Computational Electromagnetics and Model Validation

Introduction

Computational electromagnetics (CEM) converts Maxwell’s equations, material laws and boundary conditions into numerical models that can predict electromagnetic fields, currents, impedances, scattering parameters and radiation characteristics. It is indispensable when geometry, material inhomogeneity or coupling makes an analytical solution impractical.

A numerical result is not automatically a trustworthy physical prediction. Numerical approximations introduce discretization, iteration, truncation and round-off errors, while physical models introduce uncertain material data, imperfect geometry and simplified boundary conditions. Consequently, professional CEM requires both verification—determining whether the equations were solved correctly—and validation—determining whether the correct equations and assumptions were used for the intended application.

Learning Objectives

After studying this chapter, the reader should be able to:

• formulate electromagnetic boundary-value problems from Maxwell’s equations;
• derive representative FDTD update equations and apply the Courant stability condition;
• construct the electromagnetic finite-element weak form;
• formulate a representative method-of-moments integral equation;
• select an appropriate computational method for a given application;
• interpret ports, scattering parameters, return loss and radiated power consistently;
• quantify numerical convergence and distinguish error from uncertainty;
• design verification and validation activities supported by defensible evidence; and
• evaluate contemporary reduced-order, surrogate and AI-assisted electromagnetic models.

1. From Maxwell’s Equations to a Computational Model

For linear, isotropic media, the time-domain macroscopic equations are

Here, E is electric field intensity in V/m, H is magnetic field intensity in A/m, D is electric flux density in C/m², B is magnetic flux density in T, J_s is impressed current density in A/m², ρ is free charge density in C/m³, σ is conductivity in S/m, ε is permittivity in F/m, and μ is permeability in H/m.

Using RMS phasors with time dependence e^jωt, Maxwell’s curl equations become

Eliminating H gives the electric-field wave equation

This equation assumes linear media and angular frequency ω in rad/s. Anisotropic media require tensor-valued ε, μ or σ. Nonlinear and dispersive materials require additional constitutive equations.

1.1 Boundary and interface conditions

Common boundaries include:

• Perfect electric conductor (PEC): n × E = 0.
• Perfect magnetic conductor (PMC): n × H = 0.
• Material interface without surface sources: tangential E and H are continuous; normal D and B satisfy the corresponding flux conditions.
• Impedance boundary: approximates penetration into a good conductor when its curvature and thickness permit a surface model.
• Radiation boundary or perfectly matched layer (PML): represents an open domain.
• Periodic or Floquet boundary: models periodic structures with a specified phase shift.
• Waveguide port: excites and measures one or more modal fields on a reference plane.

The computational problem is incomplete unless excitation, material dispersion, source normalization and all boundary conditions are specified.

2. Discretization Families

Hybrid methods combine complementary descriptions—for example, FEM near an antenna and boundary integral equations in the surrounding unbounded region.

3. Finite-Difference Time-Domain Method

3.1 Yee staggered grid

FDTD approximates Maxwell’s curl equations directly. In the Yee grid, electric-field components are located on cell edges and magnetic-field components on cell faces. Electric and magnetic samples are also staggered by half a time step. This arrangement supplies natural centered approximations to curl operators.

Figure: Conceptual spatial staggering of field components in a Yee cell.

3.2 Explicit update equations

For a source-free, lossless, isotropic medium, Ampère’s law gives

A centered Yee-grid update is

Faraday’s law similarly gives

The remaining components follow by cyclic permutation. Conductive media require a discretization of σE; dispersive media may be represented by auxiliary differential equations, recursive convolution or equivalent pole-residue models.

3.3 Stability and numerical dispersion

For a uniform orthogonal grid in a nondispersive medium, a sufficient three-dimensional Courant–Friedrichs–Lewy condition is

where v_max is the maximum physical wave speed represented in the model. The condition controls stability, not accuracy. Numerical phase velocity remains dependent on cell size, propagation direction and frequency.

Worked example—FDTD time step. A cubic grid has ∆x=∆y=∆z=1.00\ \text{mm} and includes air, so v_max≈2.998×10⁸\ \text{m/s}. Therefore,

A 0.99 safety factor gives ∆t=1.907\ \text{ps}. Choosing the slower wave speed of a dielectric region instead of the fastest speed would produce an unsafe time step when air is also present.

PML performance must be checked separately. Too few PML cells, abrupt conductivity grading or nearby evanescent fields can create measurable reflections.

