12. Engineering Optimization and Inverse Problems

Introduction and Context

Engineering analysis traditionally asks: given a design, material, load, and operating condition, what will happen? Optimization reverses the direction of reasoning. It asks: among all feasible designs, which one best satisfies specified objectives? Inverse problems go a step further by asking: given observations of a system, what unknown parameters, inputs, geometry, or internal states could have produced them?

These questions are central to modern mechanical engineering. Aircraft structures must be light but safe, thermal systems must transfer heat efficiently while consuming little energy, and manufacturing processes must satisfy quality constraints at minimum cost. In many applications, the governing equations are available, but the design variables are numerous and the computational expense is high. In other applications, measurements are available but the underlying material properties, boundary conditions, defects, or source terms are unknown.

Optimization and inverse analysis provide systematic methods for making such decisions. They combine mathematical modelling, numerical simulation, sensitivity analysis, statistics, experimental design, and engineering judgment. Because a mathematically optimal solution may be impractical, fragile, expensive, or unsafe, engineering optimization is not merely a search for a minimum numerical value. It is a structured process for balancing performance, uncertainty, manufacturability, reliability, sustainability, and regulatory requirements.

Learning Objectives

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

1. Formulate engineering design problems using objective functions, design variables, constraints, and governing equations.

2. Distinguish continuous, discrete, single-objective, multi-objective, deterministic, and robust optimization problems.

3. Explain the operation and appropriate use of gradient-based, adjoint, heuristic, and topology-optimization methods.

4. Interpret sensitivities, Karush–Kuhn–Tucker conditions, Pareto fronts, and regularization in engineering contexts.

5. Formulate parameter-estimation and inverse-modeling problems from experimental or field data.

6. Identify ill-posedness, non-uniqueness, noise amplification, model discrepancy, and computational limitations.

7. Select suitable optimization and inverse methods for multidisciplinary engineering decisions.

8. Evaluate ethical, professional, and practical issues associated with automated engineering decisions.

From Engineering Requirements to Mathematical Formulation

A constrained design problem can be written in the general form

\[

\min_{x} J(x,p)

\]

subject to

\[

g_i(x,p)\leq 0,\qquad i=1,\ldots,m,

\]

\[

h_j(x,p)=0,\qquad j=1,\ldots,r,

\]

and

\[

x_L\leq x\leq x_U.

\]

Here, \(x\) is the vector of design variables, \(p\) represents fixed parameters or operating conditions, and \(J\) is the objective function. Inequality constraints may represent maximum stress, displacement, temperature, noise, emissions, or cost. Equality constraints may express mass balance, geometric relationships, or exact physical requirements.

The design variables may be geometric dimensions, material choices, control settings, process parameters, or fields such as density in topology optimization. A well-posed formulation begins with a clear distinction between quantities that can be changed and quantities that are fixed. Confusing a design variable with an uncertain parameter often leads to inappropriate optimization results.

For example, the design of a pressure vessel may use wall thickness and diameter as variables. The objective could be material cost, while constraints limit von Mises stress, deformation, fabrication thickness, and internal volume. If the pressure is uncertain, optimizing only for its nominal value may produce a design that is inexpensive but unsafe under realistic operating conditions. Robust or reliability-based formulations are more appropriate in that case.

Important terminology

A feasible design satisfies every constraint. The feasible region is the set of all such designs. An active constraint is satisfied as an equality at a solution; an inactive constraint has remaining margin. A local optimum is superior to nearby feasible designs, whereas a global optimum is superior to every feasible design. In nonconvex engineering problems, a numerical algorithm may find a local optimum without proving global optimality.

The objective function represents the criterion being minimized or maximized. Multiple objectives, such as mass and stiffness, generally cannot be reduced to a single answer without expressing preferences. A weighted-sum formulation,

\[

J=w_1J_1+w_2J_2+\cdots+w_kJ_k,

\]

is useful when weights have a defensible interpretation. However, weights can hide trade-offs and may fail to identify some non-convex portions of the trade-off surface. A Pareto-optimal design is one for which no objective can be improved without worsening at least one other objective. The collection of such designs is the Pareto front.

Optimality Conditions and Sensitivity

For an unconstrained differentiable problem, a stationary point satisfies \(\nabla J(x)=0\). This condition is necessary, not sufficient, because the point may be a maximum or saddle point. The Hessian, or second-derivative matrix, provides local curvature information. Positive definiteness of the Hessian at a stationary point establishes a strict local minimum under suitable assumptions.

