Numerical, Experimental, and Data-Driven Methods

Introduction

Heat and mass transfer problems often become too complex for closed-form solutions. Irregular geometries, temperature-dependent properties, mixed boundary conditions, turbulent flow, radiation, species reactions, and transient operation require a combination of mathematical modelling, numerical computation, experimentation, and data analysis. These approaches are complementary rather than interchangeable. A numerical model predicts behaviour from governing equations; an experiment measures behaviour in a real system; and a data-driven method extracts patterns, estimates parameters, or improves predictions from observations.

A credible engineering result depends on the entire chain: defining the physical problem, selecting governing equations, discretizing them appropriately, measuring relevant variables, quantifying uncertainty, validating predictions, and communicating limitations. A finely meshed simulation is not automatically accurate, and a large dataset is not automatically informative.

Learning Objectives

By the end of this chapter, the reader should be able to:

1. Formulate heat- and mass-transfer problems for numerical solution.
2. Explain the principles of finite-difference, finite-volume, and introductory computational fluid dynamics (CFD) methods.
3. Select sensors and design experiments for thermal and species-transport measurements.
4. Estimate measurement uncertainty and distinguish validation from verification.
5. Interpret infrared thermal images and laboratory or operational datasets.
6. Apply regression and basic machine-learning methods responsibly to engineering data.
7. Evaluate a combined numerical–experimental workflow in an engineering case.

1. From Physical System to Computational and Experimental Model

A heat- or mass-transfer model begins with conservation laws. For a stationary, isotropic solid with constant thermal conductivity and volumetric heat generation, the energy equation is

Here, ρ is density (kg m⁻³), c_p is specific heat capacity (J kg⁻¹ K⁻¹), T is temperature (K), t is time (s), k is thermal conductivity (W m⁻¹ K⁻¹), and q̇ is volumetric heat generation (W m⁻³). For constant k, the diffusion term is k∇²T. The equation assumes local thermal equilibrium and continuum behaviour.

A general species conservation equation for species A is

where Y_A is mass fraction, u is velocity (m s⁻¹), D_AB is mass diffusivity (m² s⁻¹), and ṁ_A''' is volumetric production or consumption of A (kg m⁻³ s⁻¹). In dilute, constant-density systems, this may be simplified to an advection–diffusion equation.

The governing equations are incomplete without initial and boundary conditions. Common thermal conditions include:

• Prescribed temperature: T = T_w.
• Prescribed heat flux: −k∂T/∂n = q''.
• Convection: −k∂T/∂n = h(T_s − T_∞).
• Radiation: −k∂T/∂n = εσ(T_s⁴ − T_sur⁴).

In the convection condition, h is the convective heat-transfer coefficient (W m⁻² K⁻¹), T_s and T_∞ are surface and fluid temperatures, ε is emissivity, and σ = 5.670 × 10⁻⁸ W m⁻² K⁻⁴. Radiation temperatures must be absolute temperatures in kelvin. Boundary conditions for species may prescribe concentration, mass fraction, or flux.

Before solving, the engineer should identify the dominant mechanisms using dimensionless groups. For example,

is the Biot number, where L_c is a characteristic length. A small Bi indicates relatively small internal conduction resistance compared with surface convection resistance; a large Bi indicates important internal temperature gradients. The Fourier number is

where α = k/(ρc_p) is thermal diffusivity. For fluid transport, the Péclet number is

for heat transfer or Pe = UL_c/D_AB for mass transfer. Large Pe indicates advection-dominated transport, which can create numerical oscillations if discretized carelessly.

2. Discretization: Turning Equations into Algebra

Discretization replaces a continuous domain by nodes, cells, or control volumes. The differential equations then become algebraic equations that can be solved by direct or iterative methods. A mesh must resolve important physical length scales, boundaries, sources, and gradients. Refining a mesh reduces discretization error but increases computational cost.

2.1 Finite-difference method

Finite-difference methods approximate derivatives at discrete points. Consider one-dimensional transient conduction with constant properties and no generation:

Using a forward difference in time and a central difference in space gives the explicit scheme

where Fo_x = α∆t/(∆x)², i identifies a spatial node, and n identifies a time level. For a uniform grid and one-dimensional diffusion, stability requires approximately Fo_x ≤ 1/2. In two dimensions with equal spacing, the corresponding restriction is Fo_x ≤ 1/4. These restrictions apply to the explicit central-difference diffusion scheme; implicit schemes have different stability properties.

