1 1 0.5 0.5

|J|

|J| 0 50 100 150 200 250

0.5 0.5 1 1 1.5

1.5

50

100

150

200

250

(a) Case ka=10

(b) Case ka=19

Fig. 3 Simulation results for the perpendicular component J compared to MoM.

Acknowledgements Financial support has been provided within the GEMS project by KTH and the Swedish Agency for Innovation Systems (VINNOVA) as parts of a collaborative research center PSCI.

References
[1] [2] [3] [4] R.F. Harrington. Field Computation by Moment Methods. New York: Macmillan (1968). S. Se, Ray Tracing Tools for High Frequency Electromagnetics Simulations, Licentiate thesis KTH (2003). Joseph B. Keller. Geometrical Theory of Diffraction. J. Opt. Soc. of Am. 52, 116130 (Feb 1962). Rao, Glisson, Wilson, IEEE Trans. Ant. Prop. 30, pp. 409-418 (1982).

Extrapolation and Modelling of Method of Moments Currents on a PEC Surface.

Sandy Se and Fredrik Bergholm1

TRITA: NA-0539 Stockholm 2005 Technical Report Royal Institute of Technology Department of Numerical Analysis and Computer Science

c Sandy Se and Fredrik Bergholm1 , November 2005 TRITA: NA-0539 Stockholm 2005
1 Authors

in random order.

Extrapolation and Modelling of Method of Moments Currents on a PEC Surface.

Sandy Se
Department of Numerical Analysis and Computer Science, NADA Royal Institute of Technology, KTH Stockholm, Sweden Email: sandy@nada.kth.se

Fredrik Bergholm
Department of Numerical Analysis and Computer Science, NADA Royal Institute of Technology, KTH Stockholm, Sweden Email: fredrikb@nada.kth.se

Abstract We present a current-based approach to high frequency approximation techniques. Our goal is to numerically model the behaviour of electromagnetic and surface current vector elds at medium-range or high frequency, using information extracted from lower frequency solutions.

I. I NTRODUCTION Method of Moments [1] (MoM) is the most common surface current based method for solving electromagnetic scattering problems. It involves the solution of a full large linear system of equations for each frequency, with surface currents1 , as unknowns. MoM can be seen as an exact brute force method which does not incorporate any knowledge about the geometric structure of the vector elds involved, thus limiting its use to electrically small problems. For high frequency electrically large problems, approximate techniques are also available, either crude current-based approximations such as Physical Optics (PO) or ray-based techniques which require known asymptotic high frequency solutions such as Geometrical Optics or the Geometrical Theory of Diffraction (GTD, UTD) [3][5]. These can provide very fast results but are approximations that do not incorporate a complete modelling of the physics and thus, typically degrade in accuracy. In the intermediate frequency range neither technique is effective and in the past decade, efforts have been concentrated on hybrid approaches, resulting in four main groups of hybrid methods. The rst group uses domain decomposition combined with high-frequency techniques. The strategy is to achieve a sparser matrix by splitting large smooth geometry parts away from small details modelled by the MoM. The pioneer schemes [6], [7] were realized for specialized geometries, by combining the MoM with GTD. For general geometries, GTD involves more complex ray tracing [8], a non robust operation which
1 We

abbreviate current density by current.

also tends to be expensive when introducing more asymptotic phenomena such as multiple interactions or creeping waves [9]. In contrast, a current-based PO is a preferable choice by making easier the integration with the current-based MoM. This explains the worldwide development of MoM-PO hybrid solvers [10][14]. In our experience, the main weakness of this method is that of displaying large errors where the PO fails, i.e. in shadow regions or near edge discontinuities. We believe that shadow regions must be properly avoided when using PO, in such a manner that the PO domain has to be reassigned for every angle of incidence, leading to more complex implementation. The second group of methods focuses on obtaining sparsity of the full matrix of the MoM with efcient algorithms such as matrix compression using Wavelet [16], Multi-grid [17] or efcient matrix vector products in the Fast Multipole Method [19] (FMM). The latter, being so far the most successful, uses block decomposition [20] where the far eld elements are regrouped into small low rank blocks [21]. This allows for larger scale problems, but the technique still needs to represent the fastest variation of the phase thus requiring the same large number of unknowns as in MoM. Schemes to reduce the total number of unknowns rst appeared for PO, with linear phase approximation scheme [25] and quadratic phase approximation [26]. A recent implementation can be found in [28] and the method described in [27] was also extended to a MoM-PO linearly phased hybrid solver [29]. The method performs faster using larger elements, but, without proper treatment of the shadow regions, it will still suffer from degradation in accuracy like the standard MoM-PO. However, in such a case, the reassignment of the PO regions will necessitate remeshing of the new MoM domain, leading to more expensive MoM-PO. Reduction of the number of unknowns in the MoM formulation can be achieved in 2D with the Integral Equation Asymptotic Phase (IE-AP) method [32], or more recently in 3D [30], with alternate basis functions which incorporate variation of phase. More complicated highorder basis functions are also available for the MoM-PO [31].

Finally, the last group of methods is the Extraction methods. They are based on extraction of characteristic properties of a known current obtained from lower frequency solutions and then extrapolation of the currents to higher frequency [33]. Such extraction-extrapolation approaches have not been much explored yet but have been reported in [34] to be possible [35]. Recently, the Asymptotic Phasefront Extraction (APE) [36] and its extension the APE-MoM [37] use a similar domain decomposition to MoM-PO with high order basis functions. APE uses GTD physical insight of the electromagnetic elds to predict the behaviour of the geometrical optics induced currents. In this paper we focus on obtaining insight about the behaviour of the surface current owing to the PO and the MoM induced currents instead of GTD. Our method rst decomposes the low or mid-range frequency surface currents expressed into triangular-type Rao-Wilton-Glisson [22] (RWG) vector basis functions in the MoM formulation, into parallel and perpendicular vector components relative to the direction of the incident magnetic eld on the surface that we call nHparallel currents and cross-currents respectively. These two currents yield systematic spatial patterns evolving over frequency closely related to the evolution of the incident magnetic eld over frequency, both on the illuminated and shadow side of the body. Such a vector eld decomposition is not standard, and yields a better understanding of the structure of the surface currents as well as the underlying physics governing their spatial variations, i.e. charge transport and charge accumulation. Section II presents the background of the decomposition on which the method is based. Different models for the induced currents are discussed in Section III for both components and illustrated in the case of a sphere. Section IV shows the use of the RWG basis functions and Section V to VI discuss the extraction procedure and how to obtain frequency dependent behaviour from low frequency MoM or FMM runs. Numerical results are presented in Section VII for the extrapolation and in Section VIII for our models, as well as in Appendix B. Conclusions are drawn in the last section. II. V ECTOR FIELD D ECOMPOSITION - A BACKGROUND For a smooth perfectly conducting body, it is well-known that the PO-currents: JP O JP O = 2 n Hinc = 0 (shadow region), (1) (2)

H 1 K

0.5

0.5

1 0.5 0 0.5 1 x 1 y 0.5 0 0.5

Fig. 1.

Streamlines are level curves transverse to Hinc .

for sufciently high wavenumber k = 2/, are a reasonably good approximation of the induced surface currents J C3 . I However, to our knowledge, there are no general error estimates telling how good this approximation actually is for arbitrarily shaped bodies. It is known that JP O currents close to shadow boundaries are quite erroneous. To alleviate this problem, some researchers in the nineties [38] apply diffusion to JP O currents to create a continuous transition into the shadow. Another technique, [24], works instead with the scattered eld 2

obtained by integrating a smooth function of JP O as integrand, extended into the shadow. It is evident that JP O explains large parts of the surface current pattern, almost irrespective of the chosen frequency k. Even low frequencies k, exhibit JP O like surface currents2 . The physical origin of the JP O currents is intuitively clear the magnetic eld, e.g. an incident planar wave creates before interactions with neighbouring surface elements Jinc = n Hinc on the illuminated side of the body, where n is the outward unit-length surface normal, and Hinc the incident magnetic eld. If we treat the sphere as transparent, then the incident magnetic eld, as seen in Fig. 2(b) along the whole streamline, can be evaluated in the shadow too. This means that the total induced surface current J = n H is related to Jinc everywhere. According to our experience, this is a good way of capturing the oscillations of J. To summarize, induced currents contain a large part of POlike currents. However, there are other currents present as well, and they are not equally easily explained. One important observation is nevertheless that JP O currents have only a limited relation to charge accumulation, to be discussed below. For an incident planar wave in the propagation direction k=(0, 0, k), we have Einc (r) and Hinc (r): Hinc = (E0 /Z0 ) eikz eiwt y Einc = x E0 eikz eiwt (3) (4)

where r = x + y + z, and Z0 = 377 ohms is the free space x y z impedance and w is the angular frequency. We refer to (0, 0, 1) as the North pole. The equator is the shadow boundary.
2 For all variables in this text, vectors but not scalars are in boldface. Vector dimension is 3 unless stated otherwise.

In this case, the PO currents run along planar sections

1

{y = Const} S, where S = {(x, y, z) : f (x, y, z) = 0} is the set of points on the surface S of the body. Hence, the PO current streamlines are level curves, see Fig. 1. Concretely, for a sphere: (x, y, z) = (cos cos , sin , cos sin ) (5)
=0
z

0.8 0.6 0.4 0.2 0 0.2 0.4

E H k

0 =

0.6 0.8 1

0.2 0.1 0 1 0.1 0.2 1 0.5 0.5 0

are the PO current streamlines obtained analytically, with (a) Parameterization of the hemisphere (b) Incident magnetic eld Hinc , as curve parameter along them, and [0, ], on the z = cos sin , < < . Eq. 4, along one streamline. illuminated side, as seen in Fig. 2(a). For complex shape, the PO current streamlines can be obtained from the intersection Fig. 2. Streamlines parameterization and incident magnetic eld. between the surface and the plane transverse to Hinc . The set of intersecting points are then ordered clockwise to form the discrete PO current streamlines. Fig. 3(b) illustrates the result of this procedure on the small aircraft code-name Eikon displayed in Fig. 3(a). Observation 1: These streamlines run as parallel sets of curves there is no convergent, divergent pattern of current lines. There is only a periodic reversal of direction of current along these streamlines, due to the periodic Hinc as seen in Fig. 2(b), making: JP O = 0 (6) (a) CAD description of the UAV. (b) Streamline = 0 on the UAV.
Streamline on Eikon 2.8 2.6 2.4 2.2 2 1.8 5.3 1.6 5.3 5.3 1.4 5.3 3 2 1 0 1 2 3 4 5.3

in most places. The conclusion is: Observation 2: Apart from the periodic variation (spatially) along the PO current streamlines, the part of the surface current eld associated with charge accumulation is the non-JP O eld. Hence, if decomposing the surface currents into two vector elds, irrespective of k, on the lit side, it is natural to let JP O C3 be one component in that eld, and J JP O be the I other component, Observation 2 then tells us that transverse charge accumulation behaviour must primarily be contained in J JP O = Jc . (7)

Fig. 3.

Unmanned Aerial Vehicle (UAV) code-name Eikon.

where T is the unit tangent along . Furthermore, J = J (J T)T (9)

In other words, we expect that time-varying charge accumulation patterns, not associated with the PO streamlines, are captured by Jc . We now make an assumption. Let J be the component of J that is parallel to the PO current streamline (cf. Eq. 5), which we denote by , and the curve is a function of the curve parameter u: (u) = (x(u), y(u), z(u)). This component is parallel to n H . In Eq. 5, the curve parameter u = . Assumption 3: The nHparallel component J behaves qualitatively like the PO current JP O modulo a slowly varying real function, on the illuminated side. We decompose J in J and J J , see Fig. 4(a), where we expect that Jc = J JP O J J carries time-varying charge not directly associated with current variations along . Denition 4: The nHparallel component is: J = (J T)T (8) 3
inc

which we call cross currents. Consequently we have dened two new entities which have complex values. From now on, we will only analyze the real part JRe of the current. The imaginary JIm can be treated analogously at a quarter of period later (t = /2), since the current is dened by J = JRe + iJIm and eit J = ei/2 J = iJ and furthermore JRe J
Im Re = JRe + JRe = J Re T + J nc

(10)

(11) (12)

= J

Im

JIm

=J

Im

T+

Im J nc

where T is the unit-length tangent and nc is the unit-length normal to the curve in the tangent plane of the body. The corresponding scalar quantities for the real components are: J Re
Re J

= JRe T = J
Re

(13) (14)

nc

Their signs are determined by the choice of the current curve travelling direction. Our underlying strategy, from now on, is to model the entities J and J (we drop the upper Re in the notations), spatially over the body, as well as over frequency k.

A. The Functionality of Cross Currents By running MoM for low and mid-range frequencies, on a sphere, we nd that on the sphere equator there are equatorial cross currents running from (e.g.) East to West poles (r=(1, 0, 0)), as well as cross currents at other latitudes running from East to West or vice versa. The intuitive interpretation behind this is that the cross currents J transport charges back and forth between East and West parts on either side of the closed body. The x-axis, connects East and West, and the E-eld runs in the xdirection. In line with Observation 2, J performs other charge accumulation patterns than those seen along the PO streamlines. Hence, a charge transport appears to run along latitude circles with respect to the North pole, as indicated in Fig. 4(b). Dene yet another component of J, namely transverse currents, denoted J : J = TL J (15) where TL is a unittangent vector along latitude circles with respect to the North pole, it is clear that J approaches J when approaches zero, and that J and J are highly correlated for [0, /4]. The standard analytical solutions, Mie solutions [2] for a sphere (which is one of the few closed bodies for which there are analytical solutions), are also expressed in latitudes and longitudes with respect to the North pole. B. What TM- and TEpolarization Tell us A planar approximation of the scattered elds on the lit side of a smooth body, leads to the PO current, see [23]. This is reasonably valid if the radius of curvature is small compared to the wavelength. The approximating assumption is simply that an incident planar wave gives rise locally to a scattered planar wave (direct reection). What is as interesting is, however, that more information can be gained from this simple approximation. There are two generic cases, the well-known TM- and TEcases, characterized by magnetic (or electric) incident elds transverse to the plane of incidence, spanned by k and n. See gures in Appendix A. The TEcase appears on the meridian = /2 running from the North pole through the forward pole (0, 1, 0). The interesting thing is that for the scattered Eeld there is no component normal to the surface, n E = 0, which means no charge along this meridian. Physically, one would then not expect any current running tangentially along the zero charge meridian, having time invariant vanishing electric eld. The PO-current is perpendicular to that meridian for roundish bodies. The conclusion is that for this kind of model in this case the constraint J ( = /2) = 0 (16)
J

(a) Decomposition along a streamline (b) Patterns on the hemisphere of nxHof the MoM current J in nxH-parallel parallel (rotating around y-axis) and and cross current J . transverse currents (rotating around zaxis). Fig. 4. Decomposition of the MoM current.

surface, the total Eeld has only a component normal to the surface, nE = s where s is the charge density, which means that we have a maximal transverse charge along this meridian, i.e. maxima in a direction perpendicular to the meridian in question. This is so, because in any other location you obtain a weighted average of the TE- and TM-cases. Physically, with no counter-gradient transport of charge, cross currents J should be expected to be zero for = 0. The conclusion is here that the constraint should be: J ( = 0) = 0 (17)

Both these constraints are also valid for the analytical Mie solutions, [2], for a sphere, so they are not just the consequence of a planar wave approximation. C. Charge Accumulation and Cross Currents From a theoretical point of view, we have already mentioned that the TE and TMpolarizations with reected planar waves may be qualitatively valid local models for mediumrange and high frequency for smooth bodies, and that they, as a byproduct, predict that there will be maximal charge accumulations on both halves of the illuminated part of the EastWest meridian. This holds for any roundish body. Observation 5: Local planar wave solutions (TM polarization) predict maximal charge accumulation along the EastWest meridian of a roundish body, since the resulting normal component of E is of maximum amplitude, in a transverse direction (latitude circle Fig. 4(b)). The conclusion is that J contributes to the transport of charge between the curves of maximal charge, = 0, [0, /2] and with opposite charge, = 0, [/2, ]. There are transverse patterns of currents converging onto the EastWest meridian, hence zero current where they change sign. D. What 2D Analytic Solutions Tell us For the TE-case on the zero meridian (from North pole by way of Front pole to the South pole), the analytic solutions for a 2D circular section of radius a are available and can be run for high k-values, see Appendix D. The total E eld oscillates 4

holds, and is expected to be valid for quite general roundish bodies, and any frequency. The TMcase appears on the meridian = 0 running from the North pole through the West pole (1, 0, 0). Here, on the

in the shadow region, but vanishes fairly quickly for ka = 29. The higher k-values, the more of the shadow zone meridian has E 0. Since J is related to E, we conclude that for this particular meridian the currents tend to the PO-currents quickly. III. M ODELLING OF S URFACE C URRENTS For roundish bodies, we suggest that both parallel and cross currents follow the uctuations of Hinc,Re . Denote the inverse free space impedance by Y = 1/377. Set E0 =1. We will concentrate on closed smooth bodies. On the shadow side, it is not clear how to specify the inuence of Hinc , which is why simple approximations such as JP O with zero values on the shadow side are used, or smoothing the JP O across the shadow boundary. However, it seems reasonable that Hinc still is a major factor behind the induced surface current pattern, and in a crude approximation, it may be worthwhile to use Hinc in the model of surface currents also on the shadow side. We investigate this empirically by studying MoM-results for low or mid-range frequencies. Such a study, resulted in the heuristic formulas given below, the so-called reference model equations. It should be noted that these reference models are not necessarily realistic models, but rather reference points for further modications, as described in later sections. A. Reference Model for Parallel Currents We model J on the previously mentioned streamline curves by: J = 2Y cos(kz), z = z(u), (lit side) (18)

The same model for the complex numbers for the phasor yields (i = 1): J Ref J =
Ref

= Y exp(ikz(u))

2Y exp(ikz(u))

(lit side). (shadow side).