4. Finite-Element Formulation

FEM partitions the domain into elements and approximates fields using local basis functions. For electromagnetic vector fields, curl-conforming Nédélec edge elements are generally preferred because they enforce tangential continuity and suppress many nonphysical modes associated with naive nodal vector interpolation.

4.1 Weak form

Consider the frequency-domain equation

under the e^jωt convention. Let the admissible field and test function v belong to H(\mathrm{curl},Ω). Multiplication by v^*, integration over Ω, and curl integration by parts give

On a PEC boundary, both the trial field and test function have zero tangential component, so the corresponding boundary term vanishes. Radiation, impedance or port boundaries contribute nonzero terms and must be incorporated consistently.

With E≈∑_m=1^Ne_mN_m, where N_m are edge basis functions, Galerkin testing yields

where

The vector e contains field degrees of freedom and f includes sources and boundary excitations. The matrices are sparse because each basis function has local support.

4.2 Eigenvalue and driven problems

In a lossless closed cavity without sources, FEM produces the generalized eigenproblem

In driven frequency-domain analysis, the matrix is solved independently at selected frequencies. Near resonance, conditioning can deteriorate; iterative solvers may require preconditioning, while direct solvers may require substantial memory.

Adaptive FEM uses an error indicator to refine elements where field gradients, geometric singularities or material contrasts make the largest contribution to the estimated error. Refining only globally can waste degrees of freedom.

5. Method of Moments and Integral Equations

MoM is especially effective for conducting antennas and scatterers in open space because the Green function satisfies the radiation condition without meshing the entire surrounding volume.

For a PEC surface S, the electric-field integral equation enforces

For the e^jωt convention, the outgoing free-space Green function is

Using the surface current density J_s, charge continuity jωρ_s+∇'_s·J_s=0, and potential representations gives

The EFIE is obtained by substituting this field into the PEC boundary condition. Expand the unknown current as

where f_n may be Rao–Wilton–Glisson basis functions on triangular meshes. Testing with functions w_m produces the dense linear system

where \mathcal{L} denotes the integral operator and V_m represents the incident-field excitation. Singular and near-singular integrals require specialized quadrature. Low-frequency breakdown, internal resonances and dense-matrix storage are important limitations. The combined-field integral equation mitigates internal-resonance difficulties for closed PEC scatterers, while fast multipole and hierarchical matrix methods reduce computational cost for large systems.

6. Sources, Ports and Observable Quantities

A model should produce observables that correspond to measurable engineering quantities. Examples include port voltages, currents, scattering parameters, input impedance, radiation pattern, directivity, gain, absorbed power and specific absorption rate.

6.1 RMS power-wave convention

Power formulas depend on the phasor convention. In this chapter, port voltages and currents are RMS phasors, and current is positive into the port. For a real positive reference impedance Z₀, define

Then |a|² and |b|² are incident and reflected powers in watts, and

If peak rather than RMS phasors are used, average power is \tfrac12\operatorname{Re}(VI^*), and corresponding power-wave normalization must include that factor. For complex reference impedances, generalized power-wave definitions are required; the simple real-Z₀ formulas must not be applied indiscriminately.

With all other ports terminated in their reference impedances,

Return loss RL is conventionally reported as a nonnegative number for |S₁₁|≤1.

Worked example—return loss and reflected power. If |S₁₁|=0.03981,

For 1 W incident power, the reflected power is

Thus 1.585 mW is reflected and 99.8415% is accepted at the port. Accepted power may subsequently be radiated, dissipated or coupled to another port; it is not automatically equal to radiated power.

7. Verification: Solving the Equations Correctly

Verification addresses numerical implementation and solution accuracy. It has two complementary forms:

1. Code verification: Does the software correctly implement the mathematical algorithm?
2. Calculation verification: Is the numerical error in this particular simulation acceptably small?

7.1 Code-verification techniques

Useful techniques include comparison with analytical solutions, benchmark problems, unit tests for element matrices and material models, conservation checks, and the method of manufactured solutions (MMS). In MMS, a sufficiently smooth artificial field is selected, substituted into the governing equations, and the resulting source and boundary data are imposed. The measured convergence rate can then be compared with the theoretical order of the scheme.

MMS tests implementation consistency but does not demonstrate that the physical model is valid.

7.2 Mesh and time-step convergence

Suppose a scalar output φ(h) is computed on successively refined meshes with representative spacing h. In the asymptotic range,

where φ₀ is the continuum estimate, C is a constant and p is the observed order. For three meshes with constant refinement ratio r=h₂/h₁=h₃/h₂>1, where mesh 1 is finest, monotonic convergence permits the estimate

This expression is unreliable under oscillatory convergence, insufficient refinement, changing topology or inconsistent solver tolerances.

