11. Process Modeling, Experimentation and Optimization

Introduction and Context

Advanced manufacturing processes are governed by interacting thermal, mechanical, electrical, chemical and metallurgical phenomena. In machining, additive manufacturing, welding, forming, surface treatment and micromanufacturing, a small change in a process variable may alter temperature, stress, material flow, tool wear, surface integrity and dimensional accuracy simultaneously. Consequently, process development based only on trial and error is expensive, slow and often incapable of identifying robust operating conditions.

Process modeling, experimentation and optimization provide a systematic alternative. A model represents the relationship between inputs, physical mechanisms and outputs. Experimentation supplies evidence for calibrating and validating that representation. Optimization uses the resulting knowledge to select process conditions that satisfy technical, economic and environmental objectives. The central aim is not merely to find a single best parameter combination, but to establish a process window that remains acceptable despite material variation, measurement error, machine drift and environmental disturbances.

This chapter treats modeling and experimentation as complementary activities. Analytical models provide insight and rapid estimates; finite-element and multiphysics models describe spatially distributed and strongly coupled phenomena; designed experiments identify influential variables and interactions; response-surface methods support prediction and optimization; uncertainty analysis evaluates confidence and robustness; and multi-objective optimization exposes trade-offs that cannot be represented by one objective alone.

Learning Objectives

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

1. Distinguish analytical, numerical, empirical and hybrid process models.

2. Formulate manufacturing problems using inputs, outputs, assumptions, boundary conditions and governing equations.

3. Explain the role of finite-element and multiphysics simulation in process analysis.

4. Design efficient experiments using screening designs, factorial designs and response-surface methods.

5. Fit, diagnose and interpret regression and response-surface models.

6. Propagate uncertainty and distinguish sensitivity from robustness.

7. Apply single- and multi-objective optimization to manufacturing decisions.

8. Develop and validate a process window through an integrated modeling–experimentation workflow.

9. Recognize numerical, statistical, managerial and ethical risks in process optimization.

11.1 The Architecture of a Manufacturing Process Model

A process model begins by defining the system boundary. The design variables or controllable factors may include cutting speed, feed, depth of cut, laser power, scan speed, layer thickness, forming force, tool geometry, coolant flow or material composition. Noise factors are variables that influence performance but cannot be controlled economically during routine production, such as initial material hardness, ambient temperature, tool-to-tool variation and machine vibration. Responses are measured outcomes: cutting force, temperature, deposition rate, dimensional error, roughness, residual stress, porosity, energy consumption or productivity.

The model must also state its assumptions. A thermal model may assume constant thermal properties, while a mechanical model may assume small deformation or rigid tooling. Assumptions reduce complexity, but they also define the domain in which predictions are credible. A model that predicts accurately under one feed rate, material and cooling condition should not automatically be extrapolated to another regime.

A useful classification is:

• Analytical models, derived from physical laws and simplified geometry, provide interpretable relationships and rapid computation.
• Empirical models, obtained from experimental data, can describe complex behavior but are usually restricted to the tested domain.
• Numerical models, such as finite-element, finite-volume and computational-fluid-dynamics models, solve governing equations for complex geometries and boundary conditions.
• Hybrid or surrogate models combine physical understanding with statistical or machine-learning approximations. A surrogate may replace an expensive simulation during optimization while retaining selected physical constraints.

The quality of a model is not determined solely by prediction accuracy. Interpretability, computational cost, extrapolation behavior, sensitivity to assumptions, ease of calibration and usefulness for decision-making are also important. A highly detailed simulation may be less useful than a simpler model if its material data are uncertain or if it requires several days for one parameter combination.

Conservation laws and constitutive behavior

Most process models are constructed from conservation of mass, momentum and energy, supplemented by constitutive equations. In a thermal process, a typical heat equation is

ρcₚ ∂T/∂t = ∇·(k∇T) + Q,

where ρ is density, cₚ is specific heat, T is temperature, k is thermal conductivity and Q represents internal or applied heat generation. Boundary conditions may include convection, radiation, conduction to tooling and a prescribed heat flux.

Mechanical behavior is commonly expressed through equilibrium equations and a constitutive law relating stress and strain. For plastic deformation, the constitutive description may include elastic response, yield behavior, strain hardening, strain-rate sensitivity and temperature dependence. In welding or additive manufacturing, thermal expansion and phase transformation may be coupled to mechanical stress and distortion.

