Computational Design, Analytics and AI

Introduction

Microwave engineering has traditionally relied on analytical electromagnetics, empirical design charts, laboratory prototyping, and numerical simulation. Modern systems—such as phased-array antennas, millimetre-wave transceivers, satellite payloads, automotive radar, and 5G/6G front ends—contain many interacting variables and stringent performance constraints. Their design therefore increasingly depends on computational methods that combine electromagnetic simulation, optimization, statistical analysis, surrogate modelling, and artificial intelligence (AI).

Computational design does not replace physical understanding. Rather, it provides a systematic means of exploring design spaces, estimating performance, identifying trade-offs, and reducing the number of costly prototypes. The central engineering challenge is to ensure that computational predictions remain physically meaningful, numerically stable, reproducible, and experimentally validated.

Learning Objectives

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

1. Explain the role of computational electromagnetics in microwave design.
2. Distinguish deterministic simulation, optimization, surrogate modelling, and machine learning.
3. Formulate microwave design as a constrained multi-objective problem.
4. Select suitable numerical methods and assess mesh, convergence, and model-form errors.
5. Construct and evaluate surrogate models using appropriate data and metrics.
6. Apply AI responsibly to inverse design, diagnosis, calibration, and signal analytics.
7. Interpret a practical computational design workflow and its limitations.

1. Computational Design as an Engineering Discipline

A microwave design problem can be represented by a vector of design variables

where variables may include geometry, substrate properties, material parameters, excitation conditions, or bias values. A simulator or measurement system maps these variables to responses:

Here, f may be a full-wave electromagnetic solver, a circuit model, a measurement campaign, or a hybrid of these. The response vector y may contain resonant frequency, bandwidth, gain, efficiency, sidelobe level, isolation, insertion loss, or power-handling metrics.

A useful computational workflow contains four linked activities:

• Modelling: translating physical structure and boundary conditions into equations.
• Simulation: solving the equations numerically.
• Analytics: extracting trends, uncertainty, sensitivities, and trade-offs.
• Decision and validation: selecting a design and comparing it with experiment.

The quality of the result depends on all four. A sophisticated AI model trained on inaccurate simulations cannot produce a trustworthy design.

2. Numerical Foundations of Microwave Simulation

2.1 Governing equations

For time-harmonic fields with the convention e^jωt, Maxwell’s curl equations in a linear medium are commonly written as

where E and H are electric and magnetic field phasors, ω is angular frequency in rad/s, μ is permeability in H/m, ε is permittivity in F/m, and J is impressed current density in A/m². For conductive materials, conductivity σ may be included through the complex permittivity

for this phasor convention. The equations apply under linear, locally defined material constitutive relations. Nonlinear, dispersive, anisotropic, or temperature-dependent materials require corresponding constitutive models.

Microwave solvers usually reduce these equations to a discretized algebraic system:

where A is the system matrix, u contains unknown field coefficients, and b represents sources and boundary excitation. The matrix depends on frequency and design variables. The system may be sparse but large, making memory management and solver selection important.

2.2 Major numerical methods

Finite-difference time-domain (FDTD) discretizes space and time on a grid. It is effective for broadband transient problems and naturally produces time-domain responses. However, curved surfaces and very small geometric features can require expensive fine grids. The Courant stability condition constrains the time step. In a simple three-dimensional Cartesian grid, a sufficient condition is approximately

where Δx, Δy, and Δz are cell dimensions and c is wave velocity in the relevant reference medium. The exact stability limit depends on the discretization and material model.

Finite-element methods (FEM) use local basis functions on a mesh of elements. They represent complex geometries and inhomogeneous materials well and are widely used for resonators, filters, and antennas. Frequency-domain FEM normally requires a new solution for each frequency, although model-order reduction can accelerate repeated evaluations.

Method of moments (MoM) converts integral equations into a matrix equation, often using Green’s functions. It is efficient for open-region radiation and conducting surfaces, but dense matrices can create substantial memory and computational demands.