(21) (22)

This means the imaginary part of J is also modelled here. B. Parallel Currents on a Sphere or Ellipsoid In the special case of a unit radius sphere, z = R sin , R = cos , we obtain the reference models for the real part: J Ref = 2Y cos(kR sin ), J Ref = Y cos(kR sin ), [0, ], [, 2], (23) (24)

which can be used for any smooth closed body. The formula can be derived as follows: Since Hinc = Y cos(kz) and by Assumption 3 with JP O = 2n Hinc on the illuminated side of the closed body, the above equation is obtained. However, and this is an important point, we consider Hinc to be a major explanatory factor behind the induced currents also on the shadow side, and with some boldness use Hinc also on the shadow side, however, not using the factor 2 since that factor is motivated by direct reection on the lit side. Whether this is a good or bad model is partly an empirical story. Our study of MoM-results indicates that this is a fairly reasonable model, abstracting away from details. The reference model for the shadow side is thus: J Ref = Y cos(kz), z = z(u), (shadow side). (19)

where the choice of determines which streamline to follow. Note here that for an incident plane wave in the z-direction, the same formula applies to an ellipsoid with semi-axis of unit length in the z-direction. This is so because z = R sin still holds. The rst equation is for the illuminated side, and the second for the shadow side, as mentioned. If giving names not only to the North and South poles of the sphere (with the North pole in (0, 0, 1)) we also have East and West poles at (1, 0, 0) and Forward and Backward poles at (0, 1, 0). We will speak of N-,S-,E-,W-,F-, and B-poles in what follows. The curves for = Const are latitude curves with respect to the F-pole. R= radius of such latitude circles. Hence = 0 is the latitude circle crossing the illuminated N-pole running from E-pole to W-pole. We will tabulate modelling results with respect to in a later section. The signicance of the F- and B-poles lies in that they are the singular cuts for the planar sections of the closed body for the PO current streamlines. A good model for the cross currents is a harder feat, but even some quite crude models seem to be of some descriptive value for simple closed bodies. C. Reference Model for Cross Currents On , we expect a-priori that an approximate model of J is: d Y cos(kz(u)) + C0 (25) du In practice, C1 is not a constant, but if we wish to to have a simple reference model, one can choose some neutral value of C1 . The logic is that cross currents J (u) vary periodically along the curve (u), with local extrema for certain positions, um , m = 1, 2, .... We expect J to be of larger magnitude where accumulation of charge takes place. Along the curve, we know that accumulation occurs at sign changes of J , and more generally at any inection points of d d2 J ( du2 J = 0). From this follows that du J attains local extrema at these inection points. If using this function for d2 J (u), and set um = {u : du2 J = 0}, we obtain expected extrema for cross currents. J (u) = C1

In what follows, our goal will be to model and t an amplitude function R(k) in (shadow side). (20) We look upon the reference model as a simple tool for obtaining currents in the shadow region at intermediate frequency. For high frequency we expect R to vanish as k increases which motivates the zero PO currents in the shadow regions. 5 J model = R(k)Y cos(kz), z = z(u),

A simpler way of achieving something similar is to replace the cos by sin in the formulas to move from J to J . We have also tried this, but achieved better results with the above formula. It should also be remembered that the main point in this simplistic reference model is that charge accumulation is oscillatory, and that cross currents as a consequence are oscillatory too. The phase of this variation is not expected to exactly follow the reference model. However, around = /2 the function for cross currents must be antisymmetric, due to the constraint Eq. 16. D. Cross Currents on a Sphere For a sphere, we set up a reference model from Eq. 25 without distinguishing between illuminated and shadow side. 1 We choose C0 = 0 and C1 = kR where the normalization factor kR originates from the inner derivative. For a spheroid type of body, our reference model for the perpendicular currents is:
Ref J () = Y sin(kR sin ) cos

Fig. 6. J from MoM at ka=19 on the lit side of 1/4 of a sphere, side view.

(26)

The curve parameter [0, 2] completes a full circle.

0.95 0.9

0.85

0.8

structures are the equatorial cross currents close to the shadow boundary and a couple of bands across the body. In the top view in Fig. 5 the direction of a narrow bundle of equatorial currents running clockwise can be seen, next to a broader band of cross currents running counterclockwise. When extracting components of surface currents using MoM, to be described more in detail in the next section, a general experience is that data are inherently noisy. The calculated J (u) do not form smooth proles along the streamlines, as a function of the curve parameter u. Such nonsmooth data are seen in Fig. 5. F. GTD Currents in Shadow Zone Along Geodesic Paths In GTD/UTD, an approximation for the surface currents are the creeping wave currents [5]. They are calculated using the GTD diffracted magnetic eld and diffraction coefcients based on canonical solutions on simple geometries. Concretely, Escat = D(k, sd ).eiksd (27)

0.75

0.7

0.65

0.6

0.55

0.5

0.45 1

0.8

0.6

0.4

0.2

0 xaxis

0.2

0.4

0.6

0.8

Fig. 5.

J from MoM at ka=10 on the lit side of 1/4 of sphere, top view.

E. Cross currents from MoM If taking the currents J from MoM, and performing a decomposition into J and J , how well are the reference models and constraints, mentioned so far, reected in such data? Let us study the example of a sphere of radius a, for two electric sizes ka = 10 and ka = 19, on the lit side, as illustrated in Fig. 5 and Fig. 6, respectively. First, we clearly see that the cross currents are zero on the meridian connecting the North and Forward poles, as stated by the TE constraint, Eq. 16. Furthermore, in Fig. 6, the cross currents diminish towards the farther-most East-West meridian in accordance with the TM constraint, Eq. 17. Other noticeable 6

where sd is the distance travelled along the geodesic, D(k, sd ) are the diffraction coefcients times the wavefront attenuation times a complex vector. Thus, the creeping wave currents are obtained from the associated magnetic elds as Jsd = n H. Looking at J and J|| , we can see systematic deviations from creeping currents being aligned with the geodesics, essentially away from the East-West meridian. Some deviations from the phasefront directions of the currents, see below, were also mentioned in [36]. Furthermore, if disregarding cross current, charge transport along PO streamlines could be an alternative to geodesic wavefront transport. G. APE-MoM Phasefronts [37] proposed another representation of the surface currents, over large smooth regions, with one or several linear phase currents as J(r ) = Ceik
(m)

(28)

where C is a complex vector amplitude assumed to be constant and k(m) is the direction of a phasefront travelling on the surface. At each point r on the surface, a local low frequency phase variation is split into m 1 different phasefronts. The phasefront extraction is achieved using a multi-dimensional complex Fourier transform j(ku , kv ) on a rectangular grid of the low frequency complex current as j(ku , kv ) =
S

J(r )eik

inc

dS

(29)

where ku , kv are the components of the projection of kinc on the surface S and J C3 is obtained from MoM. I The phasefront extraction is thus a way of describing properties of the surface currents and, in some cases, decomposing them into phasefronts of different directions. From such a representation, partial information about oscillations of the currents can be obtained, in particular phasefront directions k(m) which are not the same as directions of the currents. IV. T HE U SE OF BASIS F UNCTIONS This section explains how data needed from the MoM are extracted. This is a crucial step for obtaining estimated currents J, and components of currents, such as J and J . It is not just a technicality involving basis functions. Some additional treatment of data must also be done to avoid unneccesary numerical errors. Our need is that of evaluating currents J(r(u)) along a curve (u) from MoM or from FMM, where we expect data from FMM to be even more noisy. These currents will in subsequent sections be used for modelling or extrapolation (to higher frequency) of surface currents. A. The RWG Basis Functions in MoM The so-called RWG basis functions are briey described below. They work on a triangulated body and provide us with the behaviour of the surface currents. We use them also for the nonstandard task of vector eld decomposition into nxHparallel- and cross-currents along the streamlines, as presented below. The Method of Moments yields an estimate of the current density J on q:th triangle q using the Rao-Wilton-Glisson (RWG) basis functions, [22]. The direct output3 of MoM is just the coefcients jm of the RWG-basis functions fm (r), J(r) =
m

position of the triangle node located opposite to a common edge by r3 (m). In principle, the coefcients of the RWG-basis functions may be used for calculating estimates of J everywhere on each triangle q . By construction, summations of RWG-basis functions produce currents J in the tangent plane of each local triangle. Thereby, we replace the tangents to the JP O current streamlines, T, by the approximation Tp , where Tp is T projected orthogonally onto the local triangle along the triangle normal. Then both J from the RWG-basis functions and Tp are in the plane spanned by the edges of the local triangle. The cross currents are then calculated as: J = J (J Tp )Tp
3

(31) (32)

J(r) =
m=1

jm (r r3 (m)) em /(2Aq ) r q

where jm are the scalar coefcients of currents (from MoM) for m = 1, 2, 3, at the positions r q . The scalar em = 1 and Aq is the area of triangle q . For a given triangle, one must assign arbitrary signs em for the direction of the current ow across the m:th edge. Then, the edge of the neighbouring triangle automatically obtains the opposite sign em . Data must be provided with a list of em values over the triangles obeying the sign rule. Analogously, for the parallel component: J = (J Tp )Tp . (33)

In the experiments reported below, we have run MoM with a tessellation of 11.500 triangles, on a unit sphere. Up to k = 30, this represents at least 8 elements per wavelength. Frequencies higher than k 60 are expected to lead to rather poor currents, not resolving wavelength, with this amount of triangles. B. Avoiding Unreliable Currents Observation 6: In practice the RWG basis functions do not yield reliable values for J close to triangle nodes. This is illustrated in Fig. 7(a), and it is intuitively clear, since neighbouring triangles besides the four triangles involved in the sums of Eq. 32 do not correctly inuence values close to triangle nodes. In particular, the component J is quite sensitive, and tends to be more erroneous close to triangle nodes. This should maybe not come as a total surprise, since the computed current J = m jm fm is not smooth. More marked errors, when attempting to extract vector components such as J and J close to triangle nodes, are expected. For that reason, we form a hexagon of the triangle edge midpoints, see Fig. 7(b), given triangle nodes r1 , r2 , r3 : mij = (ri + rj )/2, i, j = 1, 2 7 i, j = 2, 3 i, j = 1, 3 (34)

jm fm (r),

r q

(30)

Given a triangle q , the index m runs over the three possible common edges to the neighbouring triangles, thus m = 1, 2, 3. The coefcients jm are associated with the edges of the triangles. We assume here a Finite Element type triangulation where two neighbouring triangles share only one common edge, thus share one jm value. For each local m, denote the
3 MoM yields complex phasors. We use the real part of the coefcients in the experiments unless stated otherwise. The imaginary part of the surface currents tells us what the pattern looks like quarter of a period later.

(a) RWG basis functions are not reliable close to triangle nodes. Fig. 7. regions.

(b) Hexagons of condence.

and analogously for J which is less crucial to smooth. Central averages are also possible. Here both j and j are integer indices for the subset of positions which has nonzero condence. For simplicity, no smoothing takes place at the end-points. After these formulas have been applied, we, as mentioned, linearly interpolate data so that the original sample points t , t = 1, 2, ..N are associated with these interpolated values. The number of sampling points on a curve is equal to N , whereas j = 1, 2, ..Nj , where Nj is the number of condent sampling points on that curve. D. Use of Symmetric Data mixing For roundish bodies that are symmetric (e.g. circles, ellipsoids, hyper-ellipsoids, etc.) one may use two symmetric streamline curves (u; ). (We denote with an extra argument indicating the chosen streamline = Const.) The point is that the data from MoM will not be the same on either side of the body, because MoM does not ensure symmetry in the estimated coefcients jm and the triangle tessellations covering the level curves = 0 are not identical. In this case, it is possible to use the condence formulas twice, rst for (u; 0 ) obtaining a condence vector conf1 RN and, secondly, another condence vector for (u; 0 ): conf2 RN , where N = number of sample points on the two curves. Let the index = 1, 2, ...N . Let J(r+ ) be the value from = 0 and let J(r ) be the value from the symmetric position on = 0 . The following simple data fusion procedure we term mixing: Denition 7: If conf1 = conf2 = 1 J = 0.5 (J(r+ ) + J(r )) else if conf1 =1 and conf2 = 0 J = J(r+ ) else if conf1 =0 and conf2 =1 J = J(r ), (42) (41) (40)

Errors when using the RWG basis functions and (b) condence

and midpoints connecting the triangle barycenters rc with each of the triangle nodes m = (rc + r )/2, = 1, 2, 3 (35)

We dene the inside of this hexagon to be data with condence equal to 1, and points outside the hexagon to be of zero condence. Hence: confidence(J(r)) = 1, confidence(J(r)) = 0, r Hexq q otherwise (36) (37)

where J(r) is the RWG evaluation of Eq. 32 and Hexq is the q:th hexagon corresponding to the q:th triangle q . This means that we only calculate J(r) values with nonzero condence, when r traverses the positions of the curve . As indicated in Fig. 7(b), long J (u) vectors often arise in the noncondent regions of the triangles. The omitted values of zero condence are replaced by values linearly interpolated from nearest values of condence equal to 1. J and J should be interpolated separately. But before the linear interpolation, some data smoothing takes place, explained in the next subsection. In all our experiments only J has been smoothed, interpolated and subjected to the condence procedure. Data for J did not seem to be in need of much improvement. Some minor jaggedness does arise, but only locally. If smoothing can be avoided, less articial errors are introduced. C. Data Smoothing Since the data for the component J () estimated from RWG-basis functions is of fairly poor quality, in particular for higher frequencies k, we also perform slight smoothing to reduce the noise level. The following smoothing (almost central averages) is done in our experiments after omitting non-condent data (but before the linear interpolation) for J : j J (j ) = = (j1 + j+2 )/2 ((J )j1 + (J )j+2 )/2 (38) (39) 8

else omit data (and linearly interpolate from condent data as before). This procedure substantially improves the quality of the estimated J and J . For nonsymmetric data other similar techniques can be used, such as using adjacent streamlines. E. Improvement of MoM The above mentioned procedure for symmetric roundish bodies actually is of some interest in itself. The mixing actually can be used as a post-processing tool in order to improve the MoM or the FMM currents. V. DATA A NALYSIS Here, data for a unit radius sphere are analyzed, by running MoM for low and medium range coefcients, varying k from 10 to 29. Since k > 2 6 the wave pattern performs more than one wavelength when passing the spherical obstacle.

The reference models for J and J may be modied as follows: J model = RI 2Y cos(kR sin ),
model J = RI Y sin(kR sin ) cos , I

TABLE I RI=1,2 (k) AND RI=1,2 (k) : M ODIFYING COEFFICIENTS OVER FREQUENCY k. I LLUMINATED SIDE (I = 1), SHADOW SIDE (I = 2). = 3 (Lit) R2 (Shadow) (Shadow) -

(43) (44)
k 10 19 29 k 10 19 29 k 10 19 29 k 10 19 29 k 10 19 29 k 10 19 29 R1

where R are examples of modifying coefcients. The superscript I indicates what interval of the curve parameter we refer to, the crudest possible interval decomposition being the intervals corresponding to the shadow and illuminated side, respectively, I = 1, 2. We omit the superscript except when specically distinguishing various intervals. Modifying coefcients may be estimated from MoM data both for J and J , both being functions of given some PO streamline = Const. By estimating these coefcients for various frequency values k, we see on one hand how large the systematic deviations are from the reference models, and on the other hand, collect information useful for extrapolation. We present the estimation results rst, and describe how the modifying coefcients have been calculated, only thereafter. A. Estimation of Modifying Coefcients In Table I, we have, for a unit sphere, estimated R (k) and R (k) for k = 10, 19, 29 for three different streamlines: = /3, = /4 and = /8. The number of sample points is N = 200. The sampling is equidistant in the curve parameter () and it is convenient to refer to an integer-valued curve parameter: t = 1, 2, ...N . For k=29 cross current data are quite noisy and the earlier described mixing has been used, also. When using mixing the table value R1 (29) = 0.71 for = /8 became 0.85. Hence, there is no veried downward trend over k. B. Shadow Side The second column of Table I shows that, in contrast to the lit side, there seem to be clear trends over k for the coefcients, such as decreasing cross currents causing R2 (k)= 1.10, 0.91, 0.62 for k = 10, 19, 29. It should be noted that the t for J around the shadow boundary, the equatorial zone, is poor, because the frequency for the peaks of J does not follow k. Here, is an exception to the rule of thumb that k and the incident H-eld governs the frequency of the variation of (components) of surface currents. The equatorial parts of the streamlines , in particular on the the shadow side, are not well-modelled by the reference models, nor by the modied ones, Eqs. 43- 44. C. Matching Peaks and Troughs Here, some more details are given as an explanation how the modifying coefcients of the previous section, concretely were calculated. We have chosen a multiple scale peak matching procedure. The motivation for peak matching procedure instead of a least square t, is that, for the cross currents, it is not easy to discern a clear correlation between the reference model and 9

, RI

0.54 0.35 0.32 R1 (Lit) 0.60 0.60 0.58 R1

R2

= 4 (Lit) R2 (Shadow) (Shadow) -

0.71 0.51 0.49 R1 (Lit) 0.82 0.83 0.85 R1

R2

= 8 (Lit) R2 (Shadow) 0.89 1.00 1.11 (Shadow) 1.10 0.91 0.62

0.96 0.78 0.76 1 (Lit) R 0.79 0.78 0.71

R2

the MoM components, see an example of a cluster plot in Appendix C. Yet, they do have qualitatively similar shape, peaks and amplitudes. Ref The reference model functions J and J Ref dened by Eqs. 26 and Eqs. 23 24, respectively, have easily dened local maxima and minima, whereas J and J from MoM uctuate a lot with many spurious local minima and maxima. A subset of these are to be pairwise associated with the easily dened local maxima and minima of the reference model. A least squares t of a line would require all the pairs of Ref (J , J ) to make a global t of the slope of the line, see Appendix C, thus increasing the inuence of the noise for estimating R . Consider the set of local maxima or minima in the disRef cretized vectors of {J (ui )}i=1...N and {J Ref (ui )}i=1...N , and denote the corresponding positions of these local extrema by argmin and argmax. These positions are subsets of t = 1, 2...N . With this notation, a list of t positions for the extrema tM and tm for the reference model of cross currents can be written: tRef M tRef m
Ref = [argmax({J })]M , M = 1, 2, .., NM

(45) (46)

Ref [argmin({J })]m ,

m = 1, 2, .., Nm

where NM and Nm are the numbers of local maxima, local Ref minima respectively for J and N is the number of samples

TABLE II RI=1,2,3 (k) AND RI=1,2,3 (k) : M ODIFYING COEFFICIENTS OVER FREQUENCY k. I LLUMINATED SIDE (I = 1), SHADOW SIDE (I = 2),
EQUATOR

procedure consists of two sets of extrema: Mp = matched extrema, p = 1, 2, ...(Nm + NM ) (52)

(I = 3).
3

k 10 19 29 k 10 19 29 k 10 19 29 k 10 19 29 k 10 19 29 k 10 19 29

R1 (Lit) 0.53 0.43 0.45 R1 (Lit) 0.86 1.17 0.79 R1 (Lit)