Constraints modify the optimality conditions. The Lagrangian is defined as

\[

\mathcal{L}(x,\lambda,\mu)=J(x)+\sum_i\lambda_i g_i(x)+\sum_j\mu_j h_j(x),

\]

where \(\lambda_i\) and \(\mu_j\) are Lagrange multipliers. The Karush–Kuhn–Tucker conditions include stationarity, primal feasibility, dual feasibility \((\lambda_i\geq0)\), and complementary slackness \(\lambda_i g_i(x)=0\). These conditions help interpret a solution: a large multiplier indicates that relaxing the associated constraint could significantly improve the objective.

Sensitivity analysis examines how outputs change when inputs or design variables change. A finite-difference estimate of a derivative is simple but may be inaccurate because of truncation error, round-off error, or solver noise. Complex-step differentiation can produce highly accurate derivatives for compatible computer codes. Automatic differentiation applies the chain rule systematically to computational programs and is increasingly used in simulation-based optimization.

The distinction between a derivative and a sensitivity is important. A derivative may describe the response of an output to a design variable while all other quantities are fixed. A total sensitivity may include the effect of state variables that themselves depend on the design. In a finite-element structural problem, the displacement field changes when thickness changes; ignoring this dependence produces an incomplete gradient.

Gradient-Based Optimization for Simulation Models

Gradient-based methods are effective when variables are continuous, the objective and constraints are reasonably smooth, and reliable derivatives are available. Steepest descent moves in the negative gradient direction, while Newton and quasi-Newton methods use curvature information to accelerate convergence. Sequential quadratic programming solves a sequence of local quadratic approximations and is particularly useful for nonlinear constrained problems. Interior-point methods maintain iterates within the feasible region by introducing barrier terms and are effective for large sparse problems.

A typical simulation-based workflow is:

1. define variables, objectives, constraints, and bounds;

2. generate a feasible initial design;

3. solve the governing state equations;

4. evaluate objectives and constraint residuals;

5. compute sensitivities;

6. determine a search direction and step length;

7. update the design and repeat;

8. verify convergence and independently check the final design.

The computational cost may be dominated by repeated state-equation solutions. For a fluid, thermal, or structural model, each optimization iteration may require one or more large numerical solves. Scaling variables and objectives is therefore important. Poor scaling causes the optimizer to take inefficient steps or to interpret numerical noise as meaningful information.

The adjoint method

When the number of design variables is very large and the number of objectives is small, the adjoint method is often preferable to computing one sensitivity solve per design variable. Suppose the state equation is

\[

R(u,x)=0,

\]

where \(u\) is the state and \(x\) is the design. For an objective \(J(u,x)\), differentiation gives

\[

\frac{dJ}{dx}=\frac{\partial J}{\partial x}+\frac{\partial J}{\partial u}\frac{du}{dx}.

\]

Differentiating the residual produces a state-sensitivity equation. The adjoint variable \(\psi\) is selected to satisfy

\[

\left(\frac{\partial R}{\partial u}\right)^T\psi=\left(\frac{\partial J}{\partial u}\right)^T.

\]

The total gradient can then be calculated without explicitly forming \(du/dx\) for every design variable. One adjoint solve is generally associated with each scalar objective or constraint, rather than each design variable. This makes adjoints valuable in aerodynamic shape optimization, computational fluid dynamics, heat transfer, and large-scale structural design.

Adjoint gradients are not automatically correct. Errors may arise from inconsistent discretization, incomplete convergence of the state solver, turbulence-model approximations, discontinuities, or incorrect treatment of boundary conditions. Gradient verification using directional finite differences is an essential professional practice.

Heuristic and Population-Based Methods

Gradient-free methods are useful when the objective is discontinuous, noisy, non-differentiable, simulation-based, mixed-integer, or characterized by many local optima. Genetic algorithms use selection, crossover, and mutation. Particle swarm optimization updates candidate solutions using individual and collective experience. Differential evolution creates trial vectors from differences among population members. Simulated annealing accepts some inferior moves early in the search to reduce the probability of premature convergence.

These methods can explore complex design spaces and do not require derivative information. Their disadvantages include large numbers of objective evaluations, sensitivity to algorithmic parameters, stochastic variability, and difficulty in certifying convergence. A heuristic result should therefore be treated as a strong candidate rather than unquestionable proof of global optimality.

