Numerical, Computational, and Experimental Thermal Engineering

Introduction and Context

Heat-transfer and thermodynamic systems are governed by conservation laws and constitutive relations. In simple geometries, these laws yield analytical solutions; in real engineering systems, however, complex geometry, variable properties, turbulence, phase change, radiation, moving boundaries, and coupled fluid flow often make analytical treatment impossible. Numerical and computational thermal engineering provide methods for approximating such systems, while experiments provide physical evidence for validating models and identifying phenomena omitted from them.

The three activities are complementary. A numerical model predicts fields such as temperature, velocity, pressure, and concentration. A computational implementation converts the model into an algorithm and solves it on a computer. An experiment measures system behaviour, estimates uncertainty, and tests whether the model is credible. Reliable engineering decisions require all three.

Learning Objectives

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

1. Formulate heat-transfer and thermodynamic problems mathematically.
2. Distinguish modelling error, discretization error, iterative error, and measurement uncertainty.
3. Explain finite-difference, finite-volume, and finite-element approaches.
4. Assess stability, consistency, convergence, conservation, and verification.
5. Design thermal experiments and quantify uncertainty.
6. Interpret validation evidence and select suitable computational tools.
7. Evaluate a coupled numerical–experimental engineering investigation.

1. Mathematical Formulation of Thermal Problems

A thermal problem normally consists of a domain, governing equations, material properties, initial conditions, boundary conditions, and quantities of interest. For a stationary, isotropic solid with constant thermal conductivity and volumetric heat generation, the energy equation is

Here, T is temperature (K), k is thermal conductivity (W m⁻¹ K⁻¹), q̇ is volumetric heat generation (W m⁻³), ρ is density (kg m⁻³), and c_p is specific heat capacity (J kg⁻¹ K⁻¹). For constant k, the equation becomes

The equation is valid when the continuum approximation applies and local thermodynamic equilibrium permits a meaningful temperature field. If properties vary strongly with temperature, the divergence form should be retained.

Typical boundary conditions include:

• Prescribed temperature: T = T_s (Dirichlet condition).
• Prescribed heat flux: -k ∂T/∂n = q''_s (Neumann condition).
• Convection: -k ∂T/∂n = h(T_s-T_∞) (Robin condition).
• Radiation: -k ∂T/∂n = εσ(T_s⁴-T_sur⁴).

In the radiation condition, ε is emissivity, σ = 5.670374419×10^-8 W m^-2 K^-4]] is the Stefan–Boltzmann constant, and absolute temperatures must be used. Boundary conditions must be physically compatible; prescribing both temperature and heat flux on the same boundary generally over-specifies the problem.

For fluid flow, the thermal field is coupled to mass, momentum, and energy conservation. A common incompressible energy equation is

where u is velocity. The dimensionless Péclet number compares advection with conduction:

where U is characteristic velocity, L is length, and α = k/(ρc_p) is thermal diffusivity. Large Pe indicates advection-dominated transport and creates greater numerical sensitivity.

2. Discretization Methods

Discretization replaces a continuous domain by a finite set of nodes, control volumes, or elements. The continuous differential equations become algebraic equations. The choice of method should reflect geometry, conservation requirements, boundary complexity, and the desired output.

2.1 Finite-difference method

Finite differences approximate derivatives at grid points. For a one-dimensional steady wall with constant properties and no heat generation,

Using a uniform grid with spacing ∆x, the central approximation is

Taylor expansion shows that the truncation error is of order O((∆x)²). Central differences are therefore second-order accurate in smooth regions. Finite differences are conceptually simple and efficient for structured grids, but irregular geometry and complex boundaries can be awkward.

2.2 Finite-volume method

The finite-volume method integrates conservation equations over a control volume. For steady one-dimensional conduction without generation,

Applying the divergence theorem gives

The net heat flux entering a control volume equals the net heat flux leaving it. This local conservation property is a major reason finite volume is widely used in computational fluid dynamics. At a face between cells, a simple approximation is

where P and E denote adjacent cell centres and k_e is the effective face conductivity. For discontinuous materials, the harmonic rather than arithmetic mean is often appropriate because it preserves resistance in series.