The equations alone do not define a complete model. Initial conditions, contact laws, friction coefficients, material-property data, geometry, loading history and numerical discretization are equally important. Incorrect or poorly justified boundary conditions can dominate the result even when the governing equations are correct.

11.2 Analytical Models: Insight Before Computation

Analytical models are valuable during early process planning because they reveal scaling laws and dominant variables. For example, a simplified cutting-force model may express force as a function of uncut chip area, material-specific cutting coefficients and tool geometry. Such a model can show why increasing feed generally raises force and power, while also indicating where the simplification fails because of built-up edge, tool wear or dynamic instability.

Analytical thermal models can estimate heat generation and temperature fields using moving heat-source approximations. In additive manufacturing, a dimensionless energy-density indicator may help compare parameter combinations, although it should not be treated as a universal predictor of porosity or microstructure. In forming, slab or upper-bound analyses provide estimates of force and deformation work. In abrasive processes, models may relate material-removal rate to abrasive concentration, pressure, relative velocity and particle characteristics.

The main advantages are speed, physical transparency and suitability for preliminary design. Their limitations arise from idealized geometry, simplified contact, homogeneous material assumptions and incomplete treatment of transient phenomena. An analytical model should therefore be used as a hypothesis-generating tool and as a benchmark for more detailed simulation, not as an unquestioned substitute for measurement.

11.3 Finite-Element and Multiphysics Simulation

The finite-element method divides a continuous domain into elements connected at nodes. The governing differential equations are converted into algebraic equations through a weak or variational formulation. The solution is then obtained for nodal temperatures, displacements, velocities or other field variables. Mesh refinement is required where gradients are high, such as near a tool edge, heat source, crack tip, contact boundary or melt-pool boundary.

Manufacturing simulations may be:

• Lagrangian, in which the mesh follows the material; useful for many solid-deformation problems but vulnerable to severe element distortion.
• Eulerian, in which material flows through a fixed mesh; useful for steady flow and some forming or fluid problems.
• Arbitrary Lagrangian–Eulerian, which allows the mesh to move independently and can reduce distortion.
• Coupled or sequentially coupled, in which thermal, mechanical, fluid, electromagnetic or metallurgical fields are solved together or passed from one analysis to another.

In metal cutting, finite-element models may predict chip formation, temperature, stress, force and tool wear indicators. In welding and additive manufacturing, thermal histories can be transferred to mechanical analyses to estimate residual stress and distortion. Electromagnetic–thermal coupling is relevant to induction heating, while fluid–thermal coupling is essential for molten-pool behavior and cooling flows.

Verification, validation and calibration

Verification asks whether the equations have been solved correctly. It includes checking element formulations, convergence, energy balance, time-step sensitivity and mesh independence. Validation asks whether the model represents the physical process adequately. It requires comparison with independent experimental observations, not merely the data used to calibrate the model.

Calibration estimates unknown parameters such as friction coefficients, heat-partition fractions, convection coefficients or material hardening constants. Calibration must avoid compensating for a structural error in the model. For example, adjusting a heat-transfer coefficient to match one temperature measurement may conceal an incorrect heat-source distribution.

Validation should compare quantities relevant to the intended decision. If the model will be used to control distortion, temperature agreement alone is insufficient; distortion, residual stress or dimensional error should also be assessed. Useful measures include mean absolute error, root-mean-square error, maximum error and prediction intervals. Visual agreement between curves is not a substitute for quantitative evaluation.

11.4 From Physical Questions to Designed Experiments

Simulation cannot eliminate experimentation. Material properties may be uncertain, sensors may be imperfect and manufacturing processes often contain mechanisms omitted from the model. Designed experiments make measurements informative by selecting factor combinations systematically rather than changing one variable at a time.

A sound experimental plan begins with a clear question. Screening asks which factors matter. Characterization seeks a predictive relationship. Optimization searches for desirable settings. Robust design examines how settings perform under noise. These objectives require different designs.

A factorial design studies combinations of factor levels. A full two-level factorial with k factors requires 2ᵏ runs and estimates main effects and interactions, subject to replication and model-resolution considerations. Fractional factorial designs reduce run count but confound selected effects. Randomization protects against time trends and hidden nuisance variables; blocking groups runs under similar conditions, such as the same batch of material or machine shift. Replication estimates pure experimental error and helps distinguish real effects from random variation.