The central approximation is second-order accurate in space because its truncation error is proportional to (∆x)². The forward time approximation is first-order accurate in time, with error proportional to ∆t. Accuracy and stability are distinct: a stable result can still be inaccurate if the grid or time step is too coarse.

An implicit scheme evaluates the spatial term at the new time level:

This requires solving a system of simultaneous equations at every time step. For linear diffusion, it is unconditionally stable in the usual von Neumann sense, but very large time steps can still produce excessive numerical error and poor resolution of transient behaviour.

Finite differences are straightforward for rectangular, structured domains. Their disadvantages include difficulty with irregular boundaries, nonuniform materials, and strict conservation when fluxes cross complex interfaces.

2.2 Finite-volume method

The finite-volume method integrates the conservation equation over each control volume. For a one-dimensional steady conduction cell with no generation, conservation requires

where e and w denote east and west faces and A is cross-sectional area. Approximating face gradients using neighbouring cell-centre values produces a balance of heat entering and leaving the cell. With generation, the integrated equation becomes

for a generic transported scalar Y_A, where Γ is the diffusion coefficient, u_n is outward normal velocity, S_A is volumetric source per unit volume, V is cell volume, and the summation is over faces.

The principal strength of finite volume is conservation: fluxes shared by adjacent cells are equal and opposite, so conservation is maintained at the discrete level. This makes it especially suitable for CFD and coupled heat–mass transfer. Common interpolation schemes include central differencing, upwind differencing, hybrid schemes, and higher-order bounded schemes. Central differencing is accurate in diffusion-dominated regions but may oscillate for high cell Péclet numbers. First-order upwind is robust and bounded but introduces numerical diffusion. Higher-order schemes reduce false diffusion but require careful treatment near steep gradients.

2.3 Introductory CFD workflow

A basic CFD analysis follows these stages:

1. Define geometry, fluid properties, sources, and operating conditions.
2. Select governing equations, such as continuity, momentum, energy, and species conservation.
3. Generate a mesh with adequate boundary and near-wall resolution.
4. Specify boundary and initial conditions.
5. Select discretization, pressure–velocity coupling, turbulence, and species models.
6. Solve iteratively until residuals and engineering quantities are acceptably converged.
7. Perform mesh and time-step studies.
8. Compare predictions with independent measurements.

For incompressible flow, the continuity equation is

and the momentum equation, in a simplified Newtonian form, is

where p is pressure, μ is dynamic viscosity, and g is gravitational acceleration. Turbulence models introduce additional assumptions and parameters. A model that predicts the mean temperature well may still predict fluctuations, local wall heat flux, or separation poorly.

Convergence should not be judged only by residual reduction. Monitor quantities such as total heat rate, pressure drop, outlet temperature, and species balance. A converged but physically incorrect solution may result from inappropriate boundary conditions, insufficient mesh resolution, or an unsuitable turbulence or radiation model.

2.4 Numerical error and verification

Three important numerical checks are:

• Consistency: the discrete equations approach the differential equations as grid spacing and time step approach zero.
• Stability: errors do not grow without bound during solution.
• Convergence: the numerical solution approaches a limiting solution under refinement.

Grid-convergence studies compare results on successively refined meshes. If a quantity changes from 82.0 to 80.8 and then 80.5 W, the result is approaching a limiting value, although the exact extrapolation requires known or estimated order of accuracy. Richardson extrapolation for a quantity φ is commonly expressed as

where φ_f and φ_c are fine- and coarse-grid results, r is the refinement ratio, and p is the observed order. This relation assumes comparable grids, asymptotic convergence, and a suitable error model. Verification asks, “Was the equation solved correctly?” Validation asks, “Are the equations and assumptions suitable for the real system?”

3. Experimental Planning and Sensor Selection

An experiment should be designed from the question, not from the instruments already available. Define the measurand—the quantity to be inferred or reported—before selecting sensors. Examples include wall temperature, local heat flux, bulk outlet temperature, relative humidity, concentration, velocity, or total heat-transfer rate.

