Mathematical Modeling, Uncertainty, and Validation

Introduction

Engineering mathematics becomes most valuable when it connects physical observations to decisions about design, control, safety, and performance. A mathematical model is a deliberately simplified representation of a physical system. In mechatronics, models may describe electrical circuits, mechanical motion, thermal exchange, sensors, actuators, or interactions among these domains.

A defensible engineering model must do more than produce numbers. It should be based on stated physical laws, use consistent units, identify assumptions, quantify uncertainty, and be compared with evidence. This chapter develops a practical workflow:

1. define the purpose and boundary of the model;
2. translate physical laws into equations;
3. check dimensions and limiting behavior;
4. nondimensionalize when useful;
5. analyze sensitivity and uncertainty;
6. identify parameters from measurements;
7. calibrate and verify the implementation;
8. validate the model against independent evidence; and
9. communicate limitations for engineering decisions.

Learning objectives

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

• formulate a dimensionally consistent model from conservation laws and constitutive relations;
• distinguish assumptions, parameters, variables, inputs, and outputs;
• use nondimensional variables to reveal dominant effects;
• calculate local sensitivity and propagate measurement uncertainty;
• estimate unknown parameters using least squares;
• distinguish verification from validation and calibration;
• design a defensible validation experiment; and
• evaluate the limitations and ethical implications of model-based decisions.

1. The structure of an engineering model

A model has a purpose, a system boundary, inputs, states, parameters, outputs, and assumptions. For a motor-drive system, for example, the applied voltage and load torque may be inputs; current and angular speed may be states; resistance, inductance, inertia, and friction coefficients are parameters; and speed or position may be outputs.

A dynamic model is often written in state-space form:

Here, x is the vector of states, u is the input vector, p is the parameter vector, y is the output vector, and t is time. The functions f and g describe state evolution and measurement output. This representation is general; linear models are a special case:

where A, B, C, and D are matrices of compatible dimensions.

A useful distinction is:

The choice of model depends on purpose. A controller may need a fast low-order model, while a thermal-safety study may require a detailed heat-transfer model. Greater complexity is not automatically greater accuracy: unnecessary parameters can make identification difficult and obscure the dominant physics.

2. Translating physical laws into equations

2.1 Conservation laws

Many engineering models begin with a balance:

For a translational mechanical system, Newton’s second law gives

where m is mass in kilograms, x is displacement in metres, and ΣF is the net force in newtons. For a mass–spring–damper system,

where c is viscous damping in N·s/m, k is stiffness in N/m, and F(t) is applied force in N. Every term has units of force: kg·m/s², (N·s/m)(m/s), and (N/m)m.

For an electrical resistor–inductor–capacitor circuit, Kirchhoff’s voltage law may produce

where L is inductance in henries, R resistance in ohms, C capacitance in farads, q charge in coulombs, i current in amperes, and v voltage in volts. The terms L di/dt, Ri, and q/C all have units of volts.

2.2 Constitutive relations and assumptions

Conservation laws are supplemented by constitutive relations describing material or component behavior. Examples include Hooke’s law, F = kx, Ohm’s law, V = IR, and Newtonian viscous force, F_d = cv. These relations are approximations valid over specified ranges.

Typical assumptions include rigid bodies, linear elasticity, constant temperature, negligible leakage, small displacement, or negligible friction. Assumptions should be written explicitly because they define the model’s domain of validity. A linear spring model may be appropriate for small deflections but inappropriate near material yielding.

2.3 Dimensional consistency

A physical equation is dimensionally consistent if all terms added or equated have identical dimensions. If a quantity has dimensions of mass M, length L, and time T, velocity has dimensions L T^-1, while force has dimensions M L T^-2. Dimensional consistency cannot prove that an equation is physically correct, but inconsistency proves that it is incorrect.

Useful checks include:

• do both sides have the same SI units?
• are arguments of exponential, logarithmic, and trigonometric functions dimensionless?
• do initial and boundary conditions use compatible units?
• do limiting cases behave physically?

For example, the expression e^-t/τ is valid only when t and τ have the same time unit, so that t/τ is dimensionless.

3. Nondimensionalization and scaling

Nondimensionalization replaces dimensional variables with ratios to characteristic scales. It reduces the number of independent parameters and identifies competing effects.

Consider the mass–spring–damper equation with a characteristic displacement X and time scale T. Define

where ξ and τ are dimensionless. Substitution gives

Divide by kX and define the natural frequency and damping ratio:

Choosing T = 1/ω_n gives

The dimensionless damping ratio ζ compares damping with critical damping. The dimensionless forcing amplitude compares applied force with the characteristic elastic force kX. Dimensionless groups often permit results from one scale of apparatus to inform another, provided the relevant groups are matched.