Worked example—observed order. Resonant frequencies computed on meshes h, 2h and 4h are respectively 9.980, 9.920 and 9.800 GHz. Then

First-order observed convergence suggests that geometry approximation, a singular field region or another first-order mechanism dominates, even if the nominal field basis has higher order.

7.3 Other verification checks

A defensible study should also examine:

• residual norms and iterative-solver tolerances;
• energy or power balance;
• reciprocity for reciprocal linear systems, such as S₂₁=S₁₂ under consistent port normalization;
• passivity, implying that a passive network cannot generate net power;
• sensitivity to domain truncation, PML thickness and port location;
• numerical quadrature and matrix conditioning; and
• reproducibility across software versions and computing platforms.

Agreement between two tools is useful but not conclusive when they share the same assumptions, algorithms or material database.

8. Validation: Representing Physical Reality Adequately

Validation compares model predictions with experimental observations for a specified context of use. A model may be valid for predicting resonance frequency within 1% but inadequate for predicting local peak electric fields or thermal deposition.

8.1 Validation hierarchy

Validation is strongest when built progressively:

1. material-property and component tests;
2. canonical subassemblies;
3. subsystem experiments;
4. full-system operational conditions.

This hierarchy helps identify which assumptions cause disagreement. Measurements should be independent of the simulation where practicable, and model parameters should not be repeatedly tuned against the same data later presented as independent validation.

8.2 Experimental comparison and uncertainty

Let y_s be a simulated quantity and y_m its measured counterpart. Their discrepancy is

If independent standard uncertainties from simulation inputs/numerics and measurement are u_s and u_m, a basic combined standard uncertainty is

This relation assumes independence and approximately standard uncertainty representations. Correlated contributions require covariance terms. Compatibility may be judged against a predefined coverage factor k, such as |δ|≤ku_c, but the choice of k and acceptance limit must be justified by the decision risk and application—not selected after seeing the result.

Validation data should include calibration traceability, fixture effects, cable and connector de-embedding, positioning tolerance, environmental conditions, repeatability and uncertainty in material properties.

Figure: Verification and validation form an iterative evidence-based process.

9. Illustrative Case: Validation of a 10 GHz Microstrip Antenna

Background

An engineering team is developing a 10 GHz microstrip patch antenna for a point-to-point sensor link. This is an illustrative case. The design requirement is a measured return loss of at least 15 dB from 9.9 to 10.1 GHz and realized broadside gain above 6 dBi.

Problem

Initial FEM simulation predicts resonance at 10.02 GHz and return loss of 24 dB. The first prototype resonates at 10.15 GHz, outside the required band. Treating this as mere measurement noise would be unjustified.

Analysis

The team identifies uncertain substrate relative permittivity, dielectric thickness, copper roughness, feed-pin dimensions, connector launch geometry and manufacturing tolerances. The original simulation used nominal permittivity and an ideal wave port placed too close to the feed discontinuity.

A mesh study shows that further refinement changes resonance by only 8 MHz; numerical discretization is therefore too small to explain the 130 MHz discrepancy. Moving the port and extending the feed line changes the prediction by 22 MHz. Material characterization using a resonator shows that the effective substrate permittivity differs from the supplier’s nominal value.

Approach and solution

The team:

1. uses edge-element FEM with local refinement at the feed and patch edges;
2. demonstrates frequency convergence and checks accepted, radiated and dissipated power balance;
3. measures substrate permittivity and loss tangent at the operating frequency;
4. models the connector launch rather than replacing it with an idealized lumped feed;
5. calibrates the vector network analyzer and de-embeds the test cable to the connector reference plane;
6. records dimensional tolerances and propagates them through a designed set of simulations; and
7. reserves a second independently manufactured antenna for validation rather than parameter calibration.

The revised model predicts 10.11 GHz for the second antenna; measurement gives 10.08 GHz. If the independently assessed standard uncertainties are 0.03 GHz for simulation and 0.04 GHz for measurement, then

The discrepancy is below one combined standard uncertainty. The measured minimum return loss is 18.5 dB and broadside realized gain is 6.3 dBi, satisfying the stated requirements.

Outcome and learning points

The original discrepancy was dominated by model-form and input uncertainty rather than mesh error. The case demonstrates that mesh refinement alone is not validation; port definitions, material characterization, fabrication tolerances and measurement reference planes can control the result. Independent validation data are also more credible than data used to tune the model.