= R2 (Shadow) 0.69 0.79 0.83 (Shadow) 1.52 1.23 0.56 R2 = 4 (Shadow) 1.20 1.05 1.29 (Shadow) 1.77 0.77 R2 = 8 (Shadow) 1.17 1.13 1.36 (Shadow) 2.33 1.74 0.70

R3 (Equator) 0.30 0.32 0.28 (Equator) 0.33 0.25 0.39 R3 (Equator) 0.63 0.46 0.43 (Equator) 0.79 0.43 0.52 R3 (Equator) 0.89 0.72 0.77 (Equator) 0.66 0.42 0.51

MRef = reference model extrema, p = 1, 2, ...(Nm + NM ) p (53) Finally, the modifying coefcients are calculated as: RI (k) =
P |Mp | P p Ref . p |Mp |

pI

(54)

R2

R3

J is treated analogously. As a byproduct, the positions tM for the chosen peaks, may also convey information on phase shifts in the sinusoidal patterns. D. Modifying Coefcients for Three Intervals It makes more sense to use three intervals for the modifying coefcients, since J and J often exhibit different deviations from the reference models in the equatorial regions (I = 3), the deep shadow region (I = 2) and the central lit region (I = 1). The equatorial regions are associated with relatively lower magnitudes of R (k) and R (k), normally without strong trends. The three intervals used in Table II are S2 = { : [ + /6, 2 /6]} (deep shadow), (56) and nally [0, 2] S1 S2 (equatorial regions). The interval termed Shadow in Table II means shadow zone, excluding the equatorial parts around the shadow boundary, and the same comment can be made for the lit zone. The modifying coefcients are less exact on the shadow side for J since, as mentioned, the frequency of the oscillations is not correctly modelled. Then, the peak matching procedure will for some items in the sum of the terms in the numerator of Eq. 54 cause some bias in the resulting ratio R (k). The - in Table II means that the automatically calculated R (10) was erroneous due to poor peak matching. The peak matching is intended for the situation when the number of peaks and troughs of both signals are the same. VI. C ONCEPTS FOR M ODELLING AND E XTRAPOLATION In this and following sections all vectors are real. Hence, J means in fact JRe and we have dropped the superscript Re . We focus on comparisons between estimated surface currents from MoM, and modelled surface currents by three models: (a) standard PO, (b) our Reference model dened by Eq. 26, Eq. 23, Eq. 24 and (c) our modied Reference Model, Eqs. 43 44. A rst test is to investigate the behaviour on a sphere, which we use for illustrating our technique. Consequently, three errors (deviations from MoM) are to be compared. Let Jmodel be the estimated currents from Eqs. 4344, using some R and R coefcients, and MoM currents and reference model currents, as before, J and JRef , respectively. We denote by [J] the concatenated vector containing all the vector currents J over the streamlines of dimension 3N . To 10 S1 = { : [/6, /6]} (lit zone), (55)

0.65 0.66 0.62 R1 (Lit) 1.09 1.23 1.49 R1 (Lit)

R2

R3

0.87 0.90 0.82 1 (Lit) R 1.12 1.48 1.09

R2

R3

of J . The positions for the extrema of J from MoM are: tM tm = =

[argmax({J })]M , M = 1, 2, .., NM [argmin({J })]m , m = 1, 2, .., Nm

(47) (48)

where M in Eq. 47 runs over all candidate maxima with NM NM , normally. Let us now match peaks. Troughs (local minima) are matched analogously, in the same manner. Peaks in J are Ref matched with peaks (peak = 1, 2, ...NM ) in J by choosing those M values within a search distance S1 as N S1 = 2(NM + Nm ) which satisfy the inequalities: peak : |tM tRef | < peak |tM tRef | peak S1 (50) (51) (49)

< 1.5 S1 .

The second inequality is used only when no tM , for the given peak, satises the rst inequality. When several peaks satisfy, e.g., the rst inequality, the highest peak is chosen. If no peak satises the rst inequality, the second inequality applies, and the highest peak there is chosen. Some obvious special cases must also be taken care of. The quantities tM and tRef are peak dened by Eqs. 47 and 45. The output from this matching

avoid the use of absolute numbers, the following socalled performances are dened: model Ref = = 1 1 [Jmodel ] [J] [JP O ] [J] ] [J] [JP O ] [J] [J
Ref

x 10

Extrapolation of MoM, Parallel component, k=60 Phi=0.3927

100 (%) (57) 100 (%) (58)

where JP O is the PO current of Eqs. 12. We use PO as a reference for model errors. In order to compute the error, it is not necessary to use J as a vector in IR3 . Considering J = J T + J nc (59)

100

200

300 u

400

500

600

700

8 6 4 2 0 2 4 6

x 10

Error Model to MoM

Error Model +stand. dev. Model stand. dev. Model

where T, nc IR3 which form an orthonormal basis, we obtain the error vectors in PO: E P O = J JP O and the error in the reference models E Ref = J JRef as E E E
PO Ref

= = =

model

(J JP O )T + (J 0)nc (60) Ref J Ref )T + (J J )nc (J (61) model model (J J )T + (J J )nc (62) (63)

100

200

300 u

400

500

600

700

Then the Euclidean norm of E model is,

model 2 ||E model (u)||2 = (J J model )2 + (J J )

Fig. 8. Case ka=60, J for = /8 using extrapolated RI=1,2,3 (k) I=1,2,3 (k) compared to MoM. Note that we do not extrapolate MoM and R currents in themselves but use low frequency values k = 10, 19, 29 to estimate the modifying coefcients for ka = 60. The deviations from MoM, called Error in the legend, is computed using Eq. 64.

due to orthonormality. Thus, the scalar error is determined solely by E model = J J model (64)
model model E = J J . PO P E O Ref Ref E .

by averaging the values in Table II, column-wise in rows 13 to 15. The other coefcients used for the same were: I=1,2,3 = 1.23 0.58 0.47 R ( = /8) (68)

(65)

Analogously for E , and E , For the global error, all the Euclidean errors must be summed, yielding the norm of [E model ] as ||[E model ]|| =
j

||E model (uj )||2 .

(66)

In our experiments, we use a smoothed version for J , Eq. 39. VII. E XTRAPOLATION One idea is to use the estimated coefcients R (k) and R (k) for extrapolation to higher frequency k. Denote the two coefcients by Rn (k), n = , . By extrapolation two things may be meant. One possibility is to calculate some average Rn (k)value (averaging over k) in Eq. 43 or Eq. 44, above, and use that approximate Rn value for a higher k value in the equations. The other possibility is to extrapolate Rn (k) as a function of k. The latter procedure is of course preferable. But in some cases, trends in Rn are so hard to calculate that it is probably safer to just use an average Rn value. The following extrapolation from k = 10, 19, 29 to k = 60 serves to illustrate the approach. The following averages were used for the nxH parallel components: RI=1,2,3 = 0.87 1.22 0.80 ( = /8) (67) 11

by averaging column-wise over k, and by a very crude trend extrapolation for the second value: 0.58 = 1/4 R (10), noting that R (29) = R (10)/3. The downward trend of R (k) is signicant, in this case. MoM for cross currents is inherently noisy, growing worse for higher frequencies, so we used mixing (Def.7) for stabilizing the values somewhat. We thus use improved MoM. When calculating the reference model and the modied reference model using these coefcients for k = 60 in the formulas, the following deviations from MoM arose: [Jmodel ] [J] Ref [J ] [J] [JP O ] [J] = = = 0.0406 0.0466 0.0557 (69) (70) (71)

and the previously dened performance value in percentage relative to the PO technique for the whole streamline were model = 27%, and for the lit side only: model = 14%, Ref = 4%. (73) Ref = 16%, (72)

Hence, the extrapolation yields about 30 % improvement over PO, in modelling surface currents and does better than the reference model. The modifying coefcients have been modelled as piecewise constant. This is of course not necessary. One may as

well model smooth transitions between different intervals, for example by using the constant levels in the interior of the intervals and use splines for connecting these constant levels in adjacent intervals. This is in our opinion a reasonable but not crucial enhancement of the approach. Furthermore, when extrapolating to the double electric size (using values from ka=10,19,29 ka=60), Fig. 8 shows that the resulting error, cf. Eq. 64 using the modifying coefcients is of the same magnitude as in many of the modelling cases, to be presented, see for instance Fig. 21. The extrapolation does not introduce more errors. Next section will go deeper into the modelling. VIII. R ESULTS OF M ODELLING The reference models can be directly applied as models of surface currents without any knowledge of data from extracting algorithms. In that sense, they remind of PO currents. Thus, it may be of interest to see whether there is any gain in using the reference models Eq. 26 and Eqs. 23 24 instead of JP O . Another alternative is that of using some data from MoMresults for some kvalues and use such data for modifying the reference models exploiting the modifying coefcients Rn which arise from comparing MoMdata with the reference model for the chosen kvalues. In the experiments, to be presented below, we have chosen k = 10, 19, 29 for calculating I=1,2,3 the three-interval modifying coefcients: Rn (k). The rst issue to clarify is whether these modied reference models Eqs. 43, 44, using Table II, imply a visible improvement over the original reference models. In Table III, we compare on one hand reference model currents JRef and J, and on the other Jmodel and J. We Ref model measure the advantage of using J or J for mimicking J, compared to JP O , using the performances dened by Eqs. 5758. This technique is meant to be used for other kvalues and other values, by dening functions RI (k; ) using sparse n data such as, e.g., k = 10, 19, 29 and = /3, /4, /8 for dening them. A. Modelling of Currents on Sphere and Ellipsoid For a sphere, the results along a PO current streamline are displayed in Fig. 9-10 and compared to MoM, for = /8. Here, we note how non-smooth the results obtained from MoM are. We present results for different kvalues, by writing kavalues, where a is the radius of the sphere. When using our rst reference model, good agreement for the parallel component J Ref is obtained at both low and higher frequency. The nxH-parallel component, like the JP O current, displays error which amplies close to the shadow boundaries located on the plots Fig. 9 and Fig. 11 around the end points Ref of the lit interval. The perpendicular component J is more difcult to model directly, as seen in Fig. 10(a) for low frequency ka = 10 and in Fig. 10(b) for higher frequency. In the shadow region, the reference model for J Ref is in phase with the MoM current, as seen Fig. 11(a) at ka = 10 and Fig. 11(b) at ka = 19 but at the south pole the amplitude 12

TABLE III P ERFORMANCE WITH R EFERENCE MODEL AND E XTRACTION M ETHOD I=1,2,3 COMPARED TO PO, OVER k USING R (k) AND RI=1,2,3 (k). = 3 Extract. (Lit) 73% 74% 69% = 4 Extract. (Lit) 51% 58% 47% = 8 Extract. (Lit) 11% 30% 17%

k 10 19 29 k 10 19 29 k 10 19 29

Extract (All) 55% 63% 63% Extract (All) 51% 48% Extract (All) 36% 36% 41%

Ref. (All) -5% 8% 0% Ref. (All) 21% 21% 15% Ref. (All) 33% 26% 28%

Ref. (Lit) -2% -1% -2% Ref. (Lit) 0% 0% -1% Ref. (Lit) -2% 2% -2%

Ref should in fact be higher. For J , modelling of phase is needed since the reference model is out of phase, as seen in Fig. 12. For an ellipsoid

(x, y, z) = (cos cos , 2.0 sin , cos sin )

(74)

the streamline is also an ellipse and the reference model for J Ref works ne and no new pattern is introduced, see Fig. 13. This is a good sign that our technique can be applied to more general roundish bodies. B. Performance Comparisons In Table III, a series of experiments for frequencies k = 10, 19, 29, for a sphere, are presented, in terms of performances (Eqs. 5758), relative to Physical Optics currents. Of course, both J and J inuence the results, whereas the latter component somewhat dominates by being roughly 34 times larger in magnitude. Consider the reference model, rst. When including the shadow region, the reference model (column Ref. All, Table III) is a 20% to 30% better model than PO, for = /4 or = /8. Ref The last column of Table III tells us that J estimates have not contributed to improve the t with MoMdata, compared to PO. The dominant part of the signal J (u), is identical for the lit side since JP O (u) = J Ref (u), so considering only J (u) the performance percentage would be exactly zero. In this case, the cross currents components are aficted with too much errors to improve over PO. Performance percentage for the reference model for the cross current is about 2%. In general, for the coefcient model, if looking at the error plots in Appendix B, for instance in Figures for = /4 k=19, or, = /8, k = 29, we have quite a good model for J (u) away from the equatorial zone, with error magnitudes around 0.3 103 , but phase errors (threefour times larger) in the sinusoidal signal in that zone outweigh what we gain in the central lit part. The model with coefcients gives only a slight improvement over PO, for the same reason.

x 10

Parallel Current Component: k=10 cut Phi=0.3927 6 From MoM Modeled

x 10

Parallel Current Component: k=19 cut Phi=0.3927 From MoM Modeled

4 2 2 |J| |J| 0 2 2 4 6 0 50 100 150 200 250 0

50

100

150

200

250

(a) Case ka=10, = /8 (lit side). Fig. 9.

(b) Case ka=19, = /8 (lit side)

Reference Model (no modifying coefcients) for the parallel component J Ref compared to MoM. Good agreement in phase.

1.5

x 10

Perpendicular Current Component: k=10 cut Phi=0.3927 2 From MoM Modeled 1.5

x 10

Perpendicular Current Component: k=19 cut Phi=0.3927 From MoM Modeled

1 1 0.5 0.5

|J|

|J| 0 50 100 150 200 250

0.5 0.5 1 1 1.5

1.5

50

100

150

200

250

(a) Case ka=10, = /8 (lit side) Fig. 10.

(b) Case ka=19, = /8 (lit side)

Ref Reference Model for the perpendicular component J compared to MoM. Good agreement for the amplitude but slightly out of phase.

For the J (u)part we have excellent t in most cases, in particular when using the Jmodel currents. This can be seen in Figures in Appendix B corresponding to k = 19 and k = 29. A general observation is that when gradually decreasing , i.e., moving towards the equator, the performance of the reference models or modied reference models increases substantially. This is so, because PO is gradually a poorer model (of the parallel component) the closer to the shadow boundary in the lit zone the surface current measurements are made. At = /8, close to the EastWest meridian in the lit zone, it is not equally easy to improve over PO. For the East-West meridian itself at = 0, we suggest using J = 0, Eq. 16, as reference model. And J = 0 also when /2. In general, in the equatorial region, the model signals are out of phase or even in counter phase compared to J , which deteriorates the achieved good result in non-equatorial 13

regions, both for the reference model and when modied with coefcients. C. Angle Measurements The total mean angle on one streamline = /8 (lit region only) over frequency k for the PO, the reference model are shown in the next table. It shows that the reference model gives improved surface current directions on the lit side compared to the PO-current directions. On the shadow side there is nothing to compare since PO currents are zero and thus lacking directions. D. General Comments It should be kept in mind that what Table III measures is how much models deviate from MoM, rather than the error relative to the true solution. For this the Mie solutions should be used.

x 10

Parallel Current Component: k=10 cut Phi=0.3927 (shadow) 5 From MoM Modeled

x 10

Parallel Current Component: k=19 cut Phi=0.3927 (shadow) From MoM Modeled

1 |J| |J| 0 50 100 150 200 u 250 300 350 400

50

100

150

200 u

250

300

350

400

(a) Case ka=10, = /8 (shadow) Fig. 11.

(b) Case ka=19, = /8 (shadow)

Reference Model (no modifying coefcients) for the parallel component J Ref compared to MoM.

x 10

Perpendicular Current Component: k=10 cut Phi=0.3927 (shadow) 2 From MoM Modeled

x 10

Perpendicular Current Component: k=19 cut Phi=0.3927 (shadow) From MoM Modeled

1.5

1.5

0.5

0.5

0.5

0.5

1.5

1.5

50

100

150

200 u

250

300

350

400

50

100

150

200 u

250

300

350

400

(a) Case ka=10, = /8 (shadow) Fig. 12.

(b) Case ka=19, = /8 (shadow)

Ref Reference Model for the perpendicular component J in shadow compared to MoM. Good agreement for the amplitude but out of phase.

TABLE IV M EAN ANGLE AND PERFORMANCE PERCENTAGE . ka PO Ref. model = /8 Perf. Ref-PO 10 11.51o 12.93o +12% 19 23.85o 17.48o +36% 29 25.33o 22.35o +13.0%

Since MoM is considered to be a fairly good algorithm, this may not be so serious, but nevertheless some of the variation of MoM are deviations from the true solutions. Locally, some of the jaggedness of the MoM solution does not replicate the true solution. Hence, some of the errors of Table III includes implicitly variations in slightly erroneous MoMdata. Good results for J , in particular for higher frequencies 14

are obtained for the coefcient model, as seen in Figures in Appendix B, for k = 19, 29 for all angles . A general lesson learned from Section B is that a better reference model for J (u) in the shadow zone is needed, because the frequency k apparently does not coincide with the actual local spatial frequency. Moreover, phase errors of the uctuations of J (u) in the lit zone should be possible to avoid with a somewhat improved reference model. Generally, the modifying coefcients are really useful only to the extent that they are stably dening some fairly constant value or some simple trend function over k, because it is desirable to infer Rn (k) values from a few sparse frequencies. In our case, we chose k = 10, 19, 29 because we noted that lower kvalues are associated with strong trends in the scattered E fareld. There is a special dynamics over k for really low frequencies, but we are more

x 10

Parallel component, k= 10, Phi=0.3927, ellipsoid 6 from MoM Modeled

x 10

Parallel component, k= 20, Phi=0.3927, ellipsoid from MoM Modeled

50

100

150

200

250

300

350

400

450

50

100

150

200

250

300

350

400

450

(a) Case ka=10, = /8 (ellipse, lit side) Fig. 13.

(b) Case ka=20, = /8 (ellipse, lit side)

Reference Model (no modifying coefcients) for the parallel component J Ref on an ellipse compared to MoM.

Jmodel

Jcross

Jmom Jpo

Fig. 14.

