What Formula 1 knows about measurement

1

A field that budgets its learning

Formula 1 is one of the few industries where the cost of learning is written into the rules.

A Formula 1 car is among the most heavily instrumented machines that operate outside a laboratory. Mercedes describes more than 250 sensors on a car during a race weekend, reporting pressures, temperatures, accelerations and displacements, and more than a terabyte of data per car over the weekend once video and ancillary information are included. The figure is large, but it is not the interesting part. Plenty of industrial sites now generate more.

What sets the sport apart is what surrounds the data. Teams compete on how quickly they can turn measurement into a better car, and the regulations limit almost every means of doing so: how many wind-tunnel runs a team may make, how much aerodynamic computation it may use, how much it may spend in a year, how much track time it gets. The volume of data is not the constraint. The opportunities to learn are.

Under those conditions, deciding what to measure, when, and to answer which question stops being a technical detail. It becomes part of the engineering of the car. That discipline is what this article is about, together with where it transfers to industry and where it does not.

2

Three reasons to measure

The same car carries sensors for three different purposes, and each purpose asks something different of the data.

Mercedes groups its onboard sensors into three categories: control, instrumentation and monitoring. Read as purposes rather than hardware, the distinction is worth keeping. Control sensors exist so the car can run. Instrumentation exists so the team can learn. Monitoring exists so the car survives. A pressure or a temperature can belong to any of the three, and the requirements change with the purpose: how fast it must be read, how precise it must be, how much of it is worth keeping and for how long.

One car, three purposes Illustration

Control
So the car can run
Read by the car's own systems, continuously. It must be fast and dependable. Most of its value is spent within milliseconds.
Instrumentation
So the team can learn
Often fitted for a test or a session and removed for racing: pressure rakes, extra probes, flow-visualisation paint. Dense and precise, and kept for analysis long after the run.
Monitoring
So the car survives
Watches components for wear, heat and failure. It needs coverage over time more than density in any one moment, and its history is what gives a reading meaning.

Industry has all three, but unevenly. Aerospace, pharmaceutical manufacturing and the advanced process industries run sophisticated diagnostic experiments. Many other long-lived assets were instrumented to run and to protect, and carry little whose only job is to improve understanding of the process. Where learning is budgeted, it is often budgeted as a project rather than as a standing part of operations. Many of the quantities that matter most are not measured directly at all. A team cannot read a tyre's internal condition during a race. It infers it from lap times, temperatures, loads and the behaviour of the car, and updates the estimate every lap.

3

The loop at four speeds

Measurement is the first step of a loop that ends in a decision, and Formula 1 runs that loop at several speeds at once.

The structure is the same at every scale. Something is measured. Hidden quantities are estimated from it. A prediction is made about what each available choice would lead to. Someone or something decides, and the result becomes the next measurement. Kalman's 1960 filter gave the estimation step its classic formal treatment: a running estimate of a system's hidden state, corrected each time a noisy measurement arrives, with its own uncertainty carried alongside.

What differs between the loops in a team is not mainly the sensors. It is the time each loop has, who closes it, and what bounds it.

Measure, estimate, decide Illustration

Measure
Estimate
Decide
Bounded by

Fig. 3.1 The same loop of measuring, estimating and deciding, at four timescales within one team. Times are orders of magnitude, not figures from any team.

Read across the four, the faster loops are closed by machines and by the driver, the slower ones by engineers and by the organisation. The slowest loop, the one that decides what the next car looks like, is the one the regulations bound most tightly. It is also the one most like an industrial investment decision.

4

Three versions of one car

A team works with three cars at once, and none of them is the truth.

There is the car in computational fluid dynamics, the scale model in the wind tunnel, and the car on track. The regulations cap the tunnel model at 60% of full size and the tunnel's air speed at 50 metres per second, so even a perfect tunnel measures a different flow from the one the race car meets. Each version disagrees with the others in its own characteristic ways, and much of a team's aerodynamic work is the discipline of knowing those disagreements: correlation, in the sport's vocabulary.

The large pressure rakes seen on cars in testing, and the fluorescent paint that shows how air moves across a surface, are instruments built for that purpose. They do not monitor the car. They test a model of it. Their job is to find the places and conditions where simulation, tunnel and track part company, so that the next round of development can trust each source where it is right and discount it where it is not.

Statistics has a precise name for what they are measuring. Kennedy and O'Hagan's Bayesian calibration framework treats the gap between a computer model and reality, the model discrepancy, as a quantity to be estimated in its own right, with its own uncertainty, rather than as error to be averaged away. The industrial digital twin faces the same problem. A twin that is not continuously checked against the plant it represents drifts from it without saying so.

5

A budget on learning

When every experiment is rationed, choosing the next one becomes an engineering decision.

The FIA's aerodynamic testing restrictions divide each year into six periods of roughly two months. In each period a team may make a set number of wind-tunnel runs, accumulate a set amount of time with the wind on, and use a set amount of computing for aerodynamic simulation. The allowance then scales with the previous championship: the champion receives 70% of the base, the tenth-placed team or a new entrant 115%. Spending is capped separately, at $215 million for 2026.