2.3 Finite-element method

Finite elements divide the domain into elements and approximate the unknown field using shape functions:

Here, N_j are shape functions and T_j are nodal temperatures. The weighted-residual or Galerkin procedure produces a matrix equation

for a steady linear problem. K is the global conductivity matrix, {T} is the vector of unknown nodal temperatures, and {F} contains heat sources and boundary contributions. Finite elements are particularly effective for complex solid geometry, structural–thermal coupling, and adaptive mesh refinement.

2.4 Boundary and immersed methods

Complex fluid domains may be treated using body-fitted meshes, immersed-boundary approaches, lattice-based methods, or particle methods. Each introduces different approximations. No method is universally superior: a conservative finite-volume method may be preferred for a flow-through heat exchanger, whereas finite elements may be more convenient for a composite solid with mechanical constraints.

3. Time Integration and Nonlinear Problems

Transient simulations require a time-marching scheme. For a semi-discrete system

an explicit Euler step is

It is inexpensive per step but conditionally stable. An implicit Euler step is

and is generally more stable for diffusion problems, although it requires solution of a linear system. The Crank–Nicolson method averages the spatial operator at times n and n+1 and is formally second-order in time for sufficiently smooth solutions, but can produce oscillations when strong transients or discontinuities exist.

For one-dimensional explicit conduction, a common stability restriction is

where Fo is the Fourier number. This limit applies to the standard one-dimensional central-difference scheme with constant diffusivity and a uniform grid; multidimensional and nonuniform cases require the appropriate modified restriction.

Nonlinearities arise from temperature-dependent properties, radiation, natural convection, and phase change. A residual equation may be written as

Newton’s method linearizes it as

where J = ∂R/∂T is the Jacobian. Under suitable smoothness and a sufficiently good initial estimate, Newton’s method can converge rapidly; otherwise damping, continuation, or a more robust fixed-point method may be needed.

4. Verification, Validation, and Uncertainty

Verification asks, “Was the mathematical model solved correctly?” Validation asks, “Does the model represent the physical system adequately for its intended use?” The distinction is essential.

4.1 Verification

Verification includes:

• Code verification: comparison with manufactured solutions, analytical solutions, or benchmark solutions.
• Solution verification: assessment of discretization and iterative errors for the selected grid and time step.
• Conservation checks: comparison of integrated heat input, heat output, and storage.

For a sequence of systematically refined grids, the observed order may be estimated from three solutions, although the calculation requires consistent refinement and comparable convergence. A practical grid-convergence study reports the mesh sizes, quantity of interest, residual criteria, and changes between meshes. A solution should not be called grid independent merely because residuals are small; residual convergence and discretization convergence are different tests.

The global energy balance for a transient control volume can be expressed as

A large imbalance indicates implementation, boundary, property, or convergence problems.

4.2 Validation

Validation compares predictions with measurements under comparable initial and boundary conditions. Agreement for one observable does not prove universal validity. Validation should include uncertainty bands, relevant operating regimes, and quantities that challenge the model, not only convenient data points.

A useful error measure is the root-mean-square error:

where N is the number of observations. RMSE has the same units as y, but it does not replace physical interpretation or uncertainty analysis.

4.3 Uncertainty propagation

For a measured quantity Y=f(X₁,X₂,...,X_n), the first-order estimate of standard uncertainty is

where u_Xi is standard uncertainty and Cov represents covariance. If inputs are independent, covariance terms vanish. For a heat-transfer coefficient estimated from

uncertainty in heat rate, area, surface temperature, and fluid temperature all contribute. Linear propagation is appropriate for small uncertainties and an approximately linear local response; Monte Carlo propagation is preferable for strong nonlinearities, non-Gaussian distributions, or correlated inputs.

5. Experimental Thermal Engineering

An experiment begins with a measurable question, not with equipment. The investigator must define the measurand, operating range, acceptable uncertainty, sampling rate, calibration procedure, and data-reduction equation.

5.1 Sensors and measurement principles

Common instruments include thermocouples, resistance temperature detectors, thermistors, infrared cameras, heat-flux sensors, pressure transducers, flow meters, and power meters. Selection depends on range, response time, spatial resolution, environmental compatibility, and loading effects.

A thermocouple measures a voltage related to a temperature difference and requires appropriate reference-junction compensation. An RTD uses the temperature dependence of electrical resistance and normally offers high repeatability. Infrared measurement is non-contact, but apparent temperature depends strongly on emissivity, reflected radiation, atmospheric attenuation, and line of sight.