Mean angle as a measure of the error.

interested in the intermediate frequencies, or higher. Another technique for extrapolation was presented in [37], APE-MoM, where the MoM is used twice, once with standard basis functions and once more multiplying the basis functions by e
i low k(m) (r rm )
high

(75)

where k(m) are extracted phasefront directions, Eq. 28. IX. C ONCLUSION It has been shown that the vector eld decomposition of the surface currents yields a better understanding of charge transport and charge accumulation. The oscillating behaviours of the current have been shown to be strongly linked to that of the incident magnetic Held. The decomposition of the surface currents into nxH-parallel and cross-current vector components yield systematic spatial patterns evolving over frequency closely related to the evolution of the incident magnetic eld over frequency. This is true both on illuminated and shadow side of the body, allowing us to formulate a reference model for each component over an entire roundish body. Currents on the shadow side, wrongly assumed to be 15

zero for all meridians by the Physical Optics, can be modelled by these fairly simple models without relying on a complex procedure such as creeping rays in the GTD formulation. However, for the TE-meridian, zero PO currents is a good approximation, as seen in Appendix D. As a side product, the perpendicular component J can be interpreted as a source of error in the PO current. Thus, the model for the cross-currents could be extended to a numerical procedure for automatic control of the error growth in the PO technique. Although we have studied a sphere, the approach should be applicable to roundish smooth bodies rather generally. This is because the nxH-parallel component J (u) can be well modelled as a function of the curve parameter only along PO streamlines. A procedure to determine the PO streamlines has been given and illustrated on the complex shape of the Eikon UAV. The reference model for the second component, the cross currents J , has been found to be not sufcient. Although it qualitatively captures the oscillating behaviour of that component of the current, the approximation is aficted with phase errors. In the lit zone, we can get phase shifting errors, whereas the shadow zone also gives rise to fewer oscillations of J . The phase of J must be better modelled but better understandings of the underlying physics are needed for this endeavor. Nevertheless, the modelling of J is less important than that of J , since the latter is about 4 times larger as a rule. Tentative extrapolation was achieved. This was done by the use of reference models modied by coefcients calculated from low frequency data from MoM. The coefcients act like diffusion coefcients applied to the amplitude of each component. A surprising result was that, in the neighborhood of the vertices, the MoM current used to compute the coefcients yield unreliable values. Corrections and smoothing of these values have been implemented. Trends over frequency in the coefcients have been studied, allowing us to extrapolate the

current to higher frequency. Our experiments show that we could easily double the electrical size of the problem. The whole approach presented in this study, as a stand alone procedure, offers an alternative way to derive high frequency approximate currents. As a processing tool plugged into an MoM-PO hybrid solver, it could offer a way to control the error of the Physical Optics current, with the potential to efciently improve the accuracy of the MoM-PO hybrid solver. ACKNOWLEDGMENT The authors would like to thank Jesper Oppelstrup for his thoughtful recommendations and suggestions. Financial support has been provided by NADA, KTH, PSCI and the National Aeronautical Research Program (NFFP) within the General ElectroMagnetic Solvers (GEMS) and Signature Modeling and Reduction Tools (SMART) projects. R EFERENCES
[1] R.F. Harrington, Field Computation by Moment Methods, New York: Macmillian, 1968. [2] J.J. Bowman, T.B.A. Senior, P.L.E. Uslenghi Electromagnetic and Acoustic Scattering by Simple Shapes, HPC, 1987. [3] Joseph B. Keller, Geometrical Theory of Diffraction, Journal of the Optical Society of America, Vol. 52(2), pp. 116-130, February 1962. [4] R.G. Kouyoumjian and P.H. Pathak, A Uniform Geometrical Theory of Diffraction for an Edge in a Perfectly Conducting Surface, Proc. IEEE, Vol. 62, pp. 1448-1481, November 1984. [5] D.A. McNamara, C.W.I. Pistorius, J.A.G. Malherbe Introduction to the Uniform Geometrical Theory of Diffraction, Artech House, London, 1989. [6] G.A. Thiele and T.H. Newhouse, A hybrid technique for combining moment methods with geometrical theory of diffraction, IEEE Trans. Antennas Propagat., vol. AP-23, pp. 62-69, Jan. 1975. [7] W.D. Burnside, C.L. Yu and R.J. Marhefka, A technique to combine the geometrical theory of diffraction and moment method, IEEE Trans. Antennas Propagat.,pp. 551-558, July 1975. [8] S. Se, Ray Tracing Tools for High Frequency Electromagnetics Simulations, Licentiate thesis TRITA-NA-0314, KTH, May 2003. [9] D.P. Bouche, F.A. Molinet and R. Mittra, Asymptotic and Hybrid Techniques for Electromagnetic Scattering, Proceedings of the IEEE, vol.81(12), Dec. 1993. [10] C.S. Kim and Y. Rahmat-Samii, Low prole Antenna Study Using the Physical Optics Hybrid Method (POHM), IEEE Antennas and Propagation Society Int. Symp., vol.3, pp. 1350-1353, June 1991. [11] U. Jakobus and F. M. Landstorfer, Improved PO-MM Hybrid Formulation for Scattering from Three-Dimensional Perfectly Conducting Bodies of Arbitrary Shape, IEEE Transactions on Antennas and Propagation vol. 43, pp. 162169, February 1995. [12] R.E. Hodges and Y. Rahmat-Samii, An iterative current-based hybrid method for complex structures, IEEE Trans. Ant. Propagat. 45 1997. [13] F. Obelleiro, J.M. Taboada, J.l. Rodriguez, J.O. Rubinos and A.M. Arias, Hybrid moment-method physical optics formulation for modeling the electromagnetic behavior of on-boaerd antennas, Microwave Opt. Tech. Letter 27,pp. 88-93, 2000. [14] J. Edlund, A Parallel, Iterative Method of Moments and Physical Optics Hybrid Solver for Arbitrary Surfaces, Licentiate Thesis, Uppsala University, 2001. [15] F.X. Canning, The impedance matrix localization (IML) method for moment-method calculations, IEEE Transactions on Antennas and Propagation, vol. 32, pp. 1830, 1990. [16] B.Z. Steinberg and Y. Leviatan, On the use of wavelet expansions in the method of moments, IEEE Transactions on Antennas and Propagation, vol. 41, pp. 610619, 1993 [17] C. Su, T.K. Sarkar, Adaptive Multiscale Moment Method (AMMM) for Analysis of Scattering from Three-Dimensional Perfectly Conducting Structures, IEEE Transactions on Antennas and Propagation, vol. 50, N0. 4, April 2002. [18] W. Wiscombe, Improved Mie scattering algorithms, Appl. Opt., Vol. 19, pp. 1505-1509, 1980

[19] N. Engheta, W. Murphy, V. Rokhlin and V. Vassiliou, The fast multipole method (fmm) for electromagnetic scattering problems, IEEE Trans. on Antennas and Propagat., vol 40, pp. 634-641, June 1992. [20] L. Greengard and V. Rokhlin, A fast algorithm for particle simulations, Journal of Computational Physics, 73(2):325-348, Dec. 1987. [21] Martin Nilsson, A fast multipole accelerated block quasi minimum residual method for solving scattering from perfectly conducting bodies, Ant. and Propag. Society International Symposium No 4, 2000. [22] S. M. Rao, D. R. Wilton, and A. W. Glisson, Electromagnetic scattering by surfaces of arbitrary shape, IEEE Trans. on Antennas and Propagat.,vol.AP-30(3), pp. 409-418, May 1982. [23] D. K. Cheng, Field and Wave Electromagnetics, Addison Wesley, Second Edition, 1998. [24] I. J. Gupta, W. D. Burnside, A Physical Optics Correction for Backscattering from Curved Surfaces, IEEE Transactions on Antennas and Propagation, vol. 35, N0. 5, May 1987. [25] A. C. Ludwig. Computation of radiation patterns involving numerical double integration. IEEE Transactions on Antennas and Propagation, pages 767769, Nov. 1968. [26] Glenn D. Crabtree. A numerical quadrature technique for physical optics scattering analysis. IEEE Transactions on Microwave Theory and Techniques, 27(5), Sep. 1991. [27] F.J.S Moreira and A. Prata. A self-checking predictor-corrector algorithm for efcient evalaution of reector antenna radiation integrals. IEEE Trans. on Ant. and Propagat., 42(2):246254, Feb. 1994. [28] S. Se and J. Oppelstrup, Physical optics and NURBS for RCS calculations, Proc. EMB 04 Computational Electromagnetics Methods and Applications, pages 90-97. SNRV, October 2004. [29] J. M. Taboada, F. Obelleiro, and J. L. Rodrgeuz. Improvement of the hybrid moment method-physical optics method through a novel evaluation of the physical optics operator. Microwave and Optical Technology Letters., 30(5):357363, 2001. [30] J. M. Taboada, F. Obelleiro, J. L. Rodrgeuz, I. Garcia-Tunon and L. Landesa, Incorporation of Linear-Phase Progression in RWG basis functions, Microwave and Optical Technology Letters., 44(2), Jan. 2005. [31] M. Djordjevic and B.M. Notaros, Higher Order Hybrid Method of Moments-Physical Optics Modeling Technique for Radiation and Scattering from Large Perfectly Conducting Surfaces, IEEE Transactions on Antennas and Propagation, Feb 2004. [32] K.R. Aberegg and A.F. Peterson, Application of the Integral EquationAsymptotic Phase Method to Two-Dimensional Scattering, IEEE Trans. on Antennas and Propagat., vol. 43, pp. 534-537, May 1995. [33] Z. Altman and R. Mittra,A technique for Extrapolating Numerically Rigourous Solutions of Electromagnetic Scattering Problems to Higher Frequencies and Thier Scaling Properties, IEEE Trans. on Antennas and Propagat., vol.47(4), April 1999. [34] Z. Altman, R. Mittra, O. Hashimoto and E. Michielssen, Efcient representation of the induced currents on large scatterers using the generalized pencil of function method, IEEE Trans. on Antennas and Propagat., vol.44(1), Jan. 1996. [35] Z. G. Figan, R. Mittra, A. Boag and E. Michielssen, A technique for solving scattering problems at high frequencies, Int. URSI Symp., Seattle WA, June 1994. [36] D. Kwon, Efcient Method of Moments formulation for large conducting scattering problems using asymptotic phasefront extraction, Ph.D. dissertation, Dept. Elect. Eng., The Ohio State University, Colombus, OH, 2000. [37] D. Kwon, R.J. Burkholder and P. H. Pathak, Efcient Method of Moments Formulation for Large PEC Scattering Problems Using Asymptotic Phasefront Extraction (APE), IEEE Trans. on Antennas and Propagat., vol. 49, No.4, April, 2001. [38] L.Takacs, A parabolic equation technique for reducing physical optics shadow boundary errors Antennas and Propagation Society International Symposium, 1990, pp:136 - 139, vol.1, May 1990

16

A PPENDIX A: TE-TM CASES

E H
k

E H E H Etan =0 H

E
k n

H E

TM case: tangent. comp. E flips total tangent H double

TE case: normal comp. H flips total tangent H double tangent. comp. E flips

Fig. 15.

TE, TM: case on a plate.

X Y

E H
k

E H H E Etan =0 H

E
k n

TM

Fig. 16.

TE, TM situations on a sphere.

A PPENDIX B: M ODELLING OF C URRENTS The following gures present a series of experiments where J and J are modelled for various frequencies k and different streamlines = /8, /4, /3. The rst six pictures correspond to the streamline = /8, close to the East-West meridian, for k = 10, 19 and 29. Then, the two next gures represent J and J evaluated on one intermediate streamline = /4, i.e. a section in the middle of the sphere, for one frequency k = 19. Finally, the last four gures represent J and J for two frequencies for k = 19 and 29 on the streamline = /3 close to the forward pole. Each gure contains two pictures, the component J or J and their respective deviations from MoM, called Error in the legends and computed using Eq. 64 and Eq. 65.

17

E=0

H E

E=0

TE

8 6 4 2 0 2 4 6

x 10

Approximations of MoM, Parallel component, k=10 Phi=0.3927

JParMoM JParmodel JParREF JPO

50

100

150

200 u

250

300

350

400

450

3 2 1 0 1 2

x 10

Error REF model to MoM, Error Model with Coefficient to MoM Error Model +stand. dev. Model stand. dev. Model Error REF +stand. dev. REF stand. dev. REF

50

100

150

200 u

250

300

350

400

450

Fig. 17.

Case ka=10, J for = /8 using RI=1,2,3 (k) and RI=1,2,3 (k) compared to MoM and PO.

3 2 1 0 1 2 3

x 10

Approximations of MoM, Perpendicular component, k=10 Phi=0.3927 JPerpMoM JPerpmodel JPerpREF

50

100

150

200 u

250

300

350

400

450

3 2 1 0 1 2 3

x 10

Error REF model to MoM, Error Model with Coefficient to MoM

Error Model +stand. dev. stand. dev. Error REF +stand. dev. REF stand. dev. REF

50

100

150

200 u

250

300

350

400

450

Fig. 18.

Case ka=10, J for = /8 using RI=1,2,3 (k) and RI=1,2,3 (k) compared to MoM.

18

6 4 2 0 2 4 6

x 10

Approximations of MoM, Parallel component, k=19 Phi=0.3927

JPar MoM JParmodel JPar REF JPO

50

100

150

200 u

250

300

350

400

3 2 1 0 1 2 3 4

x 10

Error REF model to MoM, Error Model with Coefficient to MoM

450 Error Model +stand. dev. Model stand. dev. Model Error REF +stand. dev. REF stand. dev. REF

50

100

150

200 u

250

300

350

400

450

Fig. 19.

Case ka=19, J for = /8 using RI=1,2,3 (k) and RI=1,2,3 (k) compared to MoM and PO.

x 10

Approximations of MoM, Perpendicular component, k=19 Phi=0.3927 JPerpMoM JPerpmodel JPerp

REF

50

100

150

200 u

250

300

350

400

450 Error Model +stand. dev. stand. dev. Error REF +stand. dev. REF stand. dev. REF

3 2 1 0 1 2

x 10

Error REF model to MoM, Error Model with Coefficient to MoM

50

100

150

200 u

250

300

350

400

450

Fig. 20.

Case ka=19, J for = /8 using RI=1,2,3 (k) and RI=1,2,3 (k) compared to MoM.

19

6 4 2 0 2 4 6

x 10

Approximations of MoM, Parallel component, k=29 Phi=0.3927

JPar MoM JPar model JParREF JPO

50

100

150

200 u

250

300

350

400

450

6 4 2 0 2 4

x 10

Error REF model to MoM, Error Model with Coefficient to MoM

Error Model +stand. dev. Model stand. dev. Model Error REF +stand. dev. REF stand. dev. REF

50

100

150

200 u

250

300

350

400

450

Fig. 21.

Case ka=29, J for = /8 using RI=1,2,3 (k) and RI=1,2,3 (k) compared to MoM and PO.

1.5 1 0.5 0 0.5 1 1.5 2

x 10

Approximations of MoM, Perpendicular component, k=29 Phi=0.3927 JPerp MoM JPerp model JPerp
REF

50

100

150

200 u

250

300

350

400

450 Error Model +stand. dev. stand. dev. Error REF +stand. dev. REF stand. dev. REF

3 2 1 0 1 2 3

x 10

Error REF model to MoM, Error Model with Coefficient to MoM

50

100

150

200 u

250

300

350

400

450

Fig. 22. Case ka=29, J for = /8 using RI=1,2,3 (k) and RI=1,2,3 (k) compared to MoM. Here, the t between J from MoM and J from the coefcient model (JPerpmodel in the legend) is quite good.

20

6 4 2 0 2 4 6

x 10

Approximations of MoM, Parallel component, k=19 Phi=0.7854 JPar MoM JParmodel JParREF JPO

50

100

150

200 u

250

300

350

400

450 Error Model +stand. dev. Model stand. dev. Model Error REF +stand. dev. REF stand. dev. REF

4 2 0 2 4 6

x 10

Error REF model to MoM, Error Model with Coefficient to MoM

50

100

150

200 u

250

300

350

400

450

Fig. 23.

Case ka=19, J for = /4 using RI=1,2,3 (k) and RI=1,2,3 (k) compared to MoM and PO.

x 10

Approximations of MoM, Perpendicular component, k=19 Phi=0.7854 JPerp MoM JPerp model JPerpREF

50

100

150

200 u

250

300

350

400

450 Error Model +stand. dev. stand. dev. Error REF +stand. dev. REF stand. dev. REF

3 2 1 0 1 2 3

x 10

Error REF model to MoM, Error Model with Coefficient to MoM

50

100

150

200 u

250

300

350

400

450

Fig. 24.

Case ka=19, J for = /4 using RI=1,2,3 (k) and RI=1,2,3 (k) compared to MoM.

21

6 4 2 0 2 4 6

x 10

Approximations of MoM, Parallel component, k=19 Phi=1.0472

JParMoM JPar model JPar REF J

PO

50

100

150

200 u

250

300

350

4 2 0 2 4 6

x 10

Error REF model to MoM, Error Model with Coefficient to MoM

Error Model +stand. dev. Model stand. dev. Model 400 450 Error REF +stand. dev. REF stand. dev. REF

50

100

150

200 u

250

300

350

400

450

Fig. 25.

Case ka=19, J for = /3 using RI=1,2,3 (k) and RI=1,2,3 (k) compared to MoM and PO.

1.5 1 0.5 0 0.5 1 1.5

x 10

Approximations of MoM, Perpendicular component, k=19 Phi=1.0472

JPerp MoM JPerpmodel JPerpREF

50

100

150

200 u

250

300

350

400

450 Error Model +stand. dev. stand. dev. Error REF +stand. dev. REF stand. dev. REF

x 10

Error REF model to MoM, Error Model with Coefficient to MoM

50

100

150

200 u

250

300

350

400

450

Fig. 26.

Case ka=19, J for = /3 using RI=1,2,3 (k) and RI=1,2,3 (k) compared to MoM.

22

6 4 2 0 2 4 6

x 10

Approximations of MoM, Parallel component, k=29 Phi=1.0472

JPar MoM JPar model JPar REF J

PO

50

100

150

200 u

250

300

350

x 10

Error REF model to MoM, Error Model with Coefficient to MoM

400Error Model 450 +stand. dev. Model stand. dev. Model Error REF +stand. dev. REF stand. dev. REF

50

100

150

200 u

250

300

350

400

450

Fig. 27.

Case ka=29, J for = /3 using RI=1,2,3 (k) and RI=1,2,3 (k) compared to MoM and PO.

1.5 1 0.5 0 0.5 1 1.5

x 10

Approximations of MoM, Perpendicular component, k=29 Phi=1.0472

JPerp MoM JPerp model JPerpREF

50

100

150

200 u

250

300

350

400

450 Error Model +stand. dev. stand. dev. Error REF +stand. dev. REF stand. dev. REF

x 10

Error REF model to MoM, Error Model with Coefficient to MoM

50

100

150

200 u

250

300

350

400

450

Fig. 28.

Case ka=29, J for = /3 using RI=1,2,3 (k) and RI=1,2,3 (k) compared to MoM.