Finite integration and hybrid methods combine features of several methods. Domain decomposition, physical optics, ray techniques, and circuit-field co-simulation are valuable when a complete system contains electrically large and electrically small regions.

2.3 Accuracy and verification

A numerical result is not automatically a correct result. Three distinct questions should be asked:

1. Verification: Was the mathematical problem solved correctly? Examples include mesh refinement, residual checks, energy balance, and comparison with canonical solutions.
2. Validation: Does the mathematical model represent the physical device? This requires measured material data, realistic losses, connector models, manufacturing tolerances, and experimental comparison.
3. Uncertainty assessment: How sensitive is the result to uncertain inputs and approximations?

For a computed quantity q, a mesh-convergence study may compare successive estimates q_h and q_h/2. A practical relative change is

where h is a characteristic mesh size. A small δ_q supports convergence, but does not prove physical validity. Resonant frequencies, narrow notches, and phase quantities often require stricter refinement than broad amplitude responses.

3. Formulating Microwave Design as Optimization

3.1 Objective and constraint formulation

A general design problem is

subject to

and bounds

The vector F contains one or more objectives. Constraints may specify operating frequency, impedance matching, minimum efficiency, maximum dimensions, fabrication rules, or thermal limits. Units should be normalized before combining unlike quantities.

For a single weighted objective, one possible formulation is

where fᵢ* is a target, sᵢ is a scaling quantity with the same units as fᵢ, and non-negative wᵢ are weights. This expression is useful when deviations from targets are more important than absolute values. Weight selection is an engineering decision and can bias the outcome.

3.2 Local and global optimization

Gradient-based methods are efficient when the response is smooth and derivatives are available. They can converge rapidly near a good solution but may become trapped in local minima or fail when the response has discontinuities caused by mode switching or meshing.

Derivative-free methods—including pattern search, Nelder–Mead, particle swarm optimization, and differential evolution—are useful when simulations are noisy or derivatives are unavailable. Their disadvantage is often a high number of function evaluations.

Bayesian optimization is attractive when each simulation is expensive. It uses a probabilistic surrogate, commonly a Gaussian process, to select the next point through an acquisition function. A common expected-improvement expression is

for minimization, where f_best is the best observed objective and the expectation is taken under the predictive distribution. The exact form depends on predictive mean, variance, and treatment of observation noise.

3.3 Multi-objective design and Pareto optimality

Many microwave designs involve unavoidable trade-offs: increasing bandwidth may reduce efficiency; reducing size may increase loss; suppressing sidelobes may reduce gain. A design x_a dominates x_b if it is no worse in every objective and strictly better in at least one. A design is Pareto-optimal if no feasible design dominates it. The set of such designs forms the Pareto front.

The Pareto front is often more informative than a single weighted solution because it allows designers and stakeholders to examine trade-offs explicitly.

4. Design of Experiments and Data Generation

AI and surrogate methods depend on data quality. Random sampling alone may leave important regions unexplored. Latin hypercube sampling distributes samples across each variable range, while factorial designs are useful for estimating main effects and interactions at selected levels. Adaptive sampling adds new simulations where uncertainty is large or where predicted optima occur.

The sampling range must be physically meaningful. A model trained only on nominal substrate thickness may fail when fabrication variation changes the thickness. Conversely, including impossible geometries wastes simulations and can confuse the learner.

Data preparation commonly includes:

• removing failed or physically invalid simulations after diagnosis;
• retaining failure labels when solver failure itself is informative;
• scaling variables with different units;
• separating training, validation, and test data;
• avoiding leakage between nearly identical geometry families;
• recording solver version, mesh settings, material data, and random seeds.

A random split is not always sufficient. For a family of related geometries, a grouped split—where entire geometry families are held out—gives a more realistic estimate of generalization.