A useful experimental plan identifies:

• independent variables to control, such as flow rate or heater power;
• dependent variables to measure;
• nuisance variables, such as ambient temperature or supply voltage;
• calibration procedures;
• sampling rate and duration;
• replication and randomized test order;
• safety limits and shutdown criteria;
• data storage, metadata, and quality-control rules.

3.1 Common thermal and mass-transfer sensors

Sensor choice involves range, sensitivity, accuracy, repeatability, response time, spatial resolution, environmental compatibility, calibration traceability, and disturbance of the system. A sensor should not be selected solely because its stated resolution is small. Resolution is the smallest displayed increment; accuracy concerns closeness to the true value; precision concerns repeatability.

A thermocouple attached to a surface may read below the actual surface temperature if it loses heat by convection and radiation to the surroundings. A resistance temperature detector may provide better stability but can have a slower response. A heat-flux sensor may alter the local thermal resistance. These are measurement-model issues, not merely instrument errors.

3.2 Sampling and dynamic response

A measurement system has a bandwidth. If a sensor behaves approximately as a first-order system,

where y is the true input, y_m is measured output, and τ_s is the sensor time constant. A step response reaches approximately 63.2% of its final change after one time constant. Sampling must be sufficiently fast to resolve the phenomenon of interest. Sampling below twice the highest frequency is vulnerable to aliasing; in practice, anti-alias filtering and a substantially higher sampling rate are needed when transient information matters.

4. Uncertainty Quantification

An engineering measurement should be reported as an estimate with uncertainty, for example T = 74.6 ± 1.2 °C at an approximately stated coverage level. Uncertainty is not the same as an error, because the exact error is generally unknown.

Type A uncertainty is evaluated statistically from repeated observations. For n observations with sample standard deviation s,

where u_A is the standard uncertainty of the mean, assuming independent and identically distributed observations. Type B uncertainty is evaluated from calibration certificates, manufacturer specifications, resolution, environmental information, and engineering judgement.

For a result y = f(x₁, x₂, …), the combined standard uncertainty, neglecting covariance, is

where u(x_i) is the standard uncertainty of input x_i. If inputs are correlated, covariance terms must be included. For a heat-transfer coefficient estimated from

uncertainty may arise from heat rate Q̇, area A, surface temperature T_s, and fluid temperature T_∞. Because the temperature difference is in the denominator, a small temperature difference can produce a large relative uncertainty.

An expanded uncertainty is often written U = ku_c, where k is a coverage factor, commonly near 2 for approximately 95% coverage under suitable assumptions. The selected coverage statement and distribution assumptions should be reported.

Worked Example: uncertainty in a heat-transfer coefficient

Suppose Q̇ = 500 ± 5 W, A = 0.20 ± 0.002 m², T_s = 80.0 ± 0.3 °C, and T_∞ = 25.0 ± 0.2 °C. Treat the stated uncertainties as standard uncertainties and assume independence.

The estimate is

For h = Q̇/[A∆T], the relative standard uncertainty is approximately

The terms are 0.0100, 0.0100, and approximately 0.00662. Thus the relative uncertainty is √(0.0001 + 0.0001 + 0.0000438) = 0.0150, or 1.50%. Therefore u_c(h) ≈ 0.68 W m⁻² K⁻¹. With k = 2, U ≈ 1.36 W m⁻² K⁻¹, so the result may be reported as h = 45.5 ± 1.4 W m⁻² K⁻¹, subject to the stated assumptions.

5. Thermal Images and Engineering Datasets

An infrared camera does not directly measure temperature. It measures radiative energy in a spectral band and estimates temperature using assumptions about emissivity, reflected radiation, atmospheric transmission, and camera calibration. A simplified radiative relation is

for an idealized opaque diffuse surface with negligible reflected and transmitted radiation. Real surfaces require reflected-apparent-temperature correction and appropriate emissivity settings.

Best practices for infrared thermography include:

• calibrating or checking the camera against a reference source;
• estimating surface emissivity rather than assuming it is one;
• avoiding shiny surfaces or applying a known-emissivity coating where appropriate;
• accounting for reflections from heaters, lamps, and people;
• maintaining suitable focus and viewing angle;
• recording distance, atmospheric conditions, camera settings, and lens information;
• avoiding interpretation of apparent temperature as internal temperature without a thermal model.