5.2 Calibration and experimental design

Calibration establishes the relationship between instrument output and a reference quantity. A calibration should cover the intended range and use traceable standards where required. The apparatus should include heat-loss characterization, insulation assessment, electrical safety, leak testing, and repeatability checks.

A sensible sequence is:

1. Define the measurand and governing relation.
2. Identify dominant uncertainty sources.
3. Select sensors and locations.
4. Calibrate and document instruments.
5. Establish steady or statistically stationary operation.
6. Record raw signals, time stamps, environmental conditions, and metadata.
7. Perform independent energy and plausibility checks.
8. Report processed results with uncertainty.

Sensor placement is part of the model. A thermocouple in a wall measures its own junction temperature, which may differ from the fluid temperature because of conduction along wires, radiation exchange, contact resistance, and finite response time.

5.3 Data acquisition and signal processing

Sampling frequency should resolve the relevant dynamics. Sampling too slowly causes aliasing; sampling much faster than needed increases data volume and may expose electrical noise. Anti-alias filtering, grounding, shielding, synchronization, and sensor time-response characterization are important.

A moving average can reduce random noise but may distort transients. Filtering must therefore be selected before examining results and documented as part of the method. Outlier rejection should be based on physical and statistical criteria, not on the desire to improve agreement.

6. Coupled Computational–Experimental Workflow

A robust investigation follows a traceable loop:

Figure: Iterative relationship between physical modelling, computation, experiment, and validation.

The workflow should preserve provenance: version-controlled code, documented input files, calibration records, mesh descriptions, solver settings, and raw data. Reproducibility is an engineering quality requirement, not merely a research ideal.

7. Worked Example: One-Dimensional Composite Wall

Consider a plane wall consisting of two layers in steady one-dimensional conduction. Layer 1 has thickness L₁=0.05 m and conductivity k₁=0.8 W m⁻¹ K⁻¹; layer 2 has L₂=0.10 m and k₂=0.2 W m⁻¹ K⁻¹. The inner and outer surface temperatures are 200 °C and 40 °C, respectively. Neglect contact resistance and heat loss through the sides.

The thermal resistance per unit area is

Thus,

R''_t = 0.05/0.8 + 0.10/0.2 = 0.0625 + 0.5 = 0.5625 m² K W⁻¹.

The heat flux is

Therefore,

q'' = (200-40)/0.5625 = 284.4 W m⁻²

approximately. The temperature drop across layer 1 is q''L₁/k₁ = 17.8 K, so the interface temperature is approximately 182.2 °C. This analytical result is a benchmark for a numerical model. A finite-volume solution should approach it as the mesh is refined; any persistent discrepancy indicates a formulation or implementation error rather than ordinary discretization error.

8. Illustrative Case Study: Thermal Management of an Electronic Module

Background

An illustrative engineering team must evaluate an air-cooled electronic module dissipating 120 W. The module contains localized heat-generating components mounted on an aluminium spreader. Reliability requires that the hottest component remain below 85 °C while inlet air is 25 °C.

Problem

The geometry includes narrow channels, contact resistances, temperature-dependent fan performance, and nonuniform heat generation. A purely lumped thermal-resistance model cannot resolve hot spots, but a detailed computational fluid-dynamics model is expensive and uncertain without measurements.

Analysis and approach

The team first constructs a conjugate heat-transfer model: solid conduction is coupled to air flow, and electrical power is applied as volumetric or surface heat generation according to component data. A mesh-refinement study is performed using maximum component temperature and outlet temperature as quantities of interest. The solver is checked against an energy balance and a simplified duct-flow benchmark.

A prototype is instrumented with calibrated RTDs on the spreader, thermocouples at the inlet and outlet, a differential-pressure transducer, and an electrical power meter. Air flow is measured independently. Measurements are repeated at three fan speeds and at least three power levels. Uncertainty in maximum temperature includes sensor calibration, placement, spatial nonuniformity, power measurement, and repeatability.

Outcome

The initial simulation predicts 79 °C, while the experiment gives 83 ± 2 °C. Investigation shows that the model omitted a 0.15 K W⁻¹ interface resistance between a component and the spreader. Including the measured contact resistance raises the predicted maximum to 82 °C. Mesh refinement changes the predicted maximum by less than 0.5 °C, and the remaining difference lies within combined experimental and modelling uncertainty.