Base allowance per testing period, 2026 Cited

320wind-tunnel runs

80 hwith the wind on

6 MAUhaerodynamic computation

70–115%scaled by championship position

Source FIA 2026 Formula 1 Regulations, Section F, Article F4 (Issue 11, October 2026). Six periods a year; models up to 60% scale, air speed up to 50 m/s.

The consequence is that an experiment has a price, and the question is what each one buys. This is the territory of experimental design. Howard's information value theory, from 1966, priced a measurement by how much it could change a decision. Bayesian experimental design chooses the next experiment by how much it is expected to reduce uncertainty about what matters. In structural engineering, Kammer's method placed sensors on large space structures by how much independent information each position contributed. Krause, Singh and Guestrin later gave sensor placement under a Gaussian-process model guarantees of near-optimality.

The figure below makes the trade concrete. A model of the car agrees with the track almost everywhere in its operating envelope, but misses badly in one region and modestly in another. A team has a fixed number of runs to find out where. It can sweep the envelope on a plan, sample at random, choose each run adaptively by where the gap looks largest or least known, or spend half its runs on a sweep and half adaptively.

A testing budget, spent four ways Simulation

Runs 16

Largest gap

Size error

Smaller gap

Show the true gaps

Across: ride height · Up: yawTrack below modelTrack above model

Over 200 fresh worlds, at 16 runs
Strategy	Largest gap located	Error in its size, median	Smaller gap caught

Fig. 5.1 A synthetic operating envelope (ride height across, yaw up). The gap between model and track is a large local miss, a smaller one elsewhere and a gentle overall tilt; each run measures the gap at one point with noise, and a Gaussian process estimates it everywhere. "Located": the estimate's largest gap lies within 8% of the envelope's width of the true one. "Caught": the estimate at the smaller gap's centre is within half its true size. The adaptive rule picks the point where the estimated gap plus twice its uncertainty is largest. Worlds and numbers are invented to show the mechanism. Code and results: code/formula-1/experiment. Simulation.

The result depends on the question. With 16 runs, the adaptive strategy locates the largest gap in 60% of worlds against 44% for the planned sweep, and estimates its size with a median error of 35% against 57%. It does so by spending its runs where the first signs of a gap appeared. The same choice makes it worse at the question it was not asked: it catches the smaller gap in 18% of worlds, where the sweep catches it in 33%. Splitting the budget does well on both, locating the largest gap in 64% of worlds with a median error of 23%, while matching the sweep on the smaller one. At 24 runs the pattern holds.

None of this is new mathematics. The point is that the trade is real and can be stated. A measurement plan is tuned for a question, whether or not anyone wrote the question down.

6

Where industry differs

The deeper difference is how long the information has to stay useful.

Formula 1 does not only measure for questions it already knows. Teams meet unexpected failures, track surfaces that change through a weekend, rivals whose next move they cannot see, and regulations that are rewritten every few years. What sets the sport apart is the horizon. It has a single objective that everyone agrees on, lap time. Feedback arrives every session. Test conditions can be largely repeated. And much of what a team learns about one car loses its value when the regulations change, as they did comprehensively for 2026.

Two settings for learning from a physical system
	A Formula 1 team	A long-lived industrial asset
Objective	One, agreed and measurable	Several, shifting over time, sometimes in tension
Feedback	Every session, every lap	Months or years; failures are rare
Conditions	Largely repeatable, by design	Drift with wear, supply, weather and demand
Life of the knowledge	Often a season or a regulation cycle	The life of the asset, often decades
Limit on learning	Set by the rules, the same for all	Set by budget and practice, unevenly across sectors

An industrial asset runs for decades, under objectives that change, and the question that will matter in year fifteen is rarely visible in year one. Information gathered now has to stay useful across that span, which is where the lesson of Fig. 5.1 applies with most force. A plan tuned only to today's question measures today's question well and leaves little to work with when the next one arrives. Some of the budget has to be kept for questions not yet asked.

What does transfer is the habit beneath the sport's methods: treating every measurement as having a purpose and a price, checking models against reality as a continuous activity rather than a commissioning step, and making the budget for learning explicit, so that it can be argued about and spent well.

7

Already carried over

The movement from the sport into industry is already under way, and what travels best is method.

Electronics and data
Motion Applied
Formerly McLaren Applied, renamed in 2025. It has supplied the standard control unit used by every Formula 1 team since 2008, with a contract extended to 2030, and now applies its electronics and data systems to transport and mining.
Electrification
Williams Advanced Engineering
Acquired by Fortescue in 2022. Its engineers built a 1.4 MWh battery system for a prototype 240-tonne electric mining haul truck, delivered for testing in 2023.
Medical engineering
UCL-Ventura
Engineers from Mercedes-AMG High Performance Powertrains, UCL and UCLH clinicians took a breathing aid from first meeting to first production device in under 100 hours in 2020. A refined version used up to 70% less oxygen.

In each case what moved was less a part than a way of working: rapid iteration against measurement, tight tolerances held under pressure, and teams that put the people who model a system next to the people who build and test it.

8

What we take from it

The production of knowledge can be engineered, with its own costs, constraints and performance targets.

Formula 1 shows what that looks like when the constraints are severe and explicit. Every measurement has a purpose. Models are checked against reality continuously. Experiments are chosen for what they