A thermal image should be treated as a spatial dataset. Inspect scale limits, saturation, dead pixels, gradients, and regions of interest. A visually dramatic colour map can exaggerate small differences; use numerical temperature scales and consistent limits when comparing images.

For laboratory or operational data, begin with metadata and quality checks. Confirm units, timestamps, sensor identity, calibration status, missing values, duplicated records, impossible values, and synchronization between channels. Plot raw data before filtering. Filtering can hide faults or alter peaks. Separate exploratory analysis from confirmatory analysis, and do not use future information to predict the past in a time-series model.

A useful energy-balance residual is

A persistent nonzero residual may reveal an unmeasured loss, sensor bias, incorrect property value, or faulty timing rather than random noise.

6. Data-Driven Methods for Heat and Mass Transfer

Data-driven methods range from simple regression to machine learning. Their role may be to correlate a heat-transfer coefficient, estimate an unmeasured state, detect faults, or provide a fast surrogate for a validated simulation.

A multiple linear regression model can be written as

where y is the response, x_i are predictors, β_i are fitted coefficients, and ε is the residual. Regression is useful when the relationship is approximately linear in the selected predictors and the residual assumptions are reasonable. Correlation does not establish causation, and a high coefficient of determination does not guarantee physical validity.

In heat transfer, dimensionless correlations often take the form

where Nu = hL/k_f, Re = ρUL/μ, and Pr = μc_p/k_f. The constants C, m, and n depend on geometry and flow regime. Such correlations must not be extrapolated beyond their experimental range without justification.

For data splitting, training data fit the model, validation data support model selection, and test data provide an approximately independent final assessment. For time-dependent systems, chronological splitting is usually more realistic than random shuffling. Performance measures should match the engineering purpose. Mean absolute error is

and root-mean-square error is

RMSE gives greater weight to large errors. Always compare a data-driven model with a simple baseline and inspect residuals across flow rate, temperature, geometry, and operating regime.

Physics-informed and hybrid approaches constrain a data model using conservation laws, known monotonicity, dimensional analysis, or simulation outputs. They can improve generalization, but they do not eliminate the need for reliable data and validation. AI systems may reproduce sensor bias, encode unsafe extrapolations, or obscure physical reasoning. Their predictions should therefore include applicability limits, monitoring, and human review.

7. Verification, Validation, and Reproducibility

Verification concerns correctness of implementation. It includes code tests, comparison with analytical solutions, conservation checks, manufactured-solution tests, and mesh or time-step refinement. Validation concerns agreement with reality for a specified purpose. It requires independent experimental data and a clear comparison metric.

Validation should compare like with like: local simulation temperature with local sensor temperature, area-averaged prediction with area-averaged measurement, and time-resolved prediction with time-resolved data. Account for measurement uncertainty and numerical uncertainty rather than declaring a model correct because two nominal values are close.

A reproducible study records geometry versions, mesh statistics, solver settings, property sources, boundary conditions, calibration files, raw data, processing scripts, random seeds where relevant, and software versions. Data governance also includes access control, secure storage, privacy protection where human-related data are collected, and transparent retention policies. Safety-critical decisions require traceable review and should not rely on an unvalidated black-box prediction.

8. Illustrative Case: Cooling of an Electronic Module

Background

An electronic power module dissipates 180 W through an aluminium baseplate and forced-air cooling channel. The design target is to keep the semiconductor junction below 125 °C. A CFD model predicts a maximum baseplate temperature of 96 °C, but the first prototype shows 104 °C near the downstream corner.

Problem

The discrepancy could result from mesh error, an incorrect contact resistance, nonuniform fan flow, inaccurate material properties, sensor bias, or radiation and leakage losses. The objective is not simply to adjust the model until one temperature matches; it is to identify the dominant cause and establish a validated predictive workflow.

Analysis and approach

The team first verifies the numerical model using an analytical one-dimensional conduction case and performs three mesh studies. The maximum temperature changes from 95.1 to 95.8 to 96.0 °C, indicating reasonable grid convergence for that quantity. The model is therefore verified for discretization, but not yet validated.

