From simulation to measurement

A simulation answers the question it was asked, not the question the vehicle poses. The distance between the two is where engineering judgement lives. Moving from a model to a measurement is not a hand-off; it is a loop, and each pass through it should make the model smaller in ambition and better in honesty.

Verification and validation are different questions

Verification asks whether the equations were solved correctly: does the code reproduce a known analytic result, does the answer stop changing when the time step is reduced? Validation asks whether the equations are the right ones: does the model agree with physical measurement across the range it will be used in? A model can pass the first and fail the second. The laboratory’s damping model, for instance, is checked against the closed-form solution of a mass–spring–damper, which verifies the integrator. It says nothing about whether a real damper behaves like a linear coefficient.

Write the assumptions down first

Every model leaves things out, and the useful ones say which. A quarter-car model ignores tyre compliance, friction in the damper, bushings and body pitch. Each omission is a hypothesis: that its effect is small compared with the effect being studied. Writing the list before testing turns surprises into predictions that can fail, which is what makes them informative.

Measurements have uncertainty too

A measurement is not the truth. It is another estimate, with sensor accuracy, calibration, resolution, environment and repeatability all contributing. Before comparing a simulation with a test, estimate how much two runs of the same test differ from each other. If the model and the measurement disagree by less than that scatter, there is nothing to explain. If they disagree by more, the question is whether the model is wrong, the test is wrong, or the two were not describing the same condition.

Repeatability deserves its own line in the test plan. Run the same input several times, in a randomised order, and record the conditions each time: ambient temperature, tyre or fluid temperature, battery state, wind. Drift between the first and last run is often larger than the difference between the two configurations being compared, and without a repeated baseline it cannot be separated from a real effect.

Decide what agreement means beforehand

Acceptance criteria written after seeing the data tend to fit the data. Better to decide in advance which quantities matter, such as settling time, peak temperature or change in CdA, and how close is close enough. This also guards against adjusting the model until it reproduces a single test, which looks like validation and is closer to curve fitting: it will be right for that test and unreliable for the next.

Use the instruments the question needs

The measurement plan follows from the model’s weakest assumption. If a thermal model is sensitive to airflow through a radiator, measure airflow, not only coolant temperature. If a response model depends on damping, measure the damper directly on a rig before trusting the whole vehicle. A good instrumentation plan is a list of the places where the model is most likely to be wrong.

Know the validated domain

Agreement at 60 km/h on dry tarmac is not agreement at 180 km/h in the rain. A validated model has a stated domain, and outside it the model is a guess, however polished its output. The most useful sentence in a simulation report is often the last one: this model has been compared with measurement over this range, and should not be used beyond it.

This site’s studies are fictional and its laboratory is an educational illustration. They are built that way deliberately, so that every number can be traced to an equation on the page. A real programme would add the second half of the loop: measure, compare, and let the measurement change the model.