Hybrid methods are often effective. A population-based algorithm can explore the design space, after which a gradient-based method refines a promising feasible candidate. Surrogate models, reduced-order models, or response surfaces can reduce the cost of expensive evaluations. However, surrogate predictions must be validated near the selected design, especially when constraints are safety-critical.

Topology Optimization and Structural Concept Generation

Shape optimization modifies the boundary of an existing component, whereas topology optimization determines where material should exist within a prescribed design domain. In density-based methods, each finite element is assigned a relative density \(\rho_e\) between a small lower bound and one. A common formulation minimizes compliance subject to a volume constraint:

\[

\min_{\rho} C(\rho)=F^Tu(\rho)

\]

subject to

\[

K(\rho)u=F,\qquad \sum_e \rho_eV_e\leq V^*,\qquad 0<\rho_{min}\leq\rho_e\leq1.

\]

The SIMP approach interpolates stiffness using \(E_e=\rho_e^pE_0\), where the penalization exponent \(p\) discourages intermediate densities. Filtering and projection techniques control checkerboard patterns, mesh dependence, and excessive fine-scale features.

Topology optimization is powerful for generating lightweight load paths, but a raw mathematical solution is rarely a finished component. It must be interpreted in relation to manufacturing processes, inspection access, fatigue, joints, thermal gradients, and minimum feature sizes. Additive manufacturing may enable geometries that conventional machining cannot produce, but it introduces overhang, residual-stress, support-removal, and qualification issues. A practical workflow therefore includes manufacturing constraints from the beginning, followed by geometric reconstruction, detailed analysis, and physical testing.

Parameter Estimation and Inverse Problems

In a forward problem, parameters and inputs are known and the model predicts observations. In an inverse problem, observations are used to infer unknown quantities. Let

\[

y=G(\theta)+\varepsilon,

\]

where \(\theta\) is the unknown parameter vector, \(G\) is the forward model, and \(\varepsilon\) represents measurement noise and model discrepancy. A least-squares estimate solves

\[

\hat{\theta}=\arg\min_{\theta}\|W^{1/2}(G(\theta)-y)\|_2^2,

\]

where \(W\) weights measurements according to their uncertainty or importance.

Examples include identifying elastic modulus and damping from vibration data, estimating thermal conductivity from temperature histories, determining a heat-source distribution from surface measurements, and reconstructing defects from ultrasonic or thermographic observations. The parameters may be scalar, spatially distributed, time-dependent, or categorical.

Inverse problems are often ill-posed. A solution may not exist, may not be unique, or may not depend continuously on the data. Small measurement errors can produce large changes in the inferred parameters, especially when the forward operator smooths information, as in heat conduction. Regularization stabilizes the estimation by adding prior information:

\[

\hat{\theta}=\arg\min_{\theta}\left[\|W^{1/2}(G(\theta)-y)\|_2^2+\alpha R(\theta)\right].

\]

The regularization functional \(R\) may penalize large parameter magnitudes, roughness, or deviations from a reference model. Tikhonov regularization uses commonly quadratic penalties; sparsity-promoting penalties are useful when only a few sources or defects are expected. The parameter \(\alpha\) controls the balance between data fit and prior preference and may be selected using validation, discrepancy principles, or evidence-based methods.

Bayesian inversion expresses uncertainty explicitly. With prior density \(p(\theta)\) and likelihood \(p(y|\theta)\), the posterior is

\[

p(\theta|y)\propto p(y|\theta)p(\theta).

\]

The posterior provides not only an estimate but also credible intervals and correlations. This is especially important when different parameters have similar effects on the measurements. Identifiability analysis asks whether the available data can distinguish among competing parameter values. Adding more measurements is not always sufficient; sensor location, excitation type, frequency range, and measurement quality determine the information content of an experiment.

An Illustrative Case: Thermal Design and Property Identification

Background

Consider an illustrative industrial case involving an actively cooled metallic panel used near a high-temperature process. The panel must remain below a specified temperature while using as little coolant as possible. Material conductivity is uncertain because it changes with temperature and manufacturing history.

Situation and problem

The initial design uses uniform channel spacing and a nominal conductivity obtained from a handbook. A forward finite-element model predicts acceptable peak temperature, but prototype measurements show localized hot regions. The engineering team must decide whether to redesign the channel layout, revise the material model, or both.