23

A PPENDIX C: M ODIFYING C OEFFICIENTS

1.5 x 10
3

Cluster plot Mom versus Ref least sq. fit peak matching

Cross current components (MoM)

0.5

0.5

1.5

2 1.5

0.5 0 0.5 Cross current components (Ref)

1 x 10

1.5
3

Fig. 29. Cluster of points (JRef , JM oM ) for the lit region only, = , ka = 29, the slope of a tted line in the lit interval giving the modifying 8 coefcients (R . The line can be computed either by using least square tting or peak matching tting. The latter uses only a subset of the data rather than the whole cluster of points as in least square, thus removing noise.

A PPENDIX D: 2D S OLUTION ON A C IRCULAR S ECTION Let u = Ez for E = (0, 0, Ez ). Helmholtzs equation 2 u + k 2 u = 0, u IR2 S (76)

with boundary condition uinc = uscat on the circular boundary S of radius r = a, can be solved analytically for an incident planar wave uinc to very high frequency for
+

uscat =

im
m=

Jm (ka) im 2 e Hm (kr) 2 Hm (ka)

(77)

2 where Hm is a Hankel function and Jm is a Bessel function of the rst kind.

24

(a) Case ka=29, total E-eld vanishes in the shadow zone Fig. 30.

(b) Case ka=29, scattered E-eld

2D analytical solution on a circular PEC illuminated by a planar wave from right to left.

Etot 0.3 1.5

Escat

0.2 1 0.1

0.5 0

0.1 0

0.2 0.5 0.3

0.4

50

100

150

200

250

300

350

400

50

100

150

200

250

300

350

400

(a) Case ka=29, total E-eld vanishes to zero in the shadow region Fig. 31.

(b) Case ka=29, scattered E-eld

Circular section of E-elds close to the surface, E=0 exactly on the surface.

25

PHYSICAL OPTICS AND NURBS FOR RCS CALCULATIONS

Sandy Se(I), Jesper Oppelstrup(II) Email: sandy@nada.kth.se Royal Institute of Technology, Stockholm, Sweden Email: jespero@nada.kth.se Royal Institute of Technology, Stockholm, Sweden
(II) (I)

Abstract: This paper presents a Physical Optics method using an adaptive triangular subdivision scheme for the surface integral in the computation of monostatic RCS of NURBS surfaces.1

Introduction

The computation of electromagnetic radiation patterns from complex geometries with Physical Optics (PO) involves the evaluation of a surface integral. Classically, the PO integral is evaluated numerically by quadrature formulas on triangular or quadrilateral patches which require several evaluation points per wavelength in order to catch the highly oscillatory behavior of the integrand. Alternative methods consist in decoupling the integrand into an oscillatory phase factor from an amplitude factor which is smoother. Then, the integral can be solved analytically from few points per surface provided that both factors have been approximated separately using e.g. polynomials, linear as in [1], [2] and [3], or quadratic as in [4]. Typical geometrical scenes contain complex geometries like Unmanned Aerial Vehicles (UAV) such as Eikon in Figure 1. These are small aircraft which have 8 to 10 meters in length. A simple CAD design for Eikon will consist in 109 NURBS surfaces along with 264 trimming curves. Each NURBS patch can then be discretized into smaller triangles on which the PO integral is more easily evaluated.

Figure 1: Eikon UAV.

This work is part of the CEM program at the Parallel and Scientic Computing Institute (PSCI) under the project GEMS: General Electromagnetic Solvers supported by the National Aeronautical Research Program (NFFP), the Swedish Agency for Innovation Systems (Vinnova), Saab AB and Ericsson Microwave Systems AB.
1

To do so without resolving the wavelength, a linear phase and constant amplitude approximation to the PO integral is used for each triangle. But, to insure its accuracy, we adaptively adjust the triangle density based on local variation of the phase. This adaptive subdivision scheme is the basis of the PO method used here and the gain resides in how fast it can be performed. In addition, since working with general shape objects, the scheme automatically resolves the geometry, i.e. the subdivision criteria takes into account the curvature of the object using the NURBS description of the surface. Furthermore, improvements to the basic PO formulation are made by including shadowing [5] and edge diraction. For fast edge detection, we use the trimming curves topology information available in the NURBS description. Approximate shadow regions are obtained eciently on the NURBS surfaces using raytracing techniques. Practically, a test on the outward normal at each triangle is combined with an occlusion test using ray tracing on NURBS, see [6]. One ray is launched from each triangle and traced toward the observer direction. As a result the triangle is considered illuminated if the ray path is not blocked by any other surfaces. To be ecient, this method also requires the use of topological information between NURBS and triangles, see [6] for more details. This set of algorithms can be used directly for an enhanced PO analysis on smooth curved surfaces or can be applied in an iterative MoM-PO hybrid solver for more complex problem, as in [3].

Evaluation of the radiation integral

In the frequency domain, the radiation process described by Physical Optics consists in the approximation Jpo to the currents generated from a known incident eld on a surface S, 2 scat (X ) H inc (X ), on the lit portion of S n Jpo = (1) 0, on the shadowed portion of S nscat is the unit normal vector to S at X , the position vector of a point on S and H inc is the complex magnetic eld incident on that point. PO assumes a zero surface current in shadow regions which neglects the creeping wave contribution. The scattered elds E scat from the surface currents at an observation point X can be expressed over the entire surface S by integrating the scattering component of the PO currents applied to the Greens function as follows, E scat = jkZ0
Slit

[Kscat Kscat Jpo ]G(X, X )ds

(2)

where ds is the innitesimal element of area on S, Z0 the free space impedance, Kscat and k = 2/ is the wavenumber where is the the direction pointing from S to X wavelength. The free space Greens function is ejk|XX | G(X, X ) = 4|X X | 2

(3)

Under far eld assumption and for monostatic emitter observer direction, the integral from a reecting PEC surface reduces to j ejkR E scat = E0 R
Kscat nscat e2jk(Kscat .X ) ds

(4)

Slit

where Slit is the illuminated portion of S which must be determined via a shadow region procedure. We call the PO integral the following complex function: Ipo = 1
Kscat nscat e2jk(Kscat .X ) ds

(5)

Slit

The exponential factor displays the rapidly varying phase. Finally we dene , the Radar Cross Section, as a normalization of the scattering integral Ipo such as = 4||Ipo ||2 (6)

The numerical problem consists in evaluating the double integral in Equation (5). We decouple the phase factor which is highly oscillatory, from the amplitude factor which is smoother. Then the phase can be approximated by linear function Bc (x, y) over small cells such as Bc (x, y) = Bx x + By y + B0 2Kscat X (x, y), (x, y) cellc (7)

The three coecients B can be determined using triangular cells only which are mapped into a reference conguration. The dierence between the true integral and the numerical integral can be measured as follow: =| |KA + KB | = | A(x, y)ejkBc (x,y) dxdy (A A)ejkBc dxdy +
N cell jkBc

A(x, y)ejkBc (x,y) dxdy| AejkBc (1 ejk(Bc Bc ) )dxdy| |(A A)ejkBc |dxdy
e e

(8) (9)

|KA | = |

(A A)e

dxdy| =
cell=1 Cell

(10)

|KA | |A A|KA KA h
N cell

(11) |(1 ejk(Bc Bc ) )|dxdy

e

|KB | = |

Ae

e jkBc

(1 e

e jk(Bc Bc )

)dxdy|

(12) (13)

cell=1

Cell

where KA and KB are constant and h represents a measure of the size of the cell. Note from Equation (13) that 0 when h 0 and that we only need cell of size h . The surface normal is taken as the triangle normal such as the amplitude factor becomes constant over a triangle and thus can be moved outside of the integrand. Then the linear phase approximate integral can be written Ipo 1 Ipo = cell=1 cell 3
N cell

KA h + KB kh2

(14)

1x

Ipocell = 2Sc

ejk(Bx x+By y+B0 ) dxdy

0 0

(15)

where Sc is the area of the triangular cell. The coecients are real numbers calculated to best t a plan from the values at the vertices of the triangle Bx = 2(Bc,1 Bc,3 ), By = 2(Bc,2 Bc,3 ), B0 = 2Bc,3 (16)

where Bc,i are the phase values of the i-th vertex of the triangle. Equation (15) can then be evaluated analytically [1] yielding Ipocell = 2Sc e
jkB0 j j kBx j kBy e 2 [e 2 sinc( kBx kBy ) kBy 2 kBy kBx j j kBy j kBx 2 [e 2 sinc( ) e kBx 2

sinc( kBx )], if |Bx | > |By | 2

sinc( kBy )], if |Bx | |By | 2

(17)

Special care must be applied to avoid large numerical errors when the B are both very small. The size of each cell corresponding to the number of integration points is determined on the NURBS by the subdivision scheme. The accuracy of the scheme is monitored as the integration proceeds by an error estimate. For curved geometries, if X do change much over the cell, due for instance to a change in curvature, then the cell will be further subdivided.

Adaptive Subdivision Scheme

Figure 2: Subdivision procedure necessary to resolve curved and non planar geometries. The subdivision procedure, see Algorithm 1 below, is done recursively over each triangle and is stopped when a subdivision criteria reaches a desired tolerance. The subdivision criteria has been chosen carefully such that it insures that over stationary phase regions, where the phase is varying slowly [7], the density of triangle is sparse. From an initial coarse triangle T with nodes on the surface, new nodes are inserted at the middle of each edge of T . This creates four sub-triangles T1 , T2 , T3 and T4 as illustrated on the right picture of Figure 2. Special care is taken if the triangle lies on non planar 4

Algorithm 1 Integration (Triangle T, E , Tolerance) Require: E Solve-PO-Integral (Triangle T){use Linear Phase Approximation}
if Tolerance< machine then Return E {to avoid arithmetic exceptions} else T1 , T2 , T3 , T4 Build-Sub-Triangles 2 (Triangle T) for i = 1 to 4 do Ei Solve-PO-Integral (Triangle i) {use Linear Phase Approximation} end for Enew = E1 + E2 + E3 + E4 if (||Enew E ||2 <Tolerance) then Return Enew {Subdivision criteria is reached } else for i = 1 to 4 do Ei Integration (Triangle i, Ei , Tolerance/4) {Recursive subdivision} end for Enew = E1 + E2 + E3 + E4 Return Enew end if end if

Algorithm 2 Build-Sub-Triangles (Triangle T) Require: Triangle T has nodes on NURBS, Topological information: T NURBS nurbs S Find-NURBS-Parent (Triangle T) {use Topology}
MiddlesOnT Middles-Edges (Triangle T) MiddlesOnS Project-On-NURBS (MiddlesOnT, nurbs S){use CGM } if Edge-Is-On-Trim-Curve (Triangle T, nurbs S) then MiddlesOnS Local-Renement (MiddlesOnT, nurbs S) end if T1 , T2 , T3 , T4 Create-Sub-Triangles (Triangle T, MiddlesOnS) Return T1 , T2 , T3 , T4

surfaces. The procedure resolves these geometries using local renement and projection of the new nodes on the NURBS surfaces, see middle and right pictures of Figure 2. The projection is done by minimizing the distance between the node and the surface using Conjugate Gradient Method (CGM), see [6] for details. The subdivision does not change the position of the initial nodes so only the three nodes at the middle of each edge need to be projected. The local renement corresponds to a displacement of the sub-nodes if the edges of T lie on the trimming curve of the surface. In such case, the procedure projects the nodes on the curve itself. The diculty here is to detect when a specic edge of a triangle is discretized one of the curves of the NURBS representation. To simplify the task, the initial triangulation has been connected with topological information [6] from the NURBS description so that each triangle knows on which surface or curve it belongs. Then a straight forward localization algorithm is applied in order to nd the renement point being the closest point on the curve. When all four triangles have been dened, the next step is compute to PO integral on each 5

new triangles. As before we using linear phase approximation. Then the contribution of the new triangles to the scattered eld is determined. The sum of their scattered elds scat ETnew is given by scat scat scat scat scat ETnew = ET1 + ET2 + ET3 + ET4 (18) The subdivision criteria is then the absolute error abs in the square norm between the scat scattering eld from the initial triangle ET and the sum of scattering elds from the scat new triangles ETnew : scat scat abs = ||ET ETnew ||2 (19) If the error is not small enough, the subdivision proceeds recursively and the same func1 tion call is applied to each sub-triangle with 4 of the previous tolerance. The absolute error insures that the subdivision will stop as soon as the scattering eld does not vary (stationary) between two subdivision levels or if the contribution to the scattered eld becomes small.

Figure 3: One, four and ve levels of subdivision applied to a circular plate.

Scattering from a circular plate

Figure 3 illustrates how the procedure works on a circular plate initially discretized in eight triangles. At the rst level of subdivision all the eight initial triangles are cut in four such that 32 new triangles are created, see left plot of Figure 3. The eect of local renements can been seen immediately at this rst level of subdivision. The sub-triangles extend outside the initial triangle since they have been projected on the border of the circular plate. Since the surface is plane, the integral is evaluated exactly on every triangle. The scattered eld will only be aected by a change in the geometry of the triangles T2 and T4 which have edges on the borders of the circular plate. In opposition, for sub-triangles T1 and T3 the scattered eld remains constant and they do not need to be further subdivided which explain the rose shape conguration after 4 levels of subdivision, see middle plot of Figure 3 and after 5 levels on the right plot of Figure 3. This result was validated against analytic results of the PO integral for the same circular plate when the scattering direction Kscat is varying according to an angle [0, 90] 6

RCS of 1m radius circular plate at 300Mhz : = 1m 40 POADAPTIVE POANALYTIC 20

4 erro in RCS of 1m radius circular plate at 300Mhz : = 1m

0
1

[db]

20

40
2

60
3

80

10

20

30

40 50 o [degree] ( = 0 )

60

70

80

90

10

20

30

40

50

60

70

80

Figure 4: Adaptive PO on NURBS versus analytic results for a 1 meter radius circular plate at = 1 meter and corresponding error plot per angle of incidence.
RCS of 1m radius circular plate at 1Ghz : = 0.3m 40
Absolute Error in [db] 10 erro in RCS of 1m radius circular plate at 1.0Ghz : = 0.3m POADAPTIVE = 1e2 POADAPTIVE = 1e4 POADAPTIVE = 1e6 POADAPTIVE = 1e8

POANALYTIC POADAPTIVE = 1e8 20

10

10

20

30

[db]

20
3500 3000

40 50 [degree] ( = 0o) # triangles Created per PLW

60

70

80

40
2500 # triangles 2000 1500 1000 500

POADAPTIVE = 1e8 POADAPTIVE = 1e6 POADAPTIVE = 1e4 POADAPTIVE = 1e2

60

80

10

20

30

40 50 o [degree] ( = 0 )

60

70

80

90

10

20

30

40 50 o [degree] ( = 0 )

60

70

80

90

Figure 5: Adaptive PO versus analytic results for a 1m radius circular plate at = 0.3m, plus corresponding error plot per tolerence as well as total number of sub-triangles created per angle and per tolerance. degree. For each angle, the monostatic RCS for a circular plate of one meter radius was computed both analytically and using the adaptive subdivision scheme. For the rst test, the wavelength was 1 meter and 0.3 meter for the second test. The results are displayed respectively in Figure 4 and Figure 5. The computed curve when using adaptive subdivision follows well the behavior of the analytic one, except at some dips of low decibel, singularities which anyway give no contribution to the RCS. On the right plots we can see the error with the analytic RCS 0 . With a smaller wavelength, the error remains small, see the top right plot of Figure 5, at least for all the angles [0, 50] degree and decays when the tolerance is reduced up to 1e8 . At such levels of accuracy, a maximum of 3.200 triangles is needed but this number rapidly decays with the angle, see bottom right plot of Figure 5. In comparison, classic PO code will need at least twice as many triangles to resolve the same wavelength with 10 triangles per wavelength and this regardless of the angle of incidence.

Conclusion

We have presented an adaptive triangular subdivision scheme for solving the PO integral which compute the integral on few points per surface. The innovation is that the number of integration points and the accuracy of the scheme are monitored as the integration proceeds by a measure of the local variations of the scattered elds. To illustrate the software robustness, we have applied the scheme to the Eikon UAV, see Figure 6. This shows the code works on complex geometries but results still remain to be further investigated. We have also shown how it is advantageous to preserve the native NURBS representation, given by the CAD design, along with the triangulation. In Particular, since each triangle knows on which surface it belongs, this is what allows the scheme to resolve the geometry.

Figure 6: Eikon triangulation initial and after adaptive subdivision.

References
[1] A. C. Ludwig. Computation of radiation patterns involving numerical double integration. IEEE Transactions on Antennas and Propagation, pages 767769, Nov. 1968. [2] F.J.S Moreira and A. Prata. A self-checking predictor-corrector algorithm for ecient evalaution of reector antenna radiation integrals. IEEE Transactions on Antennas and Propagation, 42(2):246254, Feb.. 1994. [3] J. M. Taboada, F. Obelleiro, and J. L. Rodr geuz. Improvement of the hybrid moment method-physical optics method through a novel evaluation of the physical optics operator. Microwave and Optical Technology Letters., 30(5):357363, 2001. [4] Glenn D. Crabtree. A numerical quadrature technique for physical optics scattering analysis. IEEE Transactions on Microwave Theory and Techniques, 27(5), Sep. 1991. [5] Nazih N. Youssef. Radar cross section of complex targets. IEEE Transactions on Antennas and Propagation, 77(5), May 1989. [6] S. Se. Ray Tracing Tools for High Frequency Electromagnetics Simulations. Licentiate thesis No. 0314, Dept. of Numerical Analysis and Computer Science, KTH, June 2003. [7] M. F. Catedra J. Perez. Application of physical optics to the rcs computation of bodies modeled with nurbs surfaces. IEEE Transactions on Antennas and Propagation, 42(10), October 1994. 8

The Rescue Wing: Design of a Marine Distress Signaling Device.

Tomas Melin
Department of Aeronautical and Vehicle Engineering, AVE Royal Institute of Technology, KTH Stockholm, Sweden Email: melin@kth.se

Sandy Se
Department of Numerical Analysis and Computer Science, NADA Royal Institute of Technology, KTH Stockholm, Sweden Email: sandy@kth.se

Abstract We present a multidisciplinary scientic analysis combining aerodynamics, ight mechanics and electromagnetics aiming at the design of a new marine distress signaling device. We show how computational uid dynamics (CFD) and computational electromagnetics (CEM) techniques have been used to assist in the design of both the ight characteristics and the radar performance of the device, as well as how its radar signature compares to popular radar reectors used on yachts and sailboats.

I. I NTRODUCTION Marine distress signaling is of paramount importance for persons in distress at sea, being the mean which lets rescue teams be aware of the situation and deploy a search and rescue mission.