When many potential factors exist, screening designs such as fractional factorial, Plackett–Burman or definitive screening designs can identify influential variables. Screening should not be interpreted as final optimization: a factor with a weak main effect may participate in a strong interaction.

Response-surface methodology

Once the important region is identified, response-surface methodology (RSM) approximates the response locally using a polynomial. A second-order model is commonly written as

ŷ = β₀ + Σβᵢxᵢ + Σβᵢᵢxᵢ² + ΣΣβᵢⱼxᵢxⱼ + ε,

where xᵢ are coded factors, β terms are estimated coefficients and ε represents unexplained variation. Central composite and Box–Behnken designs are widely used to estimate curvature and interactions with fewer runs than broad high-resolution grids.

The fitted model should be examined through residual plots, lack-of-fit tests, coefficient uncertainty, leverage and prediction error. A statistically significant coefficient is not automatically technologically important, and a non-significant coefficient may still matter if the design is small or measurement variation is large. The model is local: predictions outside the experimental region are extrapolations and should be treated cautiously.

Sequential experimentation is often superior to a single large campaign. Initial experiments identify the region of interest, a second design estimates curvature, and confirmation runs test the predicted optimum. If the response surface behaves unexpectedly, the design should be revised rather than forcing an inadequate polynomial to provide an answer.

11.5 Uncertainty, Sensitivity and Robustness

Manufacturing decisions require more than a nominal prediction. Let the response be y = f(x, z), where x denotes controllable variables and z denotes uncertain quantities. Uncertainty may arise from material properties, sensor resolution, model-form error, parameter variation, environmental conditions and run-to-run noise.

A first-order propagation approximation is

Var(y) ≈ gᵀΣg,

where g is the gradient of the response with respect to uncertain inputs and Σ is their covariance matrix. This approximation is useful near a nominal point, but Monte Carlo simulation is preferable for nonlinear models, non-normal distributions or discontinuous responses. In Monte Carlo analysis, uncertain inputs are sampled, the model is evaluated repeatedly and the resulting response distribution is summarized using means, standard deviations, quantiles or probability of failure.

Sensitivity analysis identifies which inputs contribute most to output variation. Local derivatives are efficient but may miss nonlinear interactions. Global methods, including variance-based decompositions, evaluate effects over the entire input distribution. Sensitivity results guide experimentation: variables with high influence deserve better measurement, tighter control or more detailed modeling.

A process is robust when its performance remains acceptable under credible variation. Robust optimization may minimize expected loss, worst-case loss or a weighted combination of mean and variance. Taguchi-style signal-to-noise measures provide one approach, while modern probabilistic optimization directly constrains failure probability. Robustness should be defined against engineering tolerances, not against an arbitrary desire for a small standard deviation.

11.6 Optimization and the Meaning of a Process Window

Single-objective optimization may minimize surface roughness, maximize material-removal rate or minimize energy consumption subject to constraints. A mathematical formulation is

minimize or maximize f(x), subject to gⱼ(x) ≤ 0 and hₗ(x) = 0,

where x is the vector of process variables and g and h represent inequality and equality constraints. Constraints may include machine power, tool temperature, dimensional tolerance, minimum strength, maximum residual stress or safety limits.

Manufacturing decisions usually involve competing objectives. Increasing productivity may increase roughness, tool wear or energy demand. Multi-objective optimization therefore seeks a set of Pareto-optimal solutions: no objective can be improved without worsening at least one other objective. The resulting Pareto front supports engineering judgment rather than concealing trade-offs in a single arbitrary weighting.

Common approaches include weighted-sum methods, goal programming, ε-constraint methods and evolutionary algorithms such as genetic algorithms. Surrogate-assisted optimization is useful when each finite-element simulation is expensive. However, optimization algorithms do not correct a poor model. A mathematically precise optimum based on biased data may be physically unsafe.

A process window is a region of operating conditions that satisfies performance and feasibility requirements. It should include uncertainty margins. For example, instead of reporting a nominal laser power and scan speed, an engineer may identify a region in which porosity remains below a specified limit, dimensional error stays within tolerance and energy use remains acceptable with a stated confidence level. The window is more useful for production than a single optimum because it accommodates normal variation and simplifies process control.