5. Surrogate Modelling and Analytics

A surrogate is a computationally inexpensive approximation of a costly response. Polynomial regression, radial-basis functions, Gaussian processes, support-vector regression, random forests, and neural networks are common choices.

A Gaussian-process model represents a response as a mean function plus a correlated random process. Its kernel describes similarity between input points. For training data X and observations y, a noise-inclusive covariance matrix is

where K is the kernel matrix, σ_n² is observation-noise variance, and I is the identity matrix. The predictive mean and variance depend on the selected kernel and hyperparameters. These assumptions are valuable when uncertainty estimates are needed, but a poorly chosen kernel can misrepresent sharp resonances.

Neural networks can approximate highly nonlinear mappings. For a feed-forward layer, the operation is

where z is the input vector, W is a weight matrix, b is a bias vector, and φ is an activation function. Training adjusts weights by minimizing a loss function, often using gradient-based optimization.

Model quality should be assessed using metrics appropriate to the task. For measured values y_i and predictions ŷ_i, the mean absolute error is

and the root mean-square error is

Both have the units of the response. RMSE gives greater emphasis to large errors. For a response whose mean is not close to zero, the coefficient of determination is

where ȳ is the sample mean. R² can be misleading for extrapolation and should never be used alone.

Sensitivity and uncertainty

Local sensitivity describes how a response changes with a variable:

Because units affect this value, a normalized sensitivity is often preferable:

provided x_i and y are nonzero. Global sensitivity methods, such as variance-based indices, assess effects throughout the design domain rather than at one operating point.

If independent input uncertainties are small and the response is approximately linear near its nominal point, a first-order variance estimate is

This approximation fails for strong nonlinearity, correlated inputs, or large variations. Monte Carlo or polynomial-chaos methods may then be more appropriate.

6. AI Applications in Microwave Engineering

6.1 Inverse design

Forward design predicts responses from geometry. Inverse design seeks geometry from desired responses. Because multiple structures can produce similar responses, inverse problems may be non-unique and ill-conditioned. A robust workflow first trains a forward surrogate, then optimizes through it, and finally verifies candidates using a high-fidelity solver.

Generative models and differentiable electromagnetic solvers can propose structures directly, but they require careful treatment of manufacturability, symmetry, topology, and physical constraints. A design that matches a target spectrum but violates minimum feature size is not a valid engineering solution.

6.2 Antenna and array design

AI can optimize element dimensions, spacing, phase excitation, amplitude taper, and matching networks. Array objectives may include gain, beam steering range, sidelobe level, cross-polarization, and mutual coupling. The array factor for identical elements is commonly expressed as

where N is the number of elements, a_n is the amplitude weight, k = 2π/λ is the free-space wavenumber, d is element spacing, β is progressive phase, and θ is observation angle under a specified array-axis convention. The expression assumes identical element patterns and ideal excitation; mutual coupling and embedded element patterns require a more complete model.

6.3 RF diagnosis and calibration

Machine learning can classify faults from S-parameters, near-field scans, thermal images, or production-test data. Applications include connector faults, solder defects, oscillator drift, cable degradation, and antenna detuning. Classification performance must be evaluated using confusion matrices, precision, recall, and calibration—not accuracy alone, particularly when faults are rare.

AI-assisted calibration can learn corrections between nominal circuit models and measured responses. Nevertheless, calibration standards, traceability, and uncertainty budgets remain essential in regulated or safety-critical measurements.

6.4 Signal analytics and spectrum monitoring

Deep learning can detect modulation types, interference, anomalies, and occupancy patterns in complex I/Q data. Training data must represent channel conditions, frequency offsets, noise, fading, hardware impairments, and adversarial or unanticipated signals. A classifier that performs well on laboratory data may fail in a different geographic or electromagnetic environment.

7. Physics-Informed and Hybrid Modelling

Purely data-driven models may violate reciprocity, passivity, causality, or conservation laws. Physics-informed learning incorporates equations, boundary conditions, or known constraints into training. A composite loss may be written conceptually as

where each term penalizes a different type of error and the λ coefficients control relative importance. The exact construction depends on the physical problem; it is not sufficient merely to label a neural network “physics-informed.”