In this paper we focus on a personal balloon-type device carried by the survivor, shown in Fig. 1 and called The Rescue Wing. The Rescue Wing works as a passive radar reector and visual marker assisting in localization during search and rescue operations of persons missing at sea. Its design is the fruit of a multidisciplinary study combining simulation results from aerodynamics, ight mechanics and electromagnetics, as well as data from trial ights. For aerodynamics, one of the challenges was to create an inatable light-weight structure with adequate aerodynamic characteristics. The design was implemented and modeled using the commercial CFD software FLUENT [1]. In order to assess the Rescue Wings ability to reect radar signals, electromagnetics simulations have been conducted to predict its radar cross section (RCS). The computations used the General Electromagnetic Solvers GEMS [2], a software suite developed at KTH for computational electromagnetics. II. C HARACTERISTICS OF M ARINE D ISTRESS S IGNALING On the 14:th of June 2003 a man fell overboard from a small sailing yacht outside the Halland coastline in the south of Sweden. The waves were about 1.5 to 2 meters high with a wind speed of 10 m/s. In total, ve surface vessels and three helicopters were engaged in the search and rescue mission. The search effort was hampered by the difculty of determining the position of the originating event. After three hours, one of the helicopters spotted a wave crest apparently moving in the wrong direction: the man wearing a white sweater over his life jacket. He was suffering from beginning hypothermia but was recovered unharmed. This example is typical for search and rescue missions in that the time spent on search is often many times larger than the time spent for the actual rescue. An effective signaling/localization device would lessen the search time and allow for more lives to be saved. III. P URPOSE Since the Titanic disaster, research in distress signaling has been addressed and constantly updated by the SOLAS (International Convention for the Safety of Life at Sea) [5]. These international treaties dene a number of aids including rocket ares, hand ares, buoyant smoke signals as well as

Fig. 1.

The Rescue Wing during test ight.

a number of active radio devices. The Rescue Wing can not, strictly speaking, be classied as a signaling device as it only provides a radar and an optical target, but it is intent to help localizing a person in water. The Rescue Wing is an inatable gas bag, lled with helium to provide aerostatic lift. It is shaped like a wing thus providing aerodynamic lift. As an add-on device to standard life jackets, it will be at hand when needed. In the case of a person falling overboard, or leaving the ship due to an on-board emergency, the Rescue Wing provides a mean for person in distress to communicate his position. By a simple grip-and-twist mechanism, the ination of the envelope is engaged and Rescue wing is deployed, after which no manual handling of the device is required. The Rescue Wing has been designed to operate as a displacing balloon in calm winds, and as a kite in windy conditions. It hovers 10-15 meters above the sea surface providing a radar reector target as well as a strong visual cue for detection and positioning. In the next section, we will look at the constraints restricting the design, as well as the functional specications stipulating the properties we expect from the device. IV. C ONSTRAINTS I MPOSED ON THE D ESIGN

V. A ERODYNAMICS Due to the necessarily bluff body shape of the inatable device, standard aerodynamic conceptual design tools such as handbooks [3] and panel methods [4] are not applicable. Instead, a full Navier-Stokes solver had to be employed in order to generate the aerodynamic database. Fortunately, the simplicity of the design geometry (Fig. 2) led to easy grid generation, thus enabling fast design loops. As in airplane design, key parameters are lift and drag forces L and D, and their ratio G, the glide slope. Aerodynamic results were collected in the standard way into coefcients of lift: CL , drag: CD and pitching moment CM as functions of the angle of attack . The coefcients are specied in equation 1, where q is the dynamic pressure and S the reference area. CL = L() , qS CD = D() , qS CM = M () qCM AC (1)

where CM AC is the mean aerodynamic chord of the device, see Fig. 3. The computed data was then curve tted with second degree polynomials to yield an expression suitable for the ight mechanic analysis, equations 2 and 3. The numerical error in this interpolation was small, as the aerodynamic behavior of the device was smoother than that of ordinary aircraft designs. CL = CL0 + CL + CL2 2 CD = CD0 + CD + CD2 2 (2) (3)

Together, these equations made it possible to dene and compute the glide slope G(), G=
Fig. 2. CAD model and triangulation of the Rescue Wing. is the angle between the main axis of the wing and the monostatic azimuth directions.

CL () CD ()

(4)

The design requires nding a good balance between different aspects and features of the device. These can range from size, weight, portability, up to external constraints such as environmental considerations, packaging, available material, production methods, etc. A condensed wish-list of the specications is provided below: Inexpensive and easy to manufacture. In stores, presented as an add-on safety device. When packaged, t in a cigaret box-size canister. In storage, have long shelf life and light weight. When not operating, waterproof, corrosion resistant. In stand-by operation, be attached to a life west or raft. In distress situation, easy to arm and to engage. When inated, hover at 10-15 meters above sea. In calm waters, generate aerostatic lift as a balloon. In strong winds, remain in stable ight as a kite. At long range, reect radar waves of X- or S-bands. At close range, provide a strong visual cue. 2

As shown in Fig. 3, the glide slope determines the elevation angle in steady ight when effects of gravity and buoyancy have been neglected. AC is the aerodynamic center, b the bridle point, where the leading and trailing edge tethers join into the main tether of length l going down to the anchor point A. V is the free stream velocity, M is the moment around the bridle point and s is the circle segment spanned by the bridle point at different elevation angles . The tangential plane to s, t is the reference plane when measuring the angle of incidence, i of the CM AC . The tangential force F acts in parallel with the plane t and exerts a momentum in the bridle point when the moment arm a is greater than zero. = arctan(L/D) VI. F LIGHT M ECHANICS The ight mechanics of kites is somewhat different from aircraft. In this paper only a brief overview of the method of nding an appropriate longitudinal trim will be presented. For the sake of simplicity, the weight of the line and the envelope is neglected, as are the buoyant effects. This simplication is (5)

The third case, with an incidence of i = 0.2 radians has a zero tangential force crossover elevation angle of slightly above one and has the desired glide slope behavior.
Elevation trim graph 1.5 1 0.5 0 0.5 1 i = + 0.2 rad i = 0 rad i = 0.2 rad

F [N]

0.2

0.4

0.6

0.8 [rad]

1.2

1.4

1.6

2 1.5

Fig. 3.

Variable denitions.
G []

1 0.5 0 0.5

valid for a free stream velocity range giving a high dynamic pressure resulting in the aerodynamic forces being much larger than the mass effects, i.e. L mg. The tangential force F must be zero at the equilibrium elevation angle, as described in equation 6. Additionally, as all aerodynamic forces are transmitted trough the bridle point into the tether, the moment M at the bridle point must be zero in trimmed ight according to equation 7. F = L cos() D sin() = 0 M = F a = (L cos() D sin())a = 0 (6)

1 1.5 0 0.2 0.4 0.6 0.8 [rad] 1 1.2 1.4 1.6

Fig. 4.

Trimming the elevation angle using the incidence as parameter.

(7)

The angle of attack is a function of the incidence i and the elevation angle according to equation 8 = +i 2 (8)

Equations 6, 7 and 8 form a system of equations with three unknowns: The incidence angle i, and the position of the bridle point b in relation to the aerodynamic center AC. This system is readily solved numerically, while keeping the elevation angle or, the glide slope G as high as possible. When plotting the tangential force against the elevation angle as in Fig. 4 the dependence of the tangential force of the incidence become clear. Of the three cases, the one with i = +0.2 radians clearly never crosses the line F = 0 which means that this conguration is not stable in elevation angle, as the tangential force always is negative, thus forcing the elevation angle to zero (giving zero altitude). Decreasing the incidence to zero, gives a stable conguration with a crossover F = 0 at 1. However, when examining the glide slope at the crossover in the lower graph, the glide slope there is still increasing with increasing . Having a negative G would insure a stiffer system, and be on the right side of Gmax with respect to the stall limit. 3

When a suitable angle of incidence had been attained, the bridle point could be positioned using Fig. 5. The upper panel shows the glide slope of the selected design with the G = 1 design criterion and the maximum glide slope criterion (Gmax ). These boundaries limit the useful part of the glide slope range. The G = 1 is set to ensure that the elevation of the kite always is larger than 45 degrees. In Fig. 5 the vertical distance between the aerodynamic center and the bridle point was set to CM AC /2 while varying the lateral distance h. The lower graph shows the moment coefcient around the bridle point for three different lateral positions. For h = 0, the zero moment crossover is at about 0.6 radians, which is over the Gmax limit. For the h = 0.2 CM AC , the crossover is at about 0.22 radians which is within the cruise quadrant but off the trim angle decided by equation 8 and the elevation trim. Instead, setting the lateral distance h = 0.1 CM AC yields the same trim glide slope as the elevation trim and thus positioning the bridle point on the line connecting the aerodynamic center and the anchor point. A. Altitude Stability We investigated the ow eld over an assumed sea surface in order to determine the behavior of the change in angle of attack due to the shape of the surface. In this simulation we assumed the waves being described by a series of cubic B zier e curves. The wave height is 1.3 meters and wave length is 14 meters as shown in Fig. 6. Free stream velocity was set to 10 meters per second in standard sea level atmosphere. The simulation showed the formation of recirculation bubbles in the wave troughs and an inuence on the shape of the streamlines above the wave crests.

Pitch trim graph 2 Cruise quadrant G, Glideslope, [] 1 G Design minimum limit

6 8

Change in angle of attack with wave phase and height above the sea. h = 0.5 m h=1m h=5m h =10m

1 G Max limit 0 0.2 0.4 0.6 0.8 Alpha, Angle of Attack [rad] 1 1.2 1.4
AoA [deg]

2 0.2

Cm, Bridle Momentum Coefficient, []

0.01 0.005 0 0.005

0.01 0.015 0.02 0.2 h=0 h=0.1 h=0.2

4

0.2

0.4 0.6 0.8 Alpha, Angle of Attack [rad]

1.2

1.4

6 wave phase [rad]

10

12

Fig. 5.

Positioning the bridle point in relation to the aerodynamic center.

Fig. 7. Change in angle of attack (AoA) as a function of wave phase and the height (h) above the wave crests parameter.

a is earth radius and h and H are the heights of transmitting antenna and the target respectively [6]. R= 2ah + 2aH (10)

Fig. 6. Streamlines just above the sea surface. The wave height is 1.3 meters and the wave length is 14 meters.

In order to determine this inuence quantitatively, we isolated the vertical velocity component and computed the change in angle of attack according to equation 9, where v is the vertical speed component and V is the free stream velocity. In Fig. 7 we show the dependence of the change in angle of attack with height over the sea surface, measured from the crests, as a function of the wave phase. v ) (9) V In correlation with our initial assumption, the inuence of the wavy surface diminishes with increasing altitude. At 10 meters, or about 7 wave heights the inuence is negligible. The offset from the zero line of the 10 meter curve is due to boundary layer growth. = arctan( B. Operational considerations. Radar wave clutter becomes an important consideration in the design, and operational considerations of the Rescue Wing. When investigating the radar range, the line-of-sight range R is usually taken as equation 10, ignoring refraction effects. Here 4

As the Rescue Wing has an radar cross section of about the same size as the rough seas, the signal to clutter ratio (S/C) becomes an important parameter in determining the possibility of detection of a specic target. As shown in Fig. 8 the region with low grazing angles, here called the clutter region, is where the signal to clutter ration is the lowest. Beyond the line-ofsight range, smaller waves disappear below the horizon, and gradually the (S/C) increases. Assuming innite power in the transmitter, linear wave propagation and a sufciently long tether, the detectability increases for a larger range. However, the Radar range also depends on the fourth power of the range when discussing the received reected energy. This means that we would like to keep the far region as close to the transmitting antenna as possible. One way of attaining this is to keep the height of the transmitting antenna as low as possible, thus favoring search from a small surface vessel rather than from a helicopter. C. Inatability In order for the wing to be buoyant in still air, the material employed must be extremely light-weight and also work as an efcient barrier for the gas inside. The choice fell on a composite multi-layer plastic: metallized polyethylene tereftelathe, or PET. The buoyancy criterion also imposes harder constraint on the weight of the device, as the buoyant force is much smaller than the aerodynamic forces encountered. In essence, the buoyant lifting force is proportional to the volume of the envelope, while the weight is proportional to the surface area (for a given material thickness).

inuence on performance of constraints imposed on the design is very signicant. A. Electrical Size and Shape Due to a moderate size of 2 meters in diameter, a low wing prole and an inatable round shape, we expect the Rescue Wing to have a low radar signature, in the range of the sphere, as illustrated in Fig. 9. This has been veried by the CEM simulation where the Rescue Wing was modeled as a perfect electric conductor 2.0 m 1.5 m 0.5 m in dimensions, see Fig. 2. The results given in Fig. 13 and Fig. 14 show that the RCS of the wing displays mainly low values under 0dB. B. Frequency Range VII. T HE R ADAR C ROSS S ECTION The strength of the radar reection from the Rescue Wing is measured by its monostatic radar cross section (RCS) represented by the symbol . The RCS characterizes the effective area illuminated by the incident plane wave which is directly scattered back in the direction of the emitter. Rescue radar systems generally operate at frequency ranges in the following bands: the S-band, ranging from 2 to 4 G-Hz with specic bands at a frequency of 3.0 G-Hz (3000 MHz) which has a wavelength of 10 cm, the X-band, ranging from 8 to 12 G-Hz with a typical frequency of 9.4 G-Hz (9400 MHz), is characterized by a smaller wavelength of 3.2 cm. Ships and rescue helicopters will typically carry both the Xand S-band while small vessels are limited to X-band units. S-band gives a longer range, up to 24 miles. X-band radar is more sensitive to interference from rain and waves (sea clutter) but offers greater resolution and detection of smaller targets such as the Rescue Wing. To be detected above the sea clutter, in a moderate sea using X-band radar, a return of +1dB to +5dB ( 3m2 ) must be reached as a threshold [7]. C. Radar Reectors Radar reectors [10] are designed to increase the reectivity of buoys or small craft so that they become more visible on radar. The key strategy in their design is to use at plates which produce strong backscatter at normal incidence. For instance, the commercial Davis emergency reector in Fig. 10, one of the most popular radar reectors, is composed by two vertical at disks butted together at 90o onto one horizontal disk to form rectangular corners which maximize the reections in almost all directions.

Fig. 8.

Sea clutter.

Fig. 9. Typical RCS values in m2 and in dBsm for a range of different objects in size, including the Rescue Wing. The radar detectability threshold is about +3dB.

RCS is measured in square meters, often normalized and displayed using a logarithmic scale dened in decibel squaremeters (dBsm), see Fig. 9, RCS[dBsm] = 10 log10 RCS[m2 ] 1[m2 ] (11)

The reference 0dB corresponds to the return from a sphere of 1m2 cross section. It is often convenient to display RCS values using a polar diagram. Fig. 11 shows a polar plot of a commercial radar reector (Fig. 10) and the response of the wing to signals arriving from different azimuth direction inc using a polar plot are given in Fig. 13 and Fig. 14. In common with aerodynamics design, a knowledge of the factors which affect the radar performance of the system is essential. The most signicant factors are those which inuence the propagation of the radar wave such as directional characteristics, radar operating frequency, polarization of the incident and reected waves, as well as size, shape and material covering the surface of the object. In following, we will have a closer look at these factors in order to understand how they will affect the reective performance of the Rescue Wing. Also, we will see that the 5

Fig. 10.

Typical geometry of a passive radar reector (Corner reector).

The RCS at X- and S-band are displayed in Fig. 11. Similar results in agreement with our simulation have been previously

published in a study conducted by the US Sailing Safety-AtSea Committee [11] aiming to test the efciency of various radar reectors. The small reector displays a small radar signature in accord with the tests reported in [11]. Inatable [8] [9], light and collapsible versions of such reectors are commercially available.
0 330 10 0 10 300 20 30 40 270 90 60 30

Both PO and MoM take as input a discrete mesh of the CAD geometry, where the elements must be small compared to the wavelength. An example of such discretization for = 30 cm is given in Fig. 2. F. Physical Optics and Method of Moments Since both low and high frequency methods operate on the same geometry, we can compare them at mid-frequency (run at 1 G-Hz) in order to investigate the error of the approximate PO. The results are displayed in Fig. 12. We see that the PO solution follows the MoM solution well except at directions normal to the axis of the wing, [0o 20o ] and [160o 180o ], where PO predicts too high RCS. This is mainly due to one wing shadowing the other, which is badly modeled by the PO solution [13]. However, the results show that away from normal axis incidence, i.e for about 2/3 of all aspect angles [20o 160o ], the PO solution is in good agreement with MoM.
Kite40degree, Method of Moments vs Physical Optics 1.0Ghz, wavenumber:20.95 20 MoM PO

240

120

210 180

150

Fig. 11.

RCS of the Davis Emergency (5.7in) radar reector.

15

10

RCS [dBsm]

If one or more units are inated and attached to the tether, their combined signatures would result in large RCS values well above the +1dB detectability threshold. D. Polarization Typical radar units can send and receive in both vertical and horizontal polarizations. In kite ight position, at angle of attack < 20o in horizontal elevation, or in balloon ight position at angle of attack < 45o as seen in Fig. 1, the low wing prole will make horizontal polarized reected waves dominant. Thus, only horizontal polarization is of interest for the simulation. E. RCS Calculations Tools In order to verify our assumptions on the Rescue Wing radar detection, RCS simulations are required. The GEMS [2] software suite covers a wide range of numerical methods for predicting RCS of complex threedimensional geometries using numerical evaluations of the Maxwell equations. For low frequencies and small objects in term of wavelength, exact techniques such as Method of Moments (MoM) combined with Fast Multipole Method [12] (FMM) are available. These methods numerically solve the exact integral equations for the electromagnetic elds using a set of discrete basis functions on the boundary. For high frequencies, i.e. electrically large targets, exact methods become impractical either because of computer time or memory requirement, so that approximate techniques must be used. It is well established that for monostatic RCS, the Physical Optics (PO) based on surface current approximations [14] is a good choice for computing the main reection from large, in term of wavelength, smooth surfaces. 6

10

15

20

25

30

20

40

60

80 100 Monostatic angle

120

140

160

180

Fig. 12. Exact numerical solution versus fast approximate solution of the RCS in dBsm of the Rescue Wing in ight kite position, horizontal elevation of 40o at a frequency of 1G-Hz, in function of . = 0o corresponds to a radar waves coming from normal to the axis of the wing ( = 90o ).