Learning points

The case demonstrates that agreement is not obtained by arbitrary calibration. The discrepancy was traced to a physically identifiable parameter. It also shows why local peak temperature, rather than only total heat rejection, should be validated. A model suitable for fan selection may not be sufficiently reliable for component reliability assessment.

9. Advanced Topics and Contemporary Practice

Reduced-order and surrogate models

High-fidelity simulations can be expensive in optimization or real-time control. Reduced-order models, response surfaces, Gaussian-process surrogates, and proper-orthogonal-decomposition methods approximate selected outputs over a defined parameter range. Their validity is local to the training domain and must be checked against independent cases.

Data assimilation and digital twins

A digital twin combines a physical asset, a computational model, sensor data, and an updating procedure. Data assimilation can estimate unmeasured states or uncertain parameters. The model must nevertheless obey conservation laws and maintain uncertainty estimates; a data fit alone is not proof of physical correctness.

Machine learning in thermal engineering

Machine learning can accelerate surrogate prediction, detect sensor faults, infer heat-transfer coefficients, and control thermal systems. Risks include extrapolation failure, hidden bias in training data, nonphysical predictions, and poor interpretability. Physics-informed or hybrid models are often safer than unconstrained black-box prediction when conservation and safety limits are critical.

Governance and professional responsibility

Thermal simulations influence safety, energy consumption, emissions, and equipment life. Engineers should document assumptions, disclose validation range, retain raw data, protect calibration records, and avoid presenting a single deterministic number when uncertainty is material. Regulatory requirements may apply to pressure equipment, electrical systems, hazardous fluids, laboratory safety, and environmental emissions. Computational results should be independently reviewed when they support high-consequence decisions.

10. Best Practices and Limitations

Best practice includes defining the question before choosing software; preserving conservation; using physically justified boundary conditions; checking units and property ranges; monitoring residuals and balances; performing mesh and time-step studies; validating against independent measurements; propagating uncertainty; and making code, data, and assumptions auditable.

Important limitations remain. Numerical solutions can be precise but physically wrong if the model is wrong. Experiments can be accurate but incomplete if sensors sample only a few locations. Inverse problems may be ill-conditioned: small measurement errors can cause large inferred-property errors. Turbulence, phase change, contact resistance, radiation exchange, and multiphysics coupling often involve model-form uncertainty that cannot be removed by mesh refinement. Engineering judgement is therefore required alongside computation.

Chapter Summary

Numerical, computational, and experimental thermal engineering form an integrated methodology for analysing systems that cannot be solved reliably by elementary analytical methods alone. Mathematical formulation begins with conservation laws, constitutive relations, properties, initial conditions, and boundary conditions. Finite differences, finite volumes, and finite elements transform continuous equations into algebraic systems, while time integration and nonlinear solvers determine how those systems are advanced and solved.

Verification establishes that equations and algorithms are implemented and solved correctly; validation examines whether the model represents the physical system for its intended purpose. Mesh refinement, residual monitoring, conservation checks, benchmark solutions, calibration, uncertainty propagation, and independent measurements are essential. Experiments require deliberate measurand definition, sensor selection, calibration, acquisition, and data reduction. Modern reduced-order models, digital twins, and machine learning can extend capability, but they do not eliminate physical modelling, uncertainty, or professional responsibility.

Key Terms

Mathematical model: Equations and assumptions representing a physical system.

Discretization: Replacement of a continuous domain by finite computational entities.

Finite-difference method: Derivative approximation at discrete grid points.

Finite-volume method: Conservation-law integration over control volumes.

Finite-element method: Approximation using element shape functions and weighted residuals.

Mesh: Spatial arrangement of computational cells, nodes, or elements.

Residual: Measure of how closely a numerical solution satisfies discrete equations.

Stability: Bounded response of a numerical method to perturbations.

Consistency: Agreement of a discrete equation with the governing differential equation as spacing tends to zero.

Convergence: Approach of the numerical solution to the exact or benchmark solution.

Verification: Demonstration that the equations are solved correctly.

Validation: Assessment of agreement between model predictions and physical observations.

Model-form uncertainty: Uncertainty arising from imperfect physical assumptions.

Measurement uncertainty: Estimated doubt associated with a measured result.

Calibration: Establishment of instrument response against a reference.

Péclet number: Ratio of advective to diffusive transport.

Fourier number: Dimensionless measure of transient thermal diffusion.

Conjugate heat transfer: Coupled heat transfer in solids and fluids.