10. Limitations, Governance and Contemporary Practice

10.1 Persistent technical challenges

Important challenges include multiscale geometry, highly resonant structures, dispersive and nonlinear materials, moving boundaries, plasma coupling, stochastic manufacturing variation and electrically large systems. A model may also be computationally converged but physically misleading because of an omitted cable, support structure, contact resistance or environmental interaction.

Best practice includes maintaining a model-assumption register, recording software and solver versions, preserving meshes and input files, automating regression tests, defining quantities of interest before simulation, and reporting negative or inconclusive validation results.

10.2 Reduced-order models and AI

Reduced-order models project a high-dimensional electromagnetic system onto a smaller basis, accelerating optimization and uncertainty studies. Surrogate models—including Gaussian processes, polynomial chaos expansions and neural networks—can approximate relationships between design variables and field-derived outputs.

Such models require training data that cover the intended parameter domain. Cross-validation only evaluates performance relative to the available dataset; it does not establish compliance with Maxwell’s equations or validity outside the sampled domain. Physics-informed learning can introduce governing equations into training, but conservation, passivity, causality and uncertainty still require explicit testing.

10.3 Ethical and regulatory implications

Electromagnetic predictions may support decisions involving human exposure, electromagnetic compatibility, medical devices, aerospace systems and communications infrastructure. Selective reporting of favorable meshes or measurements can conceal risk. Proprietary solvers do not remove the engineer’s responsibility to disclose assumptions, uncertainty and applicability limits.

Models used for compliance should follow the relevant jurisdictional standards for exposure, EMC, calibration and test configuration. Sensitive antenna, radar or defense models also require appropriate cybersecurity, export-control and access governance.

Chapter Summary

Computational electromagnetics transforms Maxwell’s equations into discrete systems while preserving constitutive laws, sources and boundary conditions. FDTD provides explicit broadband time-domain updates subject to stability and dispersion constraints. FEM uses a weak formulation and curl-conforming basis functions to handle complex geometries and materials. MoM converts integral equations into dense systems and is especially effective for open-region conducting structures.

Reliable interpretation requires consistent source, port and RMS phasor conventions. For example, |S₁₁|=0.03981 corresponds to 28.00 dB return loss and 0.1585% reflected power. Verification evaluates implementation and numerical error through analytical benchmarks, manufactured solutions, convergence studies and conservation tests. Validation evaluates physical adequacy for a stated context of use through independent experiments and uncertainty analysis. Trustworthy practice integrates both activities with traceable documentation, transparent assumptions and responsible use of reduced-order and AI-assisted models.

Key Terms

Computational electromagnetics: Numerical solution of electromagnetic field problems derived from Maxwell’s equations.

Verification: Assessment of whether the mathematical model has been implemented and solved correctly.

Validation: Assessment of whether a model adequately represents reality for a specified use.

FDTD: An explicit time-domain finite-difference method using staggered field samples.

Yee grid: Spatial and temporal arrangement of electric and magnetic field components used in FDTD.

Courant condition: Time-step restriction required for stability of an explicit discretization.

Numerical dispersion: Dependence of computed phase velocity on grid, frequency and propagation direction.

Finite-element method: A variational method based on elementwise basis functions and a weak formulation.

Weak form: Integral form of a differential equation used to reduce differentiability requirements.

Edge element: Curl-conforming finite element whose degrees of freedom represent tangential field behavior.

Method of moments: Weighted-residual discretization commonly applied to electromagnetic integral equations.

EFIE: Electric-field integral equation enforcing the tangential electric-field boundary condition on a conductor.

Perfectly matched layer: Artificial absorbing region used to truncate an open computational domain.

Scattering parameter: Complex ratio relating outgoing and incident modal power waves at network ports.

Return loss: Negative logarithmic measure, -20\log₁₀|S₁₁|, of input reflection magnitude.

Manufactured solution: Chosen analytical field used to generate sources for code-verification testing.

Mesh convergence: Approach of numerical outputs toward a limiting value under systematic refinement.

Uncertainty quantification: Evaluation of how uncertain inputs and model limitations affect predicted outputs.

Review Questions

Short Answer Questions

1. Distinguish code verification, calculation verification and validation.
2. Why are edge elements preferred over ordinary nodal vector elements in electromagnetic FEM?
3. State the three-dimensional FDTD Courant condition for an orthogonal grid.
4. What physical boundary condition defines the EFIE on a PEC surface?
5. Why does accepted port power not necessarily equal radiated power?