The experiment uses calibrated RTDs at five baseplate locations, a thermocouple at the air inlet and outlet, a voltage–current measurement of electrical power, and a calibrated flow meter. Sensor attachment is designed to minimize contact resistance and disturbance. The test is repeated three times after steady state is reached. The uncertainty in the temperature difference used to calculate h is reported rather than omitted.

Infrared images reveal a downstream hot region, but the surface is coated with a high-emissivity layer and the camera is corrected for reflected temperature. The image supports the RTD pattern but is not treated as an independent absolute reference without calibration.

The measured fan flow is 18% lower than the assumed design flow because of a filter pressure drop. Updating the inlet boundary condition reduces the predicted cooling performance. A second comparison shows that the aluminium-to-module contact resistance was also underestimated. After both parameters are corrected using independently measured information, the predicted maximum temperature is 103 °C, with a validation difference within the combined numerical and measurement uncertainty.

Outcome

The final model is used to assess a thicker heat spreader and a revised filter. The selected design reduces the measured maximum temperature to 91 °C under the specified test condition. The model is not claimed to be universally valid; its validation range is documented in terms of power, airflow, and ambient temperature.

Learning points

1. Verification cannot compensate for incorrect physical inputs.
2. Boundary conditions can dominate uncertainty in CFD predictions.
3. Thermal images are valuable for locating patterns but require radiometric interpretation.
4. Independent measurements of flow and contact resistance are more defensible than arbitrary calibration.
5. Validation should include uncertainty and an explicitly stated operating range.

9. Limitations and Best Practices

Numerical methods may suffer from discretization error, artificial diffusion, instability, poor mesh quality, incorrect turbulence closure, and uncertain material properties. Experiments may suffer from calibration drift, sensor placement error, heat leakage, inadequate equilibration, and correlated uncertainties. Data-driven models may suffer from biased sampling, leakage between training and test data, overfitting, concept drift, and unsafe extrapolation.

A robust workflow therefore:

• starts with a dimensional and physical assessment;
• uses conservation checks and simple limiting cases;
• designs measurements around the uncertainty budget;
• calibrates instruments and records traceability;
• performs mesh, time-step, and parameter sensitivity studies;
• separates verification from validation;
• preserves raw data and processing provenance;
• reports uncertainty, not only point estimates;
• tests models outside the fitting process;
• communicates assumptions and validity limits clearly.

Chapter Summary

Finite-difference methods approximate derivatives at discrete nodes and are effective for structured domains. Finite-volume methods integrate conservation laws over control volumes and provide strong discrete conservation, making them central to CFD. Stability, convergence, grid refinement, and conservation checks are essential numerical practices. Experiments require a defined measurand, suitable sensors, calibration, adequate sampling, and control of nuisance variables. Measurement uncertainty combines statistical and non-statistical contributions and should be propagated to derived quantities. Infrared cameras provide radiometric estimates that depend on emissivity, reflections, calibration, and viewing conditions. Data-driven models can correlate, predict, or detect faults, but must be evaluated with appropriate data splitting, residual analysis, physical constraints, and clearly stated applicability limits. Verification asks whether a model is implemented correctly; validation asks whether it represents the intended real system. Reliable engineering practice combines numerical, experimental, and data-driven evidence in a traceable workflow.

Key Terms

Discretization: Replacement of continuous equations by algebraic equations on nodes or control volumes.

Finite-difference method: Numerical method that approximates derivatives using values at discrete points.

Finite-volume method: Conservation-based method that integrates equations over finite control volumes.

CFD: Computational fluid dynamics; numerical solution of fluid-flow and related transport equations.

Mesh: Spatial subdivision used in a numerical calculation.

Numerical diffusion: Artificial spreading introduced by a discretization scheme.

Convergence: Approach of a numerical solution toward a limiting solution or residual state.

Verification: Demonstration that equations are implemented and solved correctly.

Validation: Assessment of whether a model represents a real system for a stated purpose.

Measurand: Quantity intended to be measured.

Calibration: Comparison and adjustment of an instrument against a reference.

Resolution: Smallest increment distinguishable by an instrument or recording system.

Type A uncertainty: Uncertainty evaluated statistically from repeated observations.

Type B uncertainty: Uncertainty evaluated from non-statistical information such as certificates or specifications.

Infrared thermography: Measurement and interpretation of thermal radiation to estimate surface temperature.