11.7 An Integrated Workflow

A practical workflow links physics, data and decision-making:

1. Define the product requirements, responses, constraints and acceptable risk.

2. Identify controllable factors, noise factors, measurable disturbances and likely mechanisms.

3. Build the simplest model capable of addressing the decision; use analytical estimates to check orders of magnitude.

4. Conduct screening experiments and use them to calibrate or challenge the model.

5. Refine the numerical model, perform verification tests and validate against independent data.

6. Use RSM or a validated surrogate to explore the feasible region efficiently.

7. Quantify uncertainty and conduct sensitivity analysis.

8. Perform constrained single- or multi-objective optimization.

9. Select a robust region rather than relying only on a nominal optimum.

10. Conduct confirmation experiments, document model limitations and transfer the process window to production control.

This workflow is iterative. Disagreement between simulation and experiment is not merely a nuisance; it may reveal an omitted mechanism, faulty measurement, incorrect boundary condition or changing material state.

11.8 Illustrative Case Study: Robust Optimization of Laser Powder-Bed Fusion

Background

The following is an illustrative case concerning a stainless-steel laser powder-bed fusion process. The engineering team seeks high density and acceptable surface quality while limiting build time and residual distortion.

Situation and problem

Initial production trials show that increasing laser power improves lack-of-fusion resistance at low scan speed, but excessive energy produces keyhole pores and increases distortion. The process also varies with powder-layer thickness and hatch spacing. A nominal parameter set is therefore inadequate.

Analysis and approach

The team first defines laser power, scan speed, hatch spacing and layer thickness as controllable factors. Relative density, surface roughness, build time and distortion are selected as responses. An analytical energy-balance estimate identifies implausibly low and high heat-input regions. A fractional factorial screening design then reveals that laser power, scan speed and hatch spacing dominate density, while layer thickness strongly affects roughness and build time. Interactions between power and speed are significant.

A central composite design is performed around the feasible region. A second-order response surface is fitted for density and distortion, while a separate model predicts roughness. Residual analysis identifies a curvature pattern that is not adequately described at the original low-energy boundary, so additional runs are added near that boundary rather than extrapolating.

A transient thermal finite-element model is calibrated using melt-pool width and cooling-rate measurements. The model is not used to predict every pore, because its homogenized representation cannot resolve individual powder defects. It is used to estimate thermal gradients and distortion trends. Independent confirmation builds are reserved for validation.

Powder absorptivity, thermal conductivity and layer-to-layer heat-transfer conditions are treated as uncertain. Monte Carlo analysis produces a distribution of predicted density and distortion for each candidate parameter set. The optimization problem minimizes build time and distortion while maximizing density, subject to a density requirement and a roughness limit. The resulting Pareto front contains a fast but less robust region and a slightly slower region with substantially lower probability of exceeding the distortion limit.

Outcome

The team selects a point inside the robust region rather than the mathematically fastest point. Confirmation builds at nominal settings and at controlled perturbations of power and speed meet the specified density and distortion requirements. The final production window is expressed as an allowable region with monitoring rules for powder condition, melt-pool signal and machine calibration.

Learning points

The case demonstrates that physical simulation, designed experimentation and statistical modeling answer different questions. The response surface describes measured behavior locally; the thermal model explains mechanisms and supports extrapolation cautiously; uncertainty analysis distinguishes a high-performing setting from a reliable one. Most importantly, the deliverable is a validated process window, not a single parameter combination.

11.9 Challenges, Risks and Professional Judgment

Several errors recur in process modeling and optimization. Under-designed experiments produce aliased effects and unstable conclusions. Excessive reliance on p-values can obscure practical significance. Repeated measurements are sometimes mistaken for independent replications when they share the same specimen or machine state. Simulation may appear authoritative despite uncertain material data or unverified contact conditions. Mesh refinement can improve numerical precision without improving physical accuracy.

Data leakage is a serious risk when the same experiments are used for calibration, model selection and final validation. A genuine validation set should represent the intended operating domain and remain unused during fitting. Optimization can also exploit numerical artifacts, producing parameter combinations that are mathematically attractive but physically infeasible. Constraints, manufacturability checks and confirmation experiments are essential safeguards.

Professional responsibility requires traceability of data, version control of models, disclosure of assumptions and clear communication of uncertainty. Safety-related decisions should not be based solely on an unvalidated surrogate. If optimization recommends operating near a thermal, mechanical or electrical limit, the limit must be supported by engineering standards and hazard analysis. Experimental records should preserve sensor calibration, material batch information, environmental conditions and failed trials, not only successful runs.