A second common group is the Reynolds number in fluid mechanics:

where ρ is density, U characteristic speed, L characteristic length, and μ dynamic viscosity. Re compares inertial effects with viscous effects. It is meaningful only when the model assumptions and flow regime are appropriate.

4. Sensitivity analysis

Sensitivity measures how a model output changes when an input or parameter changes. For output y = f(p), the local absolute sensitivity to parameter p_i is

Its units are output units per parameter unit. A dimensionless relative sensitivity is often more useful:

provided y and p_i are nonzero. Approximately, a small relative parameter change produces

This linear approximation is local and should not be used for large changes or strongly nonlinear models.

Worked example: static actuator deflection

Suppose a linear actuator supports a load F = 100 N with stiffness k = 2,000 N/m. The predicted deflection is

For x = F/k, the relative sensitivities are +1 with respect to F and −1 with respect to k. Therefore, a 3% increase in load and a 2% increase in stiffness give the first-order estimate

The predicted deflection increases by approximately 1%, to 0.0505 m. This analysis identifies load and stiffness as equally influential in relative terms, while their practical uncertainty may differ.

Sensitivity analysis supports sensor selection, tolerance allocation, experiment design, and model reduction. A parameter with low sensitivity may not warrant expensive measurement, whereas a highly sensitive parameter requires careful characterization.

5. Uncertainty and propagation

Uncertainty is lack of complete knowledge about a quantity; it is not the same as an error or a mistake. Random variation may be estimated statistically, while systematic effects arise from bias, calibration, environmental conditions, or an incomplete model.

If independent measured quantities x_i have standard uncertainties u(x_i), and y = f(x_1, …, x_n), the first-order uncertainty propagation formula is

For correlated inputs, include covariance terms:

The independent-input expression assumes small uncertainties and a sufficiently smooth model near the measured values. Monte Carlo propagation is preferable when the model is strongly nonlinear, input distributions are non-normal, or limits are important.

Worked example: resistance power

Power in a resistor is P = V²/R. Let V = 12.0 ± 0.2 V and R = 6.0 ± 0.1 Ω, with independent uncertainties. The nominal power is

The relative first-order uncertainty is

Thus,

and u(P) ≈ 24.0×0.0373 = 0.90 W. The result is approximately 24.0 ± 0.9 W, assuming the stated uncertainties represent standard uncertainties and no important bias has been omitted.

A reported result should match the precision of its uncertainty. False precision, such as reporting 24.000000 W from these measurements, can mislead users about confidence.

6. Parameter identification and calibration

Parameter identification estimates unknown model parameters from input–output data. Calibration adjusts a model or instrument so that its predictions agree with reference observations under specified conditions. Calibration is not validation: calibration uses data to fit parameters, whereas validation tests predictive performance using data not used for fitting.

For a linear model,

where y is the measurement vector, Φ is the known design matrix, θ is the unknown parameter vector, and ε is measurement noise. Ordinary least squares estimates θ by minimizing the sum of squared residuals:

when ΦᵀΦ is invertible. If columns of Φ are nearly dependent, the problem is ill-conditioned and parameter estimates may be highly uncertain. Weighted least squares is appropriate when measurements have unequal variances:

where W is commonly a weighting matrix proportional to the inverse noise covariance.

Identification requires informative data. If an experiment never excites a parameter’s effect, that parameter cannot be estimated reliably regardless of the numerical algorithm. Input design should cover the intended operating range while respecting safety limits.

Residual analysis is essential. Residuals should be examined for bias, trends, changing variance, autocorrelation, and dependence on operating conditions. A small average residual does not guarantee a good model if positive and negative errors cancel.

7. Verification, calibration, and validation

These terms describe different questions:

Verification concerns mathematical and computational correctness. For a simulation, it may include checking numerical convergence, conservation of energy, analytical special cases, and limiting behavior. A verified model can still be physically inadequate.

Validation is purpose-dependent. A model validated for low-speed positioning may not be valid for high-speed vibration, even if both concern the same machine. Validation should specify operating conditions, performance metrics, acceptable error, and the intended decision.

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

where y_i is the measured value, ŷ_i is the predicted value, and N is the number of observations. RMSE has the same units as the output and penalizes large errors. Mean absolute error,

is less sensitive to isolated large errors. Metrics must be interpreted with plots, uncertainty intervals, and engineering tolerances; no single metric proves validity.