Hybrid models combine a fast analytical or circuit model with learned residuals:

where r_θ is a learned correction. This can improve data efficiency if the baseline model captures dominant physics, but systematic baseline errors can also be inherited by the correction model.

8. Worked Example: Surrogate-Assisted Antenna Tuning

Consider a microstrip patch antenna whose length L, width W, and feed offset x_f are varied. A full-wave model is used to predict resonant frequency f_r and broadside realized gain G. The design target is f_r = 5.80 GHz, with gain at least 7 dBi and a restricted geometry range.

Suppose four measured or simulated candidate results are:

Candidate D meets frequency but violates gain. Candidate B has a frequency error of 0.02 GHz and satisfies gain. Candidate A has an error of 0.15 GHz, while C has an error of 0.11 GHz. If the objective is only absolute frequency error among candidates satisfying the gain constraint, B is currently best.

A surrogate is trained using geometry variables and predicts both outputs. The optimizer proposes a new geometry near the predicted optimum. That candidate must be evaluated with the full-wave solver rather than accepted solely on the surrogate prediction. If the full-wave result is f_r = 5.803 GHz and G = 7.18 dBi, it satisfies the stated targets, subject to mesh convergence and fabrication checks.

The example illustrates three principles: constraints must be enforced explicitly; a prediction is not a validation result; and the best computational design is not necessarily the best manufacturable design.

9. Illustrative Case Study: AI-Assisted Millimetre-Wave Filter Design

Background

A design team must develop a compact bandpass filter for a 28 GHz transceiver. The filter must provide low insertion loss in the passband, adequate rejection outside the band, and dimensions compatible with printed-circuit fabrication.

Problem

A conventional parameter sweep over resonator lengths, coupling gaps, substrate thickness, and dielectric constant requires thousands of full-wave simulations. Each simulation includes a fine mesh near narrow gaps and is computationally expensive.

Analysis

The team identifies seven geometry and material variables. A space-filling design generates an initial database of 600 simulations. Failed meshes are investigated rather than silently removed. Responses include centre frequency, 3 dB bandwidth, passband insertion loss, and rejection at specified offset frequencies.

Approach and solution

A multi-output surrogate is trained after scaling the variables and outputs. Data are divided by geometry family so that test structures are genuinely unseen. Cross-validation reveals that the model predicts centre frequency accurately but has larger errors near sharp transmission zeros. Adaptive sampling therefore adds simulations around these regions. A multi-objective optimizer searches for low insertion loss and high rejection while enforcing minimum gap and maximum footprint constraints.

The final five candidates are re-simulated with a higher-fidelity mesh and then fabricated. Measurements include connector de-embedding and calibration verification. The measured centre frequency differs from simulation because the manufactured substrate thickness and dielectric constant differ from nominal values. A second model trained with measured material distributions improves prediction of subsequent iterations.

Outcome

The computational workflow reduces unnecessary simulations and identifies a feasible design more efficiently than an unstructured sweep. It does not eliminate fabrication iteration, because material tolerance and assembly effects remain significant.

Learning points

1. Surrogate accuracy can vary strongly across the response surface.
2. Sharp resonances and transmission zeros require targeted validation.
3. Manufacturing variables should enter the design model early.
4. High-fidelity re-simulation and measurement are mandatory before release.
5. AI is most valuable when integrated with electromagnetic expertise and a controlled verification process.

10. Limitations, Governance, and Professional Practice

Model and data limitations

Common failure modes include insufficient sampling, extrapolation, data leakage, inappropriate normalization, overfitting, solver artefacts, and ignoring failed simulations. A low test error does not guarantee performance on a new topology, frequency band, material system, or manufacturing process.