11.10 Digital Manufacturing and Emerging Directions

Digital manufacturing increasingly links sensors, simulation and process control. Reduced-order models and machine-learning surrogates can approximate expensive simulations in real time. Digital twins combine a process model with live data, enabling state estimation, anomaly detection and adaptive control. Bayesian calibration provides a principled way to update uncertain model parameters as new observations become available. Active learning selects the next experiment where it is expected to reduce uncertainty or improve the decision most efficiently.

Physics-informed machine learning can impose conservation laws or constitutive constraints, reducing the risk of implausible predictions. Hybrid models are particularly promising where fundamental mechanisms are known but difficult-to-measure parameters remain uncertain. Nevertheless, explainability, domain shift, cybersecurity, data provenance and model drift must be addressed before such systems control safety-critical manufacturing operations.

Best Practices for Establishing Robust Process Windows

• Define responses, constraints and acceptable risk before selecting an optimization algorithm.
• Use physics to choose meaningful factors, ranges and measurements.
• Screen broadly, then characterize locally with a response-surface design.
• Randomize, block and replicate experiments appropriately.
• Separate calibration, model selection and validation data.
• Check residuals, lack of fit, mesh convergence and energy or mass balance.
• Quantify both parameter uncertainty and model-form uncertainty where possible.
• Report prediction intervals and probability of failure, not only nominal optima.
• Treat Pareto solutions as engineering choices requiring explicit priorities.
• Confirm the selected region under realistic noise and production conditions.
• Document assumptions, limitations, data lineage and change-control procedures.

Chapter Summary

Process modeling, experimentation and optimization form an integrated framework for advanced manufacturing decisions. Analytical models provide rapid physical insight, while finite-element and multiphysics models resolve coupled fields and complex geometries. Designed experiments identify influential factors, interactions and curvature efficiently; response-surface methods convert these observations into local predictive models. Verification concerns numerical correctness, whereas validation concerns agreement with physical reality.

Because manufacturing systems contain material, measurement and environmental variation, nominal predictions are insufficient. Sensitivity and uncertainty analysis identify influential sources of variation and support robust design. Single-objective optimization is useful for focused decisions, but multi-objective methods reveal trade-offs through Pareto-optimal solutions. The practical outcome should be a validated, uncertainty-aware process window that meets technical, economic and safety requirements.

Key Terms

Analytical model: A model derived from physical laws using simplifying assumptions and often expressed in closed or semi-closed form.

Boundary condition: A specified value or relationship at the boundary of a modeled domain.

Calibration: Estimation of uncertain model parameters using observations.

Design of experiments: A systematic method for planning experiments to estimate effects efficiently.

Finite-element method: A numerical technique that discretizes a continuous domain into interconnected elements.

Hybrid model: A model combining physical equations with empirical, statistical or machine-learning components.

Interaction: A condition in which the effect of one factor depends on the level of another factor.

Model validation: Evaluation of whether a model adequately represents the physical system for its intended use.

Model verification: Checking whether the mathematical model has been implemented and solved correctly.

Multiphysics model: A model coupling two or more physical fields, such as thermal, mechanical and electromagnetic fields.

Noise factor: An uncertain or uncontrolled variable that affects process performance.

Pareto-optimal solution: A solution for which improvement in one objective requires deterioration in another.

Process window: A region of operating conditions that satisfies specified performance and feasibility requirements.

Response surface: An estimated functional relationship between process factors and a response.

Robust optimization: Optimization that accounts explicitly for variation and seeks acceptable performance under uncertainty.

Sensitivity analysis: Study of how changes in inputs influence model outputs.

Surrogate model: A computationally inexpensive approximation of an expensive simulation or experiment.

Uncertainty propagation: Quantification of how input uncertainty produces output uncertainty.

Review Questions

Short Answer Questions

1. Differentiate between model verification and model validation.

2. What is the role of randomization in a designed experiment?

3. Define a process window in the context of manufacturing.

4. Why are interactions important in factorial experimentation?

5. State two reasons why an analytical model may be preferred to a detailed numerical model during early process planning.

Long Answer Questions

1. Explain how analytical, finite-element, empirical and hybrid models complement one another in process development.

2. Describe the stages involved in developing and validating a finite-element model for a thermomechanical manufacturing process.