G. RCS of the Rescue Wing at S-band and X-band We look at the contribution from all 360o angles () in an horizontal plane where the wing is resting with the nose at 0o . The RCS in all directions is quite low, as seen in Fig. 13 at S-band and in Fig. 14 at X-band in horizontal polarization.The long straight leading edges of the wing reect the most. The CAD geometry introduces a ctitious sharp edge at the nose. The sharp edge breaks the reection and that is why the RCS for around 0o is small. The real shape is smoother so we should expect the real wing to return slightly higher values at this angle.

0 330 30 20 10 0 10 20 30 40 30

300

60

270

90

240

120

210 180

150

Fig. 13.

RCS of the Rescue Wing in dBsm for S-band.

0 330 30 20 10 0 10 20 30 40 30

However, limiting factors in overpressure are envelope membrane strength and helium leak rate. The qualitative relations between these factors are still to be found. We have seen that the key factor inuencing the radar performance of a Rescue Wing is its size. However, constraints limit size due to light weight requirement and that the deated wing has to t into a cigaret box-size container. The RCS can be effectively increased though the addition of a radar reector. The aluminum material covering the wing acts a perfect electric conductor, so that radar reectors inside the wing will be ineffective. To overcome this problem, a radar reector could be attached to the wire. It should be located at least 2 meters above the see surface in order to lter out noise from the waves. However this will result in more drag forces and could lower the elevation angle of the wing, i.e. reducing its maximal altitude. ACKNOWLEDGMENT The authors would like to thank Hans Sj blom, who initio ated the work and is the holder of the Rescue wing related patents. For the RCS calculation tools, nancial support has been provided within the GEMS project by KTH and the Swedish Agency for Innovation Systems (VINNOVA) as a part of a collaborative research center PSCI. For developing the Rescue Wing, research grants has been nanced by the Carnegie foundation, Edvard Ros s foundation and KTH, dept. e of Aeronautical and Vehicle Engineering. R EFERENCES
[1] Fluent Inc, http://www.uent.com/, Fluent Inc, June, 2005. [2] GEMS, http://www.psci.kth.se/Programs/GEMS/, NADA, KTH, June 2005. [3] Roskam, et.al., Airplane design, Roskam Aviation and Engineering Corporation, Kansas, 1985. [4] Tomas Melin, http://www.ave.kth.se/divisions/aero/software/tornado/index.html, KTH, June 2005. [5] International Maritime Organisation, International convention for the safety of life at sea, SOLAS, IMO, London, 1986. [6] Skolnik, Introduction to radar systems, McGraw-Hill, 1962. [7] Kenneth Parker, Be Seen Or Be Sorry, Cruising Association, 2000. [8] Sidney Veazey, Inatable radar reectors, United States Patent 5969660, October 1999. [9] James Schaff, Steven Ball, Emergency passive radar locating device, United States Patent 6300893, October 2001. [10] John Briggs, Target Detection by Marine Radar, IEE Radar, Sonar and Navigation series 16, 2004. [11] United States Sailing Safety at Sea Committee, Radar Reector Test: http://www.ussailing.org, Safety Studies, 1995. [12] Martin Nilsson, A fast multipole accelerated block quasi minimum residual method for solving scattering from perfectly conducting bodies, Antennas and Propag. Society International Symposium No 4, 2000. [13] Sandy Se, Ray Tracing Tools for High Frequency Electromagnetics Simulations, Licentiate thesis No. 0314, Dept. of Numerical Analysis and Computer Science, KTH, June 2003. [14] Sandy Se, Jesper Oppelstrup, Physical Optics and NURBS for RCS calculations, EMB04 Computational Electromagnetics Conference Proceedings, Gothenburg, Sweden, October 2004.

300

60

270

90

240

120

210 180

150

Fig. 14.

RCS of the Rescue Wing in dBsm for X-band.

VIII. C ONCLUSION The design project has reached mid-term of its development. The next stage in preparation is to launch further trial ights. In order to reach production stage, auxiliary devices would need to be added, such as helium valves and packaging assembly. Aerodynamics simulation proves that the ight dynamics is stable but sensitive to large turbulence, to deformations in the position of bridle point and to changes in the magnitude of the incidence angle. Due to the inatable nature of the device, quite large deformations are possible. The rigidity of the structure is dependent on the internal helium overpressure. The higher overpressure, the higher rigidity.

A Modular Approach to GTD in the Context of Solving Large Hybrid Problems

Fredrik Bergholm1 , Stefan Hagdahl1,2 , Sandy Se1
1

Department of Numerical Analysis and Computing Science Royal Institute of Technology 100 44 Stockholm, Sweden E-mail: sandy/fredrikb@nada.kth.se
2

Ericsson Saab Avionics AB, Electromagnetic Technology 581 88 Linkoping, Sweden E-mail: stefan.hagdahl@esavionics.se

Abstract
In this paper we will present a ray tracer applied to Geometrical Theory of Diraction (GTD). The solver is a part of the Swedish suite of CEM solvers called General ElectroMagnetic Solvers (GEMS)1 . GTD is today a well established semi analytical method for high frequency radiation problems. During the seventies ecient GTD tools were developed and used in the academy and industry for canonical geometries. Today an abundance of GTD coecients have been solved for simple geometries. The process in examine new geometries is still going on. During the nineties GTD implementations that could treat industrial geometry, i.e. Non Uniform Rational Bezier Spline (NURBS), were developed. To exploit all these new facilities and to hybridize them with other CEM-methods there is a need for modern programming and ecient geometrical engines that treat industrial CAD in a robust way. We have developed a solver called MIRA (Modular Implementation of a GTD Ray tracer for Antenna applications) that springs from a solver called FASANT [2] developed by The University of Alcala, Spain. The solver is written in the modern F90-language and a host of new opportunities that this language gives are implemented, e.g modules, objects, derived types, overloading and linked chains.

Introduction
The solver is split up into three independent packages called Geometry, Ray and Application. The Geometry package can be executed as a stand alone solver that computes various kinds of geometrical entities. Any other software that needs NURBS related data can make use of this module. In addition, the Geometry module together with the Ray module build a stand alone solver that can serve any application that needs to access any propagation based on rays such as light, sound wave, water wave. Today the application package do antenna related computations but new application, e.g. Shooting and Bouncing Rays (SBR) via a new Application, may easily be added in the future. We have started to hybridize MIRA with a modern Method of Moment (MoM) [1] solver and it shows that the solvers data structure well support the needs from the hybrid application. Since the solver is written in F90 language it is well prepared to be compiled on large parallel computers and this will also be done in the near future. Below the three packages are discussed.
1 This work was supported in part by the CEM program at the Parallel Scientic Computing Institute at the Royal Institute of Technology.

Geometry
We here stress 5 features about the geometry. From a computational standpoint the object hierarchy, adaptive sampling and what we call a hole algorithm are important constituents. Furthermore, MIRA features multiple surface types and an explicit Bezier buer. Four basic structures in the object hierarchy are scenes, objects, surfaces and curves. (Curve = curve on a surface patch, here.) Having structures representing groups of surfaces, means that whenever surfaces are used in the model, we either have a double loop over surfaces by objects, or, have a separate object loop before double looping over surfaces by objects. This reduces computational time, e.g.: (1) In ray tracing, bounding boxes and spheres are important for reducing computations for intersection tests, and having such tests by object bounding bodies before other tests, yields a speed-up in scenes with many surfaces. (2) For diraction eld calculations in UTD, curvatures from the surfaces on either side of an edge are needed, and search for neighbouring surfaces may be limited to the same object (in most cases). Adaptive sampling refers to the need of (i) sampling entities such as normals and points on surfaces as preprocessing to avoid (many) function calls, (ii) make the sampling so that it is ecient, in the sense that, the samples are suciently dense, e.g., inside a trim curve, and that surface curvature aects sampling to avoid problems in nding reection or diraction points. Sometimes the physical material prevails outside a curve. Then the curve, whether delineating a hole or being a rim curve for an indentation/cavity, is an important source for diracted rays. Many types of antennas possess narrow cavities or indentations. Mira, in contrast to FASANT, has a built-in hole algorithm, allowing an arbitrary number of holes, Fig. 2 App. B. The algorithm determines whether points on material surfaces are inside or outside a curve, in so general a fashion, that complex structures such as a ring-like hole with material inside and yet another hole in that island of material, are permissible data. Each curve has a label i or o, short for remove-inside (=i) and remove-outside. Let p R2 be a map space (sample) point. We perform the tests p E1 E2 ... Ek = p, p I1 I2 ... I = p, where E.. is a curve with label o and I.. with label i, and curves causing p Es = , p Is = omitted. If p is in some hole(s), say Ij , j = 1, 2..., then the test Ij E1 E2 ... Ek = Ij is invoked, to nd out if the hole is in innermost material, performed as q E1 E2 ... Ek = q, with one q Ij . When calculating some entity in Mira [3] we use a transformation of spline data into Bezier surfaces (or curves): For instance, calculating (/u, /v)Cox(N (u, v)) to evaluate a B-spline N (u, v) R3 , where Cox is short for the Cox-de-Boor algorithm [4]. Some afterthought reveals that the Bezier data form a 5-dimensional matrix, call it B, which sometimes is very large. When visiting a surface (or a curve) it is natural to store the coecients of B in a buer. By requiring that the user may only use the buer by making a non-nested begin-surface call and an end-surface call, one may do memory-saving in the begin-surface and end-surface subroutines, e.g. allocations and deallocations of B. Finally, there are 3 types of surfaces in MIRA: solid, thin and transparent, where the rst is part of a body, with exterior/interior volume, the second has surface normal ambiguity, and the third a surface not participating as a scatterer but just a mathematical entity, e.g., useful when making GTD/MoM hybrids.

Ray
This section considers the contents of the Ray package and exposes its envisagable future components. In the package, a ray is dened by a starting point P and a direction D where both are in R3 . The rst thing is to use the distance t from P in the rays direction as parameter. Then we are only interested in points Q = (x,y,z) given by: Q(t) = P + t D. The internal data structure follows the above mathematical representation and allows to represent both a ray that start in the source and propagate as far as innity (semi-line or FAR eld radiation) or propagate to a nite point (segments, i.e., t belongs to a nite interval [a, b], or NEAR eld radiation). This means that irrespective of whether the observer is innitely far away or not, the data representation of the ray stays the same. A second feature is that the rays are grouped in lists, thus preparing any distribution of illumination tasks in a distributed computing environment. Linked lists are used to allow dynamic memory allocation and exible data storage. The Ray package performs the computation of R3 geometrical ray sheets that actually includes the following features: (i) reection by convex smooth surfaces, (ii) diraction by edges, (iii) multiple interaction and (iv) creeping rays launched from the shadow boundary of a convex smooth surface. 2

In a near future we will focus on two particular things. Firstly, to couple the Ray package with the NURBS trimmed representation of the surfaces in a more ecient way. Secondly we will focus on algorithms related to the following topics: (i) intersection/reection/diraction on concave/convex surfaces, (ii) diraction and creeping from an edge, (iii) multiple creeping-diracted-creeping rays, (iv) dielectric materials, (v) localization of caustics (region with high concentration of energy where standard GTD prediction fails) and (vi) sources/observers on the surface corresponding to on board antennas. The computation of characteristic points (intersection, reection, diraction, etc) is based on a numerical technique applied to the Generalized Fermats principle. This corresponds to trace a ray (Ray Tracing) between two points (source and observer) such that the optical path length is an extremum (maximum or minimum). Numerical methods are in generally fast but not reliable. They often miss to converge to the proper solution when a surface is both convex and concave or when a surface has more than one intersection with the ray. That is why we will investigate the possibility to couple a numerical method with a recursive subdivision method in order to isolate multiple solutions as well as to get good starting point for the minimization method. In addition, the MoM/PO/GTD hybrid demands an ecient method to detect hidden surfaces, ray occlusion by surface patches and surface patch visibility tests, so in this case, the modularity should save a considerable time during the assembly of all these technique together.

Application
The application package that we have implemented at present time considers antenna related computations. By convenience we also use the antenna package when we hybridize MIRA with a MoM solver. The main constituents of the application package is a set of receivers, sources and a module called attenuation that compute how much a GTD eld is attenuated a long any type of ray path. The types of sources we have implemented are plane waves, spherical harmonics, dipoles, antenna diagrams and an intrinsic type of source we call simple sources. All of these types of sources have support in GTD if they are treated in the proper way. The receivers are points, directions, antenna diagram and intrinsic simple receivers. The data structure in the package has been designed in such a way that one can easily build a complex antenna system, e.g. built up by a combination of several antenna elements distributed over any platform where each antenna element can in its turn be described by a collection of sub-sources. Receivers are built up in a similar way. When doing coupling calculations this is very useful since the user can easily build the receiving and emitting antenna systems. In addition the modularity permits one to easily add new types of sources and receivers without changing anything in the Geometry and Ray package. Probably we will in the future make use of the possibility to call other software packages inside the source routines to be able to numerically compute the radiated eld by a source. This may well come in hand when measured data is not attainable or when one needs higher accuracy. It is important to notice that this will be transparent in any other part of the code since all calls are done via interfaces. The two main advantages of the above data structure is that it well supports the already started hybridization with a MM solver and that it will, in the future, make the adaptation of the code to parallel computers relatively easy. The input data needed for the Application package are given by an object called Event which makes the links with the Ray package. An Event describes an entire generic ray path (reection path, diraction path, ...) and contains all the required geometrical information (normals, curvature at the characteristics points) needed to compute the GTD eld. The computationally most expensive part of the solver is to create these events. Having a list of all the events it is interesting to notice that they do not changes with frequency. This is in contrast to MoM which get a new set of equations for each frequency. On the other hand if the sources or receivers are moved the event data has to be re-computed. In MIRA, the event object is stored and never re-computed when frequency is changed. While hybridizing a ray tracer with a Method of Moment solver it is important to take this in consideration. Also when doing coupling calculations sometimes the receiver geometry will change with frequency and then the event object has to be re-computed. This will make such computations a bit more expensive compared to for example antenna diagram computations.

The global algorithm in MIRA (antenna application) constitutes of the following: Build the geometry. Build the sources and receivers. For all, or for one event, do: Compute a certain type of ray, e.g. a reection-diraction ray. Build the event. Do the eld computations: Compute the GTD eld from the source. Attenuate the GTD eld along the ray with the help of the event data and update the receiver. Finally Deallocate the geometry, the antennas, the events and the rays.. All eld computations are done in a module called attenuation. As argument is has an event, a source and a receiver.

Conclusions
We think that with the present MIRA one has a solver that is relatively robust and general both in terms of geometry and ray tracing. Thanks to the features presented in this article, MIRA can be used as a platform for future implementations of various kinds of GTD coecients. As a suggestion we think that it could be interesting to call either numerically constructed databases for dierent kinds of diraction coecients or to call numerical methods, e.g FDTD solvers, for computation of GTD coecient. With the modern geometry data structure that well supports the visualization and industrial CAD tools, that exist in the industry, MIRA will serve as an excellent starting point for new applications.

Acknowledgment
The authors would like to thank J. Gustafsson and C. Lundstedt at Ericsson Saab Avionics AB for providing us with interesting test geometries.

References
[1] J. Edlund et. al. An investigation of hybrid techniques for scattering problems on disjunct geometries. (ibid), 2000. [2] J. Perez et. al. Analysis of antennas on board arbitrary structures modeled by nurbs surfaces. IEEE Transactions on Antennas and Propagation, AP-45:10451052, June 1997. [3] M.F. Catedra et. al. Fasant theoretical foundations. technical report, signal theory and comm. dept., univ. of alcala, 1998. [4] G. Farin. Curves and surfaces for computer aided geometric design. acad. press, 1998.

Appendix A:

Figure 1: Direct and reected rays on a generic satellite. This model is composed of 90 NURBS surfaces and consisted of an union of simple objects, mostly cones and cylinders. A point source is situated above the satellite (view on the right) . The shortest direct and reected path between the source and the observer points underneath the satellite are searched for in the ray tracer. The direct and reected GTD eld are summed up at the observer points.

Figure 2: Direct and reected rays on four semi-transparent spheres modelled with 8 NURBS. One source point is placed in the middle, four observer points behind each sphere. At each observer point it reaches three rays: one direct ray from the source passing through the semi-transparent surfaces of the sphere, and two indirect reections coming from bounces on facing spheres.

Figure 3: Diracted rays on a generic cube. The diracted paths between one source point and one observer point above the cube occur at the middle of each edge. Diraction from the corners of the cube can be also identied in the ray tracer.

Figure 4: Diracted rays on a generic cube, top view.

Figure 5: Near to Far eld reection on a generic Gripen-like aircraft built up with large and complex surfaces. The model is composed of 50 NURBS. An antenna has been placed under the aircrafts left wing and act as a source point for the rays. The shortest reected paths from the wing, from both the upper side and the bottom side of the fuselage are displayed for several far eld directions.. Such a computation is performed by MIRA in a few minutes.

Figure 6: Near to Far eld reection on a generic aircraft, side view.

Appendix B:

 U

D 5 # 7  # v @ 5  8aCq&o3Cq taHAHhxCgAAHPH3HQ23AHCA3HA3XH8t23y3W32Hgx332CHp3C3 5 # 7 I @ 7 5 g  l I % @ % # 7 @ # I I S { D 5 @ @ I # I #" f S 5  5 I % w 5 @ @ I # g7" f @ v w @  " # v 5 S 5 # f  5 # 7  " I I  % # @" v " #% @ "" T T @ B 7 # S # i  " f @ @ # @ x223220HC!32P32333dt2a8HCHCf4Cs3Yr3z2Vr23ap3rAd! y w T % @  B 7 T g  7 I @ 7 5 {  @ g  T # T 7  B 7 l S @ v #  # 7 # @ w @ I % qx~Ye}aoV|XCCCahsQHpz(x8PHay3xpHHdU!2aAHCHgh3@ y w  v # 5  7 5 # I 5  I 5 I @  I 5 @ B # 5 7 # 5 @ @ B 7 v T # u D S @  I S g""  I 7    qx&UHU4ahHR8e3CHAH32R3H3223arCrUtCHHC3&288!a3S @2sH34Hp2aCCAHah2rCqdp43u6a4Hgh3mhUaHAHhs3PaoVnam23a7 S 7  # I 5 # 7 I @ 7 5 % # f  5 B 7 7  B 7 # @ I % @l g  5 # 7 I @ 7 5 5 # B 7 T g  l " 5  5 # 7 I @ 7 5 B 7 I " 5 B 7 5  I i D # 7 @ 5 #" f " 5  7 5  7 I  I % @ B I B T 5 # 5  7 CCCahfaeH8k3CH3Xfj3a32Hd23aX2thHHdCgA3At3VRf4HC3S # 7 " f   T g #" # # 7 5 S" f  5 @ y D 7%  I  B 7 #% y w  S 5  B @ B 7 #% @ S ae23&!dad3233!3AHQ8aC8&aqYRq39H33VYxC3H 233(qx$3Xp2aRAPCH23aqxx8vauthHAHH " 5  7  S 5  y w @ T g" 7 I   # D @ " 5  7 S 5  y w % # 5 # 7  7 5 @ " f # S  @  B B @ B i D 5 5  g f S 7  I S 5 " # S   # % 5 # 7  5  ""  # % B 33$s3rAH3X3qp332WHC$WhCa8H3e2d3cbY$2a83432`YXC33@ B 7 5 T # S  B @ @ 7%  I  I 5 E D B @  % # 5 # 7  5  "" ) '  % # "       3WVU36AR!Q8H2PAH3CGFC33A98(8643210(&$!