Analysis

The design variables are channel widths, spacing, and local wall thickness. The objective combines coolant pumping power, material mass, and temperature non-uniformity. Constraints impose a maximum temperature, minimum manufacturable wall thickness, pressure limits, and geometric accessibility. Thermocouple measurements and infrared images are used to estimate temperature-dependent conductivity and an uncertain contact-resistance parameter.

A regularized least-squares inverse problem is first solved using data from several controlled heat-load tests. Smoothness is imposed on the conductivity function, while physically plausible bounds are enforced. The calibrated model is then used in a constrained design optimization. An adjoint calculation supplies gradients of the peak-temperature constraint and pumping-power objective with respect to many channel parameters.

Approach and solution

The team uses a two-stage strategy. A broad design exploration with a low-order surrogate identifies promising channel patterns. A gradient-based constrained optimization then refines the best candidates using the calibrated high-fidelity model. Robust constraints are evaluated for a range of conductivity and heat-load values rather than only at nominal conditions. The final geometry is simplified to satisfy machining limits and subjected to a separate mesh-converged analysis.

Outcome

The important outcome is not simply a lower predicted temperature. The combined procedure reveals that the original hot spots were caused primarily by an underestimated contact resistance, not by a uniformly low material conductivity. Calibrating the model prevents an unnecessary and costly change in material specification. The redesigned channels improve thermal uniformity while maintaining pressure and manufacturing limits. Before production release, the design requires physical validation because the inverse estimate and numerical model remain uncertain.

Learning points

The case illustrates four principles. First, optimization cannot compensate for a systematically incorrect model. Second, inverse estimation should include regularization and uncertainty assessment. Third, gradients are especially valuable when many geometric variables influence a small number of objectives. Fourth, a numerically optimal design must be translated into a manufacturable and testable engineering artifact.

Challenges, Risks, and Common Errors

A frequent mistake is optimizing an objective that is easy to compute but poorly represents the engineering purpose. For example, minimizing mass without fatigue, buckling, or manufacturing constraints can create an unsafe design. Another mistake is treating a constraint violation caused by numerical tolerance as insignificant when the constraint represents a safety limit.

Model-form uncertainty is distinct from measurement noise. More accurate measurements cannot remove errors caused by missing physics, inappropriate constitutive laws, or incorrect boundary conditions. Calibration data should be separated from validation data; using the same data for both can create unjustified confidence.

Optimization may also suffer from premature convergence, dependence on the initial point, hidden infeasibility, poor variable scaling, mesh dependence, and noisy gradients. Discrete choices such as material grade, fastener type, or machine availability should not be forced into a continuous formulation without verification. In multi-objective problems, presenting one weighted solution without showing the trade-off information may conceal important alternatives.

Uncertainty propagation should be considered when the solution is sensitive to loads, material properties, tolerances, or environmental conditions. Robust optimization seeks good performance over a range of conditions, while reliability-based design constrains the probability of failure. These approaches may require many model evaluations and therefore benefit from surrogate modelling, reduced-order modelling, or adaptive sampling.

Professional, Ethical, and Governance Considerations

Engineering decisions remain the responsibility of qualified professionals even when software produces an apparently optimal answer. An optimization report should record model assumptions, units, solver settings, convergence measures, constraint margins, derivative checks, mesh studies, and independent verification. Reproducibility requires version-controlled models, documented data preprocessing, and traceable design changes.

Safety constraints should not be weakened merely to obtain a lower cost or a more attractive performance metric. Measurement data used for inverse modelling must be obtained and stored responsibly, particularly when equipment condition, proprietary processes, or operational security is involved. Automated or machine-learning-assisted optimization should be scrutinized for extrapolation beyond its training data and for hidden bias introduced by incomplete datasets.

Where regulations or professional standards specify allowable stresses, reliability levels, testing procedures, or traceability requirements, these obligations take precedence over unconstrained mathematical objectives. Human review is especially important when optimization changes load paths, introduces unfamiliar materials, or produces geometries that are difficult to inspect.

Digital Implementation and Emerging Directions

Modern optimization platforms increasingly connect computer-aided design, finite-element analysis, computational fluid dynamics, laboratory data, and manufacturing information. Differentiable programming and automatic differentiation make it possible to obtain gradients through parts of a simulation workflow. Physics-informed machine learning can provide fast approximations or incorporate governing equations into data-driven models, although it does not eliminate the need for validation.