Aliasing: Misrepresentation of signal frequencies caused by inadequate sampling.

Regression: Statistical estimation of a relationship between response and predictors.

Surrogate model: Reduced computational model trained or constructed to approximate a more detailed model.

Numerical uncertainty: Uncertainty associated with discretization, iterative solution, and numerical approximations.

Review Questions

Short Answer Questions

1. Distinguish numerical verification from experimental validation.
2. What physical meaning does the Biot number provide?
3. Why can first-order upwind differencing produce numerical diffusion?
4. Define Type A and Type B uncertainty.
5. Why must emissivity be considered when interpreting infrared images?

Long Answer Questions

1. Explain the finite-volume method and show why it is naturally conservative.
2. Compare explicit and implicit finite-difference schemes for transient conduction, including stability and computational implications.
3. Describe a complete CFD workflow from geometry definition through validation.
4. Explain how sensor response time and sampling rate affect transient thermal measurements.
5. Discuss how a heat-transfer correlation should be developed, tested, and limited to its valid range.

Analytical / Application Questions

1. A one-dimensional explicit conduction calculation uses α = 1.0 × 10⁻⁵ m² s⁻¹ and ∆x = 0.01 m. Determine the largest time step permitted by the one-dimensional stability condition Fo ≤ 1/2.
2. Design an experiment to determine the convective heat-transfer coefficient of air flowing over a heated plate. Identify controlled variables, measurements, calibration requirements, uncertainty sources, and validation checks.
3. A CFD model predicts a wall temperature of 92 °C while the measured value is 98 ± 2 °C. Propose a structured investigation that separates mesh error, boundary-condition uncertainty, sensor uncertainty, and physical-model inadequacy.

Multiple Choice Questions with Answers

Question 1

Which numerical method integrates a conservation equation over a control volume?

A. Finite-volume method

B. Finite-difference method

C. Polynomial interpolation only

D. Dimensional analysis

Correct Answer: A. Finite-volume method

Question 2

For one-dimensional explicit central-difference diffusion, the usual stability requirement is:

A. Fo ≥ 1

B. Fo ≤ 1/2

C. Fo = 2

D. Fo ≤ −1/2

Correct Answer: B. Fo ≤ 1/2

Question 3

A large Péclet number generally indicates that transport is dominated by:

A. Radiation

B. Advection

C. Chemical equilibrium

D. Solid conduction only

Correct Answer: B. Advection

Question 4

Which quantity is directly associated with the uncertainty of the mean from repeated independent observations?

A. u_A = s/√n

B. u_A = ns

C. u_A = s²/n

D. u_A = √n/s

Correct Answer: A. u_A = s/√n

Question 5

What is the primary purpose of a mesh-refinement study?

A. To establish numerical convergence and estimate discretization effects

B. To calibrate an infrared camera

C. To increase sensor bandwidth

D. To determine fluid emissivity

Correct Answer: A. To establish numerical convergence and estimate discretization effects

Question 6

Which factor can strongly bias an infrared temperature estimate of a shiny surface?

A. Reflected radiation and incorrect emissivity

B. The number of CFD iterations only

C. The density of aluminium only

D. The sample standard deviation alone

Correct Answer: A. Reflected radiation and incorrect emissivity

Question 7

For h = Q̇/[A(T_s − T_∞)], which situation tends to produce especially large relative uncertainty in h?

A. A very small temperature difference

B. A large, independently calibrated temperature difference

C. Zero uncertainty in all measurements

D. A perfectly known heat-transfer area

Correct Answer: A. A very small temperature difference

Question 8

Validation primarily asks whether:

A. The code syntax is correct

B. The model represents the real system adequately for its intended purpose

C. The mesh contains the fewest cells possible

D. The residual plot has any slope

Correct Answer: B. The model represents the real system adequately for its intended purpose

Question 9

Why is chronological splitting often preferred for time-series prediction?

A. It prevents future information from entering the training process

B. It always increases the number of samples

C. It removes all sensor noise

D. It guarantees zero prediction error

Correct Answer: A. It prevents future information from entering the training process

Question 10

A sensor with a first-order time constant reaches approximately what fraction of a step change after one time constant?

A. 10.0%

B. 37.0%

C. 63.2%

D. 100.0%

Correct Answer: C. 63.2%