Explainability and reproducibility

Engineers should retain the dataset, feature definitions, software versions, solver settings, mesh criteria, random seeds, trained-model parameters, and validation reports. Explainability tools such as permutation importance or sensitivity analysis can reveal dominant variables, but correlation does not establish causation.

Security and intellectual property

Microwave design datasets may contain proprietary geometries, defence-related information, or customer data. Access control, encryption, model-sharing policies, and supply-chain assessment are necessary when cloud tools are used. AI-generated designs must also be checked for patent conflicts and licensing restrictions.

Safety and regulatory considerations

AI cannot override electromagnetic compatibility, spectrum, exposure, reliability, export-control, or product-safety requirements. In high-consequence applications, human approval, traceable evidence, conservative margins, and independent verification are essential.

11. Contemporary Directions and Best Practices

Current research includes differentiable full-wave solvers, neural operators, reduced-order models, active learning, digital twins, uncertainty-aware optimization, and foundation models for scientific data. These methods are promising but remain sensitive to domain shift and physical constraints.

A robust professional workflow is:

1. Define requirements, variables, constraints, and acceptance tests.
2. Build the simplest physically credible model.
3. Verify solver implementation and establish mesh convergence.
4. Generate a documented, representative design-of-experiments dataset.
5. Train and compare several surrogate models using grouped validation.
6. Quantify prediction uncertainty and identify extrapolation.
7. Optimize with manufacturing and regulatory constraints included.
8. Re-evaluate selected designs using high-fidelity simulation.
9. Fabricate, calibrate, measure, and compare with predictions.
10. Update the model using trustworthy experimental evidence.

Chapter Summary

Computational design in microwave engineering integrates electromagnetic modelling, numerical simulation, optimization, analytics, surrogate modelling, and AI. Maxwell’s equations provide the physical foundation, while numerical methods transform field problems into solvable algebraic systems. Verification, validation, convergence analysis, and uncertainty assessment are necessary for credible results.

Optimization formalizes requirements through objectives, constraints, and bounds. Multi-objective methods expose trade-offs through Pareto fronts. Surrogates and active learning reduce the cost of repeated simulation, but their reliability depends on representative data and disciplined validation. AI supports inverse design, antenna and filter optimization, fault diagnosis, calibration, and signal analytics. Physics-informed and hybrid models can improve data efficiency and physical consistency.

The essential professional principle is that AI is an engineering aid, not a substitute for electromagnetic theory, measurement, manufacturing knowledge, or accountable technical judgement.

Key Terms

Computational electromagnetics: Numerical solution of electromagnetic field problems.

FDTD: A time-domain method based on finite spatial and temporal differences.

FEM: A numerical method using local basis functions over a finite-element mesh.

Method of moments: An integral-equation technique that expands unknown currents or fields in basis functions.

Mesh convergence: Stabilization of a computed result as discretization is refined.

Verification: Demonstration that equations and numerical implementation are solved correctly.

Validation: Demonstration that the model represents the physical system adequately.

Design variable: A controllable parameter defining a design.

Objective function: A quantity minimized or maximized during optimization.

Constraint: A condition that feasible designs must satisfy.

Pareto front: Set of non-dominated solutions in a multi-objective problem.

Surrogate model: An inexpensive approximation of a costly simulation or experiment.

Gaussian process: A probabilistic model that provides predictions and uncertainty estimates.

Active learning: Sequential selection of informative new data points.

Inverse design: Determination of structure or parameters from desired responses.

Physics-informed learning: Machine learning constrained by physical equations or principles.

Domain shift: Change between the training environment and the deployment environment.

Data leakage: Unintended use of test or future information during model training.

Digital twin: A continuously updated computational representation of a physical asset.

Uncertainty quantification: Analysis of variability and lack of knowledge in predictions.

Review Questions

Short Answer Questions

1. Distinguish verification from validation in computational electromagnetics.
2. What is the purpose of a surrogate model in microwave design?
3. Define a Pareto-optimal solution.
4. Why should a surrogate prediction be re-evaluated using a high-fidelity solver?
5. State two risks associated with applying AI to microwave engineering.