Digital twins combine measurement streams with updating models to monitor condition and support decision-making. Inverse methods are used for online estimation of degradation, boundary conditions, and latent states. Active learning and optimal experimental design select the next measurement or simulation that is expected to reduce uncertainty most effectively. Multifidelity optimization combines inexpensive approximate models with occasional high-fidelity evaluations.

Future engineering workflows are likely to be more integrated rather than divided into separate analysis, optimization, and testing stages. Nevertheless, increased automation makes verification, uncertainty quantification, interpretability, and governance more important. Speed without trustworthy evidence is not a substitute for engineering quality.

Best-Practice Workflow

A defensible optimization or inverse-analysis study should:

• define the engineering decision and success criteria before selecting an algorithm;
• distinguish design variables, uncertain parameters, measured quantities, and fixed inputs;
• formulate all safety, physical, regulatory, and manufacturing constraints explicitly;
• verify the forward model through analytical checks, limiting cases, mesh studies, or experiments;
• scale variables and objectives and test sensitivities independently;
• use global exploration, local refinement, or hybrid methods according to problem structure;
• quantify uncertainty and test the selected design outside the calibration data;
• inspect active constraints, robustness, manufacturability, and lifecycle implications;
• preserve an auditable record of data, assumptions, algorithms, and decisions.

Chapter Summary

Engineering optimization determines the best feasible design according to stated objectives, while inverse problems infer unknown parameters, inputs, or states from observations. Both depend on a credible forward model and a careful mathematical formulation. Gradient-based methods are efficient for smooth large-scale problems, and adjoint methods make gradients practical when there are many design variables but few objectives. Heuristic methods are useful for discontinuous, noisy, mixed, or highly nonconvex problems, often in combination with local refinement.

Topology optimization generates material layouts but must be followed by manufacturing interpretation and verification. Inverse problems are frequently ill-posed and require regularization, uncertainty analysis, and attention to identifiability. Multi-objective, robust, and reliability-based methods help represent realistic engineering trade-offs. Professional practice requires independent validation, transparent assumptions, safety governance, and recognition that numerical optimality does not by itself establish engineering acceptability.

Key Terms

Objective function: A scalar measure of performance that an optimization seeks to minimize or maximize.

Design variable: A controllable quantity that defines a candidate engineering design.

Constraint: A mathematical or physical condition that a feasible design must satisfy.

Feasible region: The set of all designs satisfying the specified constraints and bounds.

Active constraint: A constraint that is satisfied at equality at a solution.

Sensitivity: The change in a response caused by a change in an input or design variable.

Gradient: The vector of first derivatives of a scalar function with respect to its variables.

Adjoint method: A sensitivity technique that obtains gradients efficiently for many design variables and few objectives.

KKT conditions: Necessary optimality conditions for constrained nonlinear optimization under suitable regularity assumptions.

Pareto front: The set of nondominated solutions representing trade-offs among multiple objectives.

Heuristic optimization: Search based on rules or stochastic population processes rather than exact derivatives.

Topology optimization: Optimization of material distribution within a prescribed design domain.

Inverse problem: Estimation of unknown causes, parameters, or states from observed effects.

Ill-posed problem: A problem lacking existence, uniqueness, or stable dependence on data.

Regularization: Incorporation of stabilizing penalties or prior information into an inverse problem.

Identifiability: The extent to which available data can uniquely distinguish unknown parameters.

Surrogate model: An inexpensive approximation used in place of an expensive simulation.

Robust optimization: Optimization that seeks satisfactory performance under uncertainty.

Reliability-based design: Design that constrains the probability of failure or noncompliance.

Model discrepancy: Error caused by limitations or incorrect assumptions in the mathematical model.

Review Questions

Short Answer Questions

1. Distinguish between a design variable, an uncertain parameter, and a constraint.

2. What is the meaning of an active constraint at an optimum?

3. Why are adjoint methods valuable in large-scale simulation-based optimization?

4. Define regularization in the context of inverse problems.

5. What is meant by a Pareto-optimal design?

Long Answer Questions

1. Formulate a constrained optimization problem for a lightweight pressure vessel and explain the meaning of each term.

2. Explain the KKT conditions and discuss how Lagrange multipliers can support engineering decision-making.

3. Compare gradient-based and heuristic optimization methods with respect to derivatives, computational cost, and problem suitability.