3. Explain response-surface methodology, including the purpose of coded variables, quadratic terms and confirmation experiments.

4. Discuss uncertainty propagation, sensitivity analysis and robust optimization in establishing reliable manufacturing conditions.

5. Explain why multi-objective optimization is generally more appropriate than single-objective optimization for advanced manufacturing processes.

Analytical / Application Questions

1. A machining process has low roughness at low feed but poor productivity, while high feed increases force and tool wear. Propose an experimental and optimization strategy that identifies a robust operating region rather than a nominal optimum.

2. A simulation predicts acceptable distortion, but confirmation builds show large variation. Identify possible sources of discrepancy and design a verification, validation and uncertainty-analysis plan.

3. For an additive process, increasing energy input improves density until a threshold, after which keyhole porosity and distortion increase. Explain how you would use screening, response-surface modeling, Monte Carlo analysis and Pareto optimization to select production parameters.

22. Multiple Choice Questions with Answers

Question 1

Which statement best distinguishes verification from validation in process simulation?

A. Verification compares simulation with experiments, while validation checks mesh quality.

B. Verification checks whether equations are solved correctly, while validation checks adequacy against physical observations.

C. Verification estimates uncertain parameters, while validation minimizes the objective function.

D. Verification is performed only for analytical models, while validation is used only for empirical models.

Correct Answer: B. Verification checks whether equations are solved correctly, while validation checks adequacy against physical observations.

Question 2

Why is a response-surface model generally considered local?

A. It can only be fitted with one experimental observation.

B. Its coefficients are always physically universal constants.

C. Its predictive accuracy is principally supported within the experimental region used to fit it.

D. It cannot include interaction terms.

Correct Answer: C. Its predictive accuracy is principally supported within the experimental region used to fit it.

Question 3

A process factor has a weak main effect but participates in a strong interaction. What is the appropriate interpretation?

A. The factor can always be discarded.

B. The factor may strongly affect the response at particular levels of another factor.

C. The experiment must have zero measurement error.

D. The response surface is necessarily invalid.

Correct Answer: B. The factor may strongly affect the response at particular levels of another factor.

Question 4

Which experimental practice primarily protects against time trends in machine or environmental conditions?

A. Randomization

B. Mesh refinement

C. Parameter calibration

D. Objective weighting

Correct Answer: A. Randomization

Question 5

What is the main purpose of Monte Carlo uncertainty analysis?

A. To replace all physical experiments permanently

B. To calculate a single deterministic optimum without constraints

C. To estimate the distribution of outputs resulting from uncertain inputs

D. To eliminate model-form error automatically

Correct Answer: C. To estimate the distribution of outputs resulting from uncertain inputs

Question 6

A Pareto-optimal solution is one for which:

A. All objectives achieve their individual independent optima.

B. No objective can improve without worsening at least one other objective.

C. The process has zero experimental variation.

D. The solution always minimizes production cost.

Correct Answer: B. No objective can improve without worsening at least one other objective.

Question 7

Why should confirmation experiments be conducted after optimization?

A. Optimization algorithms cannot use constraints.

B. They test whether predicted performance is achieved under actual process conditions.

C. They make uncertainty analysis unnecessary.

D. They convert a numerical model into an analytical model.

Correct Answer: B. They test whether predicted performance is achieved under actual process conditions.

Question 8

Which situation is most appropriate for a surrogate model?

A. The original model is inexpensive and instantaneous.

B. The governing physics is completely unknown and no data exist.

C. Repeated evaluation of a validated but computationally expensive simulation is required.

D. The process has no measurable responses.

Correct Answer: C. Repeated evaluation of a validated but computationally expensive simulation is required.

Question 9

In robust process design, noise factors are important because they:

A. Are always controllable through software.

B. Represent variations that can cause performance to depart from nominal predictions.

C. Are equivalent to equality constraints.

D. Eliminate the need for replication.

Correct Answer: B. Represent variations that can cause performance to depart from nominal predictions.

Question 10

A thermal model is calibrated to one temperature measurement by changing a heat-transfer coefficient, but it predicts distortion poorly. What is the most likely concern?

A. The model may be compensating for an incorrect physical structure or heat-source representation.

B. The model has too many independent validation data sets.

C. The response surface contains no curvature.

D. Randomization has been excessive.

Correct Answer: A. The model may be compensating for an incorrect physical structure or heat-source representation.