 HhdRh62a8aA2rx3q2a86d!32!3AH3V HA3HC!23q3!Coe23aC3A3Cz0!a2A2fHAYea3RH86!H3xpHHd3q tC3z2ha2A2VC8aq23tp&t8eq86Vht2aaq23CdHVzH234je

HA3q32A833AH3Ve234jp32

 Ak!2C28kCq2C2CHC3&HA33nHCrx C3th3 3p38GHa0$h!4AH3sC3AehaAd!2C3CC6CA 3ftHCHqCC3 C3C0HCC p 03YAH3!!ax8qACasA3CHp932AHHp&a2A29H8aasa 33CHdXCaCC p r2Cq48dHCgAr3AHaA8A!as34ax!

CahHC$CaCC6CAAYqH4H2a ahCn3qf3Hd3Ra2Hd3RAY!32!HcC693eCau93H93q9

HQ8aC23axh3HCsah2nH8te23VH34H 2sC6Ca83e33&Q8HApq$sHh342n63Cah(etauk3CHV rA3pe2XHA3RHCq6s2a8H3!3!haC4CzHut3HA3a3q

Architecture and Geometric Algorithms in MIRA, a Ray-based Electromagnetic Wave Simulator

Sandy Sefi, sandy@nada.kth.se Department of Numerical Analysis and Computer Science, Royal Institute of Technology KTH, SE:100 44 Stockholm, Sweden

Abstract In this paper we describe the basic architecture of the user-oriented electromagnetic simulator called MIRA which determines the near and far field from transmitting antennas and plane waves. The simulator is a part of the Swedish code development project GEMS: General ElectroMagnetic Solvers supported by PSCI, KTH, Saab Avionics AB, and NFFP and continues in the NFFP 473 project SMART. Keywords: Design Software, Geometrical Theory of Diffraction, Ray Tracing, Shadow Detection

INTRODUCTION

An electromagnetic wave simulator is used in Aerospace and Telecom areas to assist in the analysis of installed antenna performance and Radar Cross Section (RCS). Its aim is to predict the field emitted by antenna models on board large structures - buildings, aircrafts or ships - as realistically as possible. In the high-frequency band, asymptotic approximations to the Maxwell equations are suitable to evaluate the electromagnetic field. One of the powerful asymptotic methods known is the Geometrical Theory of Diffraction (GTD). The GTD [1] is based on Geometric Optics and Diffraction Theory. It assumes that all waves are well-formed and are locally plane waves. This enables Ray Tracing algorithms to be used with the following advantages: 1. It supplies a method to asymptotically compute the interaction of an antenna with a structure, when classical integral formulation methods become computationally too expensive. Ray-Tracing is geometric. The computational demands is not dependent on the electrical size of the structure but only on the complexity of the geometry. There is no runtime penalty in increasing the frequency. Ray-Tracing is not memory intensive. It is relatively easy to parallelize the underlying algorithm.

2.

3. 4.

MIRA is essentially a Ray-Tracer based on Fermats principle and works on complex CAD geometries where the objects are described by trimmed NURBS - Non Uniform Rational B-spline - and rational Bezier patches. NURBS is a standard for parametric surface representation [2]. Its Object Oriented design has been developed to support hybridization techniques with other numerical methods in the frequency domain such as Method of Moments (MoM) or Physical optics (PO). Our first hybrid application uses the Ray-Tracer module of MIRA as a stand-alone solver coupled with a PO/MoM solver [3] to determine accurately the shadow regions of the PO domain when it is illuminated by an antenna. The Method of Moments is used locally to compute the input behavior around the antenna. A similar technique is used when the wave is scattered. Below we will describe the different steps of the ray-based electromagnetic wave simulation, after that we will consider the basic features of the software architecture. Finally, we will give a brief overview of the Ray/PO hybrid.

THE RAY-BASED WAVE SIMULATION

The following gives a brief description of the physical process of the high-frequency wave propagation which is simulated: 1. Input of the geometry data (scene) consists of a collection of (i) surfaces in a trimmed NURBS format subdivided into a combination of Rational Bezier Patches by using the Cox De Boor algorithm [2], (ii) ElectroMagnetic (EM) sources and (iii) receivers. 2. The EM sources emit an incident field characterized by a certain direction, amplitude, phase and frequency. The field is represented by a wave. 3. The wave is absorbed, scattered, or reflected by the surfaces as it travels through the scene. The EM field is attenuated as the wave travels and the attenuation is proportional to the traveled distances. 4. The wave reflected or diffracted from a surface depends on properties of the surface (curvature, transparency, surface material...) as well as on incident illumination.

Simplified Input Scene (1.)

Incident field

Trimmed NURBS

Wave Propagation (2.)

Incident E ij

(3.)

Reflected E ij

Diffracted E ij

Figure 1 5.

High-frequency wave propagation.

The linearity of the Maxwell equations allows the superposition of the total EM field from the sum of independent electromagnetic contributions. Diffracted Reflected Incident Total = E ij + E ij + E ij E ij

6. 7.

The Total field E ij will be found as sum of the direct contributions from the source i, reflected and diffracted contributions reaching the receiver j. Each contribution is represented by rays traced from the source i to the receiver j

Total

The determination of the rays is done following Fermats principle. This corresponds to tracing a ray between the source and receiver such that the optical path length reaches an extremum (maximum or minimum). The characteristic points (intersection, reflection, diffraction, etc.) are found by numerical conjugate gradient techniques applied to Fermats principle. This part represents the most time consuming task in the solver since the ray-path determination process requires both rational Bezier surface point evaluations and computationally expensive evaluations of derivatives on NURBS surfaces.

PROGRAM DESIGN AND ARCHITECTURE

To design an architecture of a computer code is the task of choosing between various available alternatives and at the same time considering possible future developments and trends. The main problem is to design an architecture that is general enough to be useful for more than a single task, while providing an appropriate framework for development and efficient implementations. In this context, our aim is to develop a code capable of performing the calculations on fully realistic industrial models, to easily support future requirements or functionalities and to allow hybridization with other computational electromagnetic tools. HISTORY AND BACKGROUND The starting point for the architecture of MIRA was the FASANT code which was developed at Cantabria University by F. Ctedra and co-workers. From a computational electromagnetic point of view, this is a good old - quite reliable - code. Its functionalities and comparison with measurements can be found in [4]. A good comparison with the NEC/BSC code can be found in [5]. Apart from possessing these qualities, we found the Fortran 77 FASANT code difficult to modify and we decided to redesign the code by going back to the fundamental theory [6] and do a complete re-implementation using object oriented design with a Fortran 90 implementation. However, it is important to remark here that in the course of redesigning FASANT, we also introduced new algorithms. For example, we added diffraction algorithms, based on a search of combinations of local minima and local maxima, diffraction corner detection, double diffraction path determination, better visibility preprocessing taking into account transparent surfaces or dielectric sheets, etc.... The results and the description of the new features can be found in [7]. FEATURES OF THE ACTUAL ARCHITECTURE The following design features were leading up to a wave simulator architecture powerful and flexible enough to support hybrid applications: 1. Modular design

This work follows ideas from the Computer Graphics Society in the area of physical rendering processes [8]. These ideas [9] stipulate that a modular design should permit (i) the same general architecture to be configured for use in various application areas, (ii) the development of new algorithms and (iii) the implementation of variations of existing techniques reusing the already available environment. To fulfill it in our context, we have structured the architecture by a clear division of the complete system into subsystems called modules. The different modules have to be as independent of each other as possible and require well defined interfaces between them. The interfaces must be general enough not to limit the range of possible implementations. One major obstacle on this

theme are dependencies between modules, which need to be minimized. The details of this modular approach can be found in [10].
Other Modular Solver

Object Orientation
Classes inheritance

HYBRID CODE

Limited modules dependencies

Fa

Software with Modular Design Figure 2 2. Features of the software architecture

Object orientation.

Each object (3D surfaces, bounding curves, rays, antennas...) has well-defined interfaces consisting of a set of operations or methods that may be invoked by other objects. It also has internal attributes that represent the state of the object. The set of methods and attribute variables is described by the class of the object. Classes are manually organized in inheritance since automatic inheritance is not supported by Fortran 90 language. A class inherits behavior and attributes from its parents. In this way, the internal state of an object is encapsulated by the interface and can be manipulated by other objects only through the methods offered. This permits a clear separation between the interface of an object and its implementation. It defers the handling of implementation details from the early stages of the analysis and design process and therefore allows for better abstraction. 3. Good accuracy

The accuracy is only limited by the algorithms used and the parameters chosen, but not by the architecture itself. Since accuracy most often is coupled with the computational cost of the calculations, the architecture allows various options for trading accuracy against computational complexity. One way to offer this is to allow for different solution strategies for various parts of the system (c.f. flexibility). The generic characteristics of the new structure permit the coexistence of different geometrical representation. For instance, the initial free-formed surfaces (NURBS, rational Bezier etc....) coexist with the set of the plane facets/triangles constructing the mesh. Each surface is labeled and each

Ac

Flexibility switch

st

cu

rat

facet/triangle knows on which surface it belongs. All these links between data represent an important step toward improving the accuracy. In fact, the solver will work faster with only a facet description but will need to go back to geometrical information for tangents, radius of curvature, etc., to make precise field calculations. 4. Flexibility

Flexibility generally comes at a cost. In most cases a specific interface can be implemented with better performance than a general interface. In each case we have to decide if the additional flexibility is worth the price imposed by it or if a restricted interface offers substantially better performance. Sometimes, this might require the implementation of both the general interface and an interface with better performance but restricted functionality. We believe that in many cases and in the long run it is more important to provide a flexible interface than the implementation with the highest performance. 5. Handling of complex models

To be useful for real world applications an architecture must handle large and complex data sets. A fully realistic model is typically composed by more than one thousand NURBS surfaces. This requires the minimization of the storage requirements within the architecture. Practically, for the representation of the geometry data structure, linked chains of buffers dynamically allocated are preferred to one huge list of surfaces stored in a fixed sized matrix. 6. Modern programing style

Good coding allows for more efficient research and development. Current coding practices are in strict opposition to old fashion programming style characterized by goto loops, code duplication instead of subroutine calls and global variables used in many different parts of the code. Instead we assist on variable names that mean something, parentheses and white-space to make the code readable. 7. Robust design

We have significantly improved code quality and reliability thanks to systematic checking of input arguments, parameters and tolerances, division by zero etc.... 8. Efficiency thanks to topological information

The topology of an object is the list of the connections between faces (surfaces), edges (bounding curves) and vertices (corners) of a geometrical object. Most recent CAD products include topological facilities. A remaining problem is that engineers are used to exchanging data using IGES1 formats likes which do not transfer the topology [11]. In this context, we were faced with the task of extracting the necessary topological information directly from the geometry. The topology is used to improve the global efficiency of the computations. For instance, when examining diffraction phenomena, the topology permits the following edge classification: an edge can be (i) free: a boundary without neighboring surface, (ii) C0: continuity everywhere between neighboring surfaces, (iii) C1: not the same tangent plane for two surfaces or (iv) material discontinuity. This classification permits the removal of fictitious edges and speeds up the diffraction localization search [12] by directly focusing on the correct edge (C1). Edges are defined as Bezier curves. The classification is obtained from the study of normal discontinuities at the boundary curves between NURBS. The topology is pre-computed in a table. The table contains labeled information such as parentship relations with the surfaces and neighboring edge relations.

IGES, the Initial Graphics Exchange Specification, is a widely used CAD data exchange specification.

9.

Performance

Maybe the most important factor is to obtain acceptable performance. The Ray tracer does this by using simple algorithms, avoiding redundant calculation and unnecessary computations. For example, the measured execution time of the determination of the direct illumination for 100.000 rays launched through a scene composed of 400 NURBS surfaces, is less than 3 minutes on one node of an IBM RS6000 with 160 MHz, 256 MB of RAM. For comparison, on a very small example of a direct near field and diffraction calculation at six receivers illuminated from one source around a cylinder (only 6 NURBS), FASANT gets 0.37 Mflops (Million floating-point operations per second) for a total execution time (wall clock time) of 97,7 seconds. On the same problem, MIRA runs at 5.16 Mflops during a total execution time of 6.5 seconds, which makes it 15 times faster than FASANT. These results differ so much since we are using a new detection of diffraction path based on direct search of local minima and maxima (Fermats principle) instead of running a minima search and use an algorithm for discarding false solution. In addition this results also illustrates how important the topology can be to improve the code performance.

HYBRIDIZATION: SHADOW DETERMINATION

The coupling of the Ray Tracer module of the wave simulator to a PO solver reuses the ray module to determine accurately the parts of the non illuminated PO domain situated in the shadow regions of an emitting source. The hybridization improves the PO approximation by removing from the study all the geometrical details residing inside the shadow cast by a PO object on itself (self shadow) or by any blocker (other eventual object) between the PO domain and the emitting source (occlusion). THE HYBRID ALGORITHM To solve this problem, a test ray is launched from the barycenter of each PO facet/triangle. The ray is then traced back toward the point source. If the ray path to the source is not blocked by other surfaces, then the facet is directly illuminated, the surface current distribution is computed with J = 2nxH . A PO approximation of the scattered field can be obtained from the surface current J by integrating the PO fields and finally compute the RCS of the object under study. In our first version, only the incident field is used to illuminate (no reflection nor diffraction). This task remains however highly time consuming since a huge number of rays has to be traced. In the future, indirect illumination of the PO domain coming from the reflection of a ray on a surface, can easily be added. The two main characteristics of this algorithm are: 1. High computational complexity. There are (i x j) occlusion tests to perform. During one occlusion, several intersection tests must be executed with all the surfaces facing the points i or j. Each intersection test requires heavy NURBS derivations and surface point evaluations. The independency of computations between each source i and receiver j gives natural parallelization.The details of the parallelization realized with MPI library calls and implemented on the IBM SP with Power2 processors can be found in [13].

2.

ILLUSTRATION For the assumptions and algorithm presented above, the current distribution on a realistic aircraft is obtained. The tested geometries are realistic industrial models and have been

borrowed from Saab Avionics. The geometry is composed by 82.736 triangles/receivers that facetize 368 trimmed NURBS surfaces. Figure 3 represents the current distribution without taking into account the occlusion (this picture is given as reference to illustrate the PO solver working by its own). Figure 4 represents the illumination taking into account self shadow and occlusion. A single source, a plane wave perpendicular to the nose of the plane, illuminates 82.000 receivers placed on the aircraft surface.

Figure 3 PO only without Ray-tracing: current surface distribution on an aircraft from a plane wave pointing vector parallel to the aircraft center axis.

Figure 4 PO and shadow: illumination taking into account the self shadow of the aircraft. The dots represent triangles illuminated by a plane wave. In such orientation, half of the wing surfaces are occluded.

CONCLUSION

In this paper we have proposed an architectural approach to the implementation of an high frequency electromagnetic wave simulator. Its Object Oriented design is powerful and flexible enough to support a number of algorithms, yet can handle large and complex bodies. It has been developed to support hybridization techniques. In particular, we have given a detailed description of our recent Ray/PO hybrid. The method described combines several new concepts: use of topological information, double geometry representation: NURBS & Triangulation and accurate shadow detection taking into account self occlusion to a powerful Applied Electromagnetics software. Future improvements will focus on efficiency aspects for industrial applications.

References
[1] [2] [3] J. B. Keller, Geometrical Theory of Diffraction, J. of Optical Soc. of Americ., Vol. 52, No. 2, pp. 116-130, February, 1962. Gerald Farin, Curves and Surfaces for Computer Aided Geometric Design, Fourth Edition, Acad. Press, 1996. Johan Edlund, Stefan Hagdahl and Bo Strand, An investigation of hybrid techniques for scattering problems on distinct geometries, Technical Report PSCI No. 2000:08, KTH, March, 2000. J. Prez, F. Saiz, J., O. Gutierrez, I. Gonzalez, M.F. Ctedra, I. Montiel, J. Guzmn, FASANT: Fast Computer Tool for the Analysis of Antennas OnBoard Antennas, IEEE Antennas and Propagation Magazine, Vol. 41, No. 2, June, 1999. J. Prez, J.A. Saiz, O. Conde, R.P. Torres, M.F. Ctedra, Analysis of Antennas on Board Arbitrary Structures Modelled by NURBS Surfaces, IEEE Transactions on Antennas and Propagation, Vol. 45, No. 6, June, 1997. M.F. Ctedra et. al., FASANT (Version S.4) Theoretical Foundations, Technical report, Signal Theory and Comm. Dept., Univ. of Alcal, 1998. Sandy Sefi, Design and Architecture of MIRA, a Ray-based Electromagnetic Code, DRTs thesis, University Joseph Fourier, Grenoble, France, October 2000. Philipp Slusallek, Hans-Peter Seidel, Vision: An architecture for global illumination calculations, IEEE Transactions on Visualization and Computer Graphics, 1(1):77--96, March, 1995. Philipp Slusallek, Vision - An Architecture for Physically Based Rendering, PhD thesis, University of Erlangen, IMMD IX, Computer Graphics Group, April, 1995. Fredrik Bergholm, Stefan Hagdahl, Sandy Sefi, A modular Approach to GTD in the Context of Solving Large Hybrid Problems, proceeding publication in AP2000 Millennium Conference on Antennas & Propagation, Davos, Switzerland, April, 2000. Patrick Chenin, Geometric Modelling for ElectroMagnetic Simulation, LMC-IMAG Unival S.A., August 1998. Fredrik Bergholm, Locating Diffraction Points with Ray-based Methods, Technical Report PSCI, KTH, December, 2000. Sandy Sefi, Parallelization of a Ray-based Electromagnetic Wave Simulator using MPI, PDC High Performing Computing Course Report, KTH, August, 2001.

[4]

[5]

[6] [7]

[8]

[9]

[10]

[11] [12] [13]