Reduced-order model: Lower-dimensional approximation of a higher-fidelity model.

Digital twin: Data-connected computational representation of a physical asset.

Review Questions

Short Answer Questions

1. Distinguish verification from validation in thermal engineering.
2. State the physical meaning of the Péclet number.
3. Why is the finite-volume method described as locally conservative?
4. What is the purpose of a calibration procedure?
5. Name four major sources of uncertainty in a heat-transfer experiment.

Long Answer Questions

1. Explain the formulation and numerical solution of a transient conduction problem, including initial conditions, boundary conditions, time integration, and stability.
2. Compare finite-difference, finite-volume, and finite-element methods with respect to geometry, conservation, and typical applications.
3. Discuss a complete verification and validation strategy for a conjugate heat-transfer simulation.
4. Explain uncertainty propagation for an experimentally determined heat-transfer coefficient.
5. Discuss the opportunities and risks of machine learning and digital twins in thermal engineering.

Analytical / Application Questions

1. A two-layer wall has L₁=0.04 m, k₁=1.2 W m⁻¹ K⁻¹, L₂=0.08 m, and k₂=0.16 W m⁻¹ K⁻¹. The boundary temperatures are 180 °C and 30 °C. Determine the steady heat flux and interface temperature, neglecting contact resistance.
2. A one-dimensional explicit conduction code uses α=8×10^-6 m² s⁻¹ and ∆x=0.01 m. Determine the largest time step permitted by Fo≤1/2, and explain why a smaller step may still be useful.
3. Design a validation experiment for a compact heat exchanger. Specify the measurands, sensors, calibration needs, energy-balance check, uncertainty sources, and computational quantities that should be compared with measurements.

Multiple Choice Questions with Answers

Question 1

Which method is based directly on integrating a conservation equation over control volumes?

A. Finite-difference method

B. Finite-volume method

C. Finite-element method

D. Monte Carlo method

Correct Answer: B. Finite-volume method

Question 2

For one-dimensional explicit central-difference conduction, which condition is commonly required for stability on a uniform grid?

A. Fo ≥ 1

B. Fo = 0

C. Fo ≤ 1/2

D. Fo ≤ 2

Correct Answer: C. Fo ≤ 1/2

Question 3

What is the primary purpose of verification?

A. To prove that all experiments are unbiased

B. To determine whether the physical model is universally valid

C. To establish that the equations and numerical implementation are solved correctly

D. To replace calibration of sensors

Correct Answer: C. To establish that the equations and numerical implementation are solved correctly

Question 4

In the relation Pe=UL/α, a large Péclet number indicates that which transport mechanism is relatively dominant?

A. Advection

B. Conduction

C. Radiation

D. Thermal storage only

Correct Answer: A. Advection

Question 5

A wall has total thermal resistance per unit area 0.5 m² K W⁻¹ and a temperature difference of 100 K. Its heat flux is:

A. 0.005 W m⁻²

B. 0.5 W m⁻²

C. 50 W m⁻²

D. 200 W m⁻²

Correct Answer: D. 200 W m⁻²

Question 6

Which temperature scale must be used in the Stefan–Boltzmann radiation relation?

A. Celsius only

B. Fahrenheit only

C. Absolute temperature

D. Any scale with the same increment

Correct Answer: C. Absolute temperature

Question 7

What does grid refinement primarily assess?

A. Sensor calibration error

B. Discretization error

C. Electrical grounding quality

D. Fluid chemical composition

Correct Answer: B. Discretization error

Question 8

If covariance between measurement inputs is known to be nonzero, uncertainty propagation should:

A. Automatically discard all measurements

B. Include covariance terms

C. Set every uncertainty to zero

D. Use only the largest input value

Correct Answer: B. Include covariance terms

Question 9

Which quantity is a direct check on a thermal simulation’s conservation performance?

A. Number of graphics cards used

B. Difference between integrated heat input, output, and storage

C. Colour of the contour plot

D. File size of the mesh

Correct Answer: B. Difference between integrated heat input, output, and storage

Question 10

Why can a model that predicts total heat rejection accurately still be inadequate for electronics reliability?

A. Total heat rejection is never measurable

B. Reliability may depend on local peak temperature and hot spots

C. Reliability depends only on mesh cell count

D. Peak temperature is unrelated to heat transfer

Correct Answer: B. Reliability may depend on local peak temperature and hot spots