Long Answer Questions

1. Derive the frequency-domain electric-field wave equation under the e^jωt convention, including conductivity.
2. Derive representative Yee-grid updates for one electric and one magnetic field component and discuss stability and numerical dispersion.
3. Develop the FEM weak form of the vector wave equation and explain the roles of boundary terms and curl-conforming spaces.
4. Formulate the PEC electric-field integral equation and explain its method-of-moments discretization.
5. Design a verification and validation plan for a microwave filter, including convergence, measurement uncertainty and acceptance criteria.

Analytical / Application Questions

1. An FDTD model uses cells of 0.8 mm, 1.0 mm and 1.2 mm and contains air. Calculate the maximum stable time step and propose a justified operating time step.
2. A two-port simulation reports |S₁₁|=0.08, |S₂₁|=0.94 and negligible radiation. Calculate reflected and transmitted power fractions and assess whether the remaining fraction is physically plausible as dissipation.
3. Three successively refined FEM meshes with refinement ratio 2 predict resonances of 5.120, 5.180 and 5.210 GHz from coarsest to finest. Estimate the observed order and discuss what additional evidence is required before claiming validation.

Multiple Choice Questions with Answers

Question 1

A one-port model has |S₁₁|=0.10. What is its return loss?

A. 10 dB

B. 20 dB

C. 40 dB

D. 0.1 dB

Correct Answer: B. 20 dB

Question 2

Which statement correctly distinguishes validation from verification?

A. Validation tests numerical convergence only.

B. Validation proves that a model is exact for every application.

C. Validation assesses physical adequacy for a specified context of use.

D. Validation is identical to checking solver residuals.

Correct Answer: C. Validation assesses physical adequacy for a specified context of use.

Question 3

For a cubic FDTD grid with cell size ∆ and maximum wave speed v_max, the three-dimensional Courant limit is

A. ∆t≤∆/(v_max\sqrt{3})

B. ∆t≤\sqrt{3}∆/v_max

C. ∆t≤v_max∆/3

D. ∆t≤1/(v_max∆\sqrt{3})

Correct Answer: A. ∆t≤∆/(v_max\sqrt{3})

Question 4

Which finite-element basis is generally appropriate for the electric field in a three-dimensional curl-curl formulation?

A. Piecewise constant scalar basis

B. Nédélec curl-conforming edge basis

C. Fourier basis restricted to one node

D. Discontinuous scalar nodal basis without constraints

Correct Answer: B. Nédélec curl-conforming edge basis

Question 5

Why is a conventional surface MoM model attractive for an isolated PEC antenna in free space?

A. It eliminates all matrix storage.

B. It requires no basis functions.

C. It meshes only the conducting surface and incorporates radiation through the Green function.

D. It is always explicit in time.

Correct Answer: C. It meshes only the conducting surface and incorporates radiation through the Green function.

Question 6

Under the chapter's RMS convention and a real reference impedance, an incident power wave has |a|=2\ \sqrt{W}. What incident power does it represent?

A. 1 W

B. 2 W

C. 4 W

D. 8 W

Correct Answer: C. 4 W

Question 7

A mesh-converged simulation disagrees substantially with an experiment. Which conclusion is most defensible?

A. The experiment must be wrong.

B. Mesh convergence alone does not establish physical validity.

C. The governing equations have been experimentally proved incorrect.

D. Further mesh refinement must remove every discrepancy.

Correct Answer: B. Mesh convergence alone does not establish physical validity.

Question 8

What is the primary purpose of the method of manufactured solutions?

A. To determine measured material permittivity

B. To validate a model against operational hardware

C. To test code implementation and observed convergence order

D. To replace all physical experiments

Correct Answer: C. To test code implementation and observed convergence order

Question 9

For independent standard uncertainties u_s=0.03\ \text{GHz} and u_m=0.04\ \text{GHz}, what is the combined standard uncertainty?

A. 0.01 GHz

B. 0.035 GHz

C. 0.05 GHz

D. 0.07 GHz

Correct Answer: C. 0.05 GHz

Question 10

Which practice most directly preserves the independence of validation evidence?

A. Using the same measurement repeatedly to tune and validate the model

B. Selecting the acceptance threshold after observing results

C. Reserving a separate experiment that was not used for parameter calibration

D. Omitting measurement uncertainty from the comparison

Correct Answer: C. Reserving a separate experiment that was not used for parameter calibration