4. Explain density-based topology optimization, including penalization, filtering, and manufacturing considerations.

5. Discuss why inverse heat-conduction problems are ill-posed and how regularization and Bayesian inference address this difficulty.

Analytical / Application Questions

1. A thermal component has ten thousand geometry variables and one temperature objective. Propose a sensitivity strategy and justify whether an adjoint method is appropriate.

2. An identified material property changes drastically when a small amount of measurement noise is added. Diagnose the likely cause and propose two remedies.

3. A topology-optimized bracket has excellent simulated stiffness but cannot be machined or inspected. Develop a revised workflow that addresses the design, manufacturing, and validation issues.

22. Multiple Choice Questions with Answers

Question 1

Which formulation most appropriately represents a constrained engineering optimization problem?

A. Maximize every design variable independently without restrictions

B. Minimize an objective subject to governing equations, inequalities, equalities, and bounds

C. Estimate measurement noise without using a forward model

D. Select the design with the largest number of active constraints

Correct Answer: B. Minimize an objective subject to governing equations, inequalities, equalities, and bounds

Question 2

Why is an adjoint method especially efficient when optimizing a CFD model with millions of design variables and one scalar objective?

A. It eliminates the need to solve the governing equations

B. It requires one independent sensitivity solve for every design variable

C. It can obtain the gradient with a cost mainly associated with the objective rather than the number of variables

D. It guarantees that the optimization has a global solution

Correct Answer: C. It can obtain the gradient with a cost mainly associated with the objective rather than the number of variables

Question 3

At a constrained optimum, a large positive Lagrange multiplier for an inequality constraint generally indicates that:

A. The constraint is irrelevant to the solution

B. Relaxing the constraint could significantly improve the objective

C. The constraint must be removed from the model

D. The objective has no sensitivity to the constraint

Correct Answer: B. Relaxing the constraint could significantly improve the objective

Question 4

Which situation most strongly favors a heuristic optimization method over a conventional gradient-based method?

A. A smooth convex problem with accurate analytical gradients

B. A small quadratic problem with linear constraints

C. A mixed discrete-continuous problem with discontinuous simulation responses

D. A well-scaled unconstrained problem with a known Hessian

Correct Answer: C. A mixed discrete-continuous problem with discontinuous simulation responses

Question 5

In density-based topology optimization, the primary purpose of penalizing intermediate densities is to:

A. Encourage designs that approach solid-or-void material distributions

B. Increase the number of gray elements in the solution

C. Remove the volume constraint

D. Replace the structural equilibrium equations

Correct Answer: A. Encourage designs that approach solid-or-void material distributions

Question 6

What makes an inverse problem ill-posed?

A. It always has exactly one stable solution

B. It may lack existence, uniqueness, or stable dependence on measured data

C. It contains no measurement data

D. It can only be solved using a genetic algorithm

Correct Answer: B. It may lack existence, uniqueness, or stable dependence on measured data

Question 7

A regularization term in a parameter-estimation objective is introduced primarily to:

A. Guarantee that the measurements contain no noise

B. Stabilize the estimate by incorporating prior preferences or smoothness

C. Increase model discrepancy deliberately

D. Make every parameter independently identifiable

Correct Answer: B. Stabilize the estimate by incorporating prior preferences or smoothness

Question 8

A design that cannot improve any one objective without worsening another is called:

A. Infeasible

B. Dominated

C. Pareto-optimal

D. Unconstrained

Correct Answer: C. Pareto-optimal

Question 9

Why should calibration and validation data generally be separated in inverse modelling?

A. To ensure that the model is never compared with measurements

B. To prevent assessment of predictive performance using data already fitted

C. To increase the number of unknown parameters

D. To remove the need for uncertainty analysis

Correct Answer: B. To prevent assessment of predictive performance using data already fitted

Question 10

A topology-optimized component satisfies stress constraints but violates minimum feature-size and inspection requirements. The most appropriate response is to:

A. Accept the result because numerical optimality overrides manufacturing constraints

B. Ignore the inspection issue if the mesh is refined

C. Include manufacturing and inspection constraints, then re-optimize and validate the revised design

D. Replace all constraints with a mass objective

Correct Answer: C. Include manufacturing and inspection constraints, then re-optimize and validate the revised design