Long Answer Questions

1. Compare FDTD, FEM, and the method of moments with respect to geometry, frequency range, computational cost, and typical microwave applications.
2. Explain how a constrained multi-objective antenna optimization problem can be formulated and solved.
3. Discuss the importance of mesh convergence, material modelling, and measurement validation.
4. Describe a complete data-generation and validation strategy for a surrogate model.
5. Explain how physics-informed or hybrid models can improve upon purely data-driven models.

Analytical / Application Questions

1. Design a computational workflow for optimizing a compact 28 GHz bandpass filter while accounting for manufacturing tolerances.
2. A neural surrogate has low average RMSE but large errors near transmission zeros. Analyse possible causes and propose a validation and adaptive-sampling strategy.
3. Develop a decision framework for determining whether AI should be used for antenna inverse design, RF fault diagnosis, or spectrum classification in a safety-critical product.

Multiple Choice Questions with Answers

Question 1

Which activity primarily establishes whether a numerical solver has correctly solved the specified mathematical problem?

A. Validation

B. Verification

C. Fabrication

D. Market analysis

Correct Answer: B. Verification

Question 2

In a constrained microwave optimization problem, which condition defines a feasible design?

A. It has the largest number of variables

B. It satisfies all specified constraints and bounds

C. It is always the globally optimal design

D. It requires no experimental validation

Correct Answer: B. It satisfies all specified constraints and bounds

Question 3

Which metric has the same units as the predicted microwave response?

A. R²

B. Normalized sensitivity

C. RMSE

D. Correlation coefficient

Correct Answer: C. RMSE

Question 4

What is the principal advantage of Bayesian optimization for expensive electromagnetic simulations?

A. It eliminates the need for physical constraints

B. It selects new evaluations using a predictive surrogate and acquisition rule

C. It guarantees a global optimum in one iteration

D. It requires no initial data

Correct Answer: B. It selects new evaluations using a predictive surrogate and acquisition rule

Question 5

A design is Pareto-optimal when:

A. Every objective is individually maximal

B. It is dominated by at least one feasible design

C. No feasible design is better in all objectives and strictly better in one

D. All objectives have identical numerical values

Correct Answer: C. No feasible design is better in all objectives and strictly better in one

Question 6

For the phasor convention e^jωt, which complex-permittivity expression represents conductivity in the stated convention?

A. ε_c = ε + jσ/ω

B. ε_c = ε − jσ/ω

C. ε_c = σ − jωε

D. ε_c = εω/σ

Correct Answer: B. ε_c = ε − jσ/ω

Question 7

Why can a random train-test split be misleading for related microwave geometries?

A. It always increases mesh density

B. Nearly duplicate geometry families may appear in both sets

C. It prevents normalization

D. It removes all measurement noise

Correct Answer: B. Nearly duplicate geometry families may appear in both sets

Question 8

In the uncertainty approximation Var(y) ≈ Σᵢ(∂y/∂x_i)²Var(x_i), which assumption is important?

A. The response is approximately linear near the nominal point

B. All variables must be dimensionless

C. The response must be exactly constant

D. The solver must use FDTD

Correct Answer: A. The response is approximately linear near the nominal point

Question 9

What is a major limitation of purely data-driven inverse design?

A. It cannot represent any nonlinear response

B. The inverse mapping may be non-unique and may produce impractical geometries

C. It always requires analytical closed-form solutions

D. It cannot use measured data

Correct Answer: B. The inverse mapping may be non-unique and may produce impractical geometries

Question 10

Which action is most appropriate before releasing an AI-designed microwave component?

A. Accept the surrogate output without further checks

B. Remove all constraints to increase design freedom

C. Perform high-fidelity simulation, fabrication checks, and measurement validation

D. Use only training-set performance

Correct Answer: C. Perform high-fidelity simulation, fabrication checks, and measurement validation