Data used for calibration and validation should be separated. If data are scarce, repeated experiments at different operating conditions are preferable to randomly splitting highly correlated time-series samples. Validation conditions should include the loads, temperatures, speeds, and disturbances relevant to deployment.

8. Illustrative case: identifying and validating a temperature model

Background

A small enclosed motor controller overheats during prolonged operation. An engineer proposes a lumped thermal model in which the controller has uniform temperature T and exchanges heat with ambient air at temperature T_a.

Problem and model

With generated power P and thermal resistance R_θ, the model is

where C_θ is thermal capacitance in J/K and R_θ is thermal resistance in K/W. The product τ = R_θC_θ is the thermal time constant. For constant P and T_a, the solution is

The assumptions are uniform component temperature, constant parameters, constant power, and heat loss represented by a linear resistance. Radiation, airflow changes, and spatial temperature gradients may violate these assumptions.

Identification

A step-power test applies P = 20 W in an ambient temperature of 25 °C. The measured steady temperature is approximately 85 °C, so the estimated thermal resistance is

Suppose the temperature reaches 63.2% of its final temperature rise after 480 s. For a first-order response, that time is τ, so τ̂ = 480 s. Therefore,

Verification and validation

The implementation is first verified by comparing its numerical solution with the analytical expression above. Energy-balance checks confirm that the rate of stored thermal energy equals generated heat minus heat loss.

For validation, a separate test uses 15 W at a different ambient temperature and a changed airflow condition. The measured and predicted temperature histories are compared using RMSE, maximum error, and whether the model correctly predicts the safety limit. If the model fits the calibration test but consistently underpredicts temperature under increased airflow restriction, the likely causes include a changing thermal resistance, radiation, or nonuniform heating.

Outcome and learning points

The model may be suitable for early warning and controller design if its error remains below the safety margin in the intended operating range. It should not be used without further evidence for certification or extreme environments. The case illustrates that parameter fitting, code verification, and physical validation are separate activities. It also shows why a model’s purpose and operating domain must accompany every prediction.

9. Limitations, governance, and professional practice

Models necessarily omit some phenomena. Linearization may fail near saturation, friction may depend on direction and speed, sensors may have drift, and parameters may vary with temperature or aging. Extrapolation beyond validated conditions is a major source of engineering failure.

Safety-critical decisions require conservative treatment of uncertainty. Engineers should preserve traceability from data and assumptions to conclusions, document versioned code and parameter sets, and report unfavorable as well as favorable validation results. Measurements should be collected ethically and safely, with attention to privacy when human activity or personal data are involved.

AI and machine-learning models can assist parameter estimation, surrogate modeling, anomaly detection, and experiment design. However, a high predictive score does not establish physical validity. Data leakage, distribution shift, opaque features, and lack of extrapolation control are serious risks. Physics-informed or hybrid models can combine conservation laws with data-driven corrections, but the correction must still be tested independently and bounded where safety matters.

Best-practice workflow

1. State the engineering decision and acceptable error.
2. Define system boundaries, inputs, outputs, and assumptions.
3. Derive equations from conservation and constitutive laws.
4. Check units, signs, limiting cases, and initial conditions.
5. Nondimensionalize to identify dominant effects.
6. Perform sensitivity analysis before collecting expensive data.
7. Design informative experiments and quantify measurement uncertainty.
8. Identify parameters using appropriate estimation methods.
9. Verify equations, code, numerical convergence, and conservation properties.
10. Validate against independent data across the intended operating domain.
11. Report residuals, uncertainty, limitations, and possible failure modes.
12. Monitor performance after deployment and update the model responsibly.

Chapter Summary

Mathematical modeling converts physical laws and engineering assumptions into predictive relationships. A sound model distinguishes inputs, states, parameters, outputs, and residuals; maintains dimensional consistency; and states its domain of applicability. Nondimensionalization exposes the ratios governing behavior and can simplify both analysis and experimentation.

Sensitivity analysis identifies influential quantities, while uncertainty propagation estimates how imperfect measurements affect predictions. Least-squares identification and calibration use data to estimate parameters, but require informative experiments and careful residual analysis. Verification asks whether equations and implementations are correct; validation asks whether the model represents the real system adequately for a specified purpose. Independent validation, transparent reporting, and conservative treatment of uncertainty are essential for defensible mechatronics decisions.

Key Terms

Model: A simplified mathematical representation of a physical system.

System boundary: The defined region or components included in a model.

State: A variable containing information needed to describe future system behavior.

Parameter: A model quantity treated as fixed or slowly varying during analysis.

Constitutive relation: An equation describing material or component behavior.

Dimensional consistency: The requirement that terms being added or equated have compatible units.

Nondimensionalization: Replacement of dimensional variables by ratios to characteristic scales.

Dimensionless group: A unit-free combination of variables, such as Reynolds number.

Sensitivity: Change in an output caused by change in an input or parameter.

Uncertainty: Incomplete knowledge represented quantitatively or qualitatively.

Covariance: A measure of joint variation between two uncertain quantities.

Parameter identification: Estimation of unknown model parameters from observations.

Calibration: Adjustment of parameters or instruments using reference data.

Residual: Difference between a measured and predicted value.

Verification: Demonstration that a model or implementation was constructed correctly.

Validation: Evaluation of whether a model adequately represents reality for a stated purpose.

RMSE: Root-mean-square error between predictions and observations.

Extrapolation: Prediction outside the conditions supported by available evidence.

Review Questions

Short Answer Questions

1. Distinguish a state variable from a parameter in a dynamic model.
2. Why must the argument of an exponential function be dimensionless?
3. What physical comparison is represented by the damping ratio?
4. State one difference between verification and validation.
5. Why can a small mean residual fail to demonstrate model adequacy?

Long Answer Questions

1. Explain a systematic process for deriving and checking a mass–spring–damper model.
2. Describe how nondimensionalization can reduce model complexity and support scale-independent experimentation.
3. Explain first-order uncertainty propagation, including the role of covariance.
4. Compare calibration, parameter identification, verification, and validation.
5. Discuss how assumptions and operating-domain limits should be communicated in a safety-related model.

Analytical / Application Questions

1. A sensor measures voltage as 5.00 ± 0.05 V and resistance as 10.0 ± 0.2 Ω. Use first-order propagation for P = V²/R to estimate nominal power and standard uncertainty, assuming independence.
2. Design a validation experiment for a thermal model of a motor controller. Specify inputs, measured outputs, calibration and validation data, error metrics, and acceptance criteria.
3. A fitted model performs well at low speed but underpredicts vibration at high speed. Analyze at least three possible physical, measurement, or identification causes and propose tests to distinguish them.

Multiple Choice Questions with Answers

Question 1

Which quantity is normally a model parameter?

A. Applied voltage that is deliberately varied

B. Rotor inertia estimated for a motor

C. Measured shaft position at each time

D. Instantaneous prediction residual

Correct Answer: B. Rotor inertia estimated for a motor

Question 2

In the equation m d²x/dt² + c dx/dt + kx = F, what are the units of k?

A. N/m

B. N·s/m

C. kg/m

D. m/N

Correct Answer: A. N/m

Question 3

What is the principal purpose of nondimensionalization?

A. To remove all experimental uncertainty

B. To guarantee a model is physically valid

C. To reveal governing ratios and reduce independent quantities

D. To convert every model into a linear model

Correct Answer: C. To reveal governing ratios and reduce independent quantities

Question 4

For y = F/k, the relative sensitivity of y to k is

A. +1

B. 0

C. k/F

D. −1

Correct Answer: D. −1

Question 5

For independent variables, first-order variance propagation is based on

A. Only the largest input uncertainty

B. The squared partial derivatives and input variances

C. The product of all input means

D. The number of model states only

Correct Answer: B. The squared partial derivatives and input variances

Question 6

If P = V²/R, V = 12 V and R = 6 Ω, the nominal power is

A. 2 W

B. 18 W

C. 24 W

D. 72 W

Correct Answer: C. 24 W

Question 7

Which activity primarily asks whether computer code correctly implements the equations?

A. Validation

B. Verification

C. Calibration

D. Extrapolation

Correct Answer: B. Verification

Question 8

Why should calibration and validation data generally be separated?

A. To increase the number of fitted parameters

B. To prevent all measurement noise

C. To test predictive performance on data not used for fitting

D. To make dimensional analysis unnecessary

Correct Answer: C. To test predictive performance on data not used for fitting

Question 9

For a first-order thermal response, the time to reach approximately 63.2% of the final temperature change equals

A. Half the time constant

B. One time constant

C. Two time constants

D. The thermal resistance

Correct Answer: B. One time constant

Question 10

Which result is most indicative of poor parameter identifiability?

A. Inputs excite several independent system behaviors

B. The design matrix has nearly dependent columns

C. Residuals are plotted with units

D. The model has a stated system boundary

Correct Answer: B. The design matrix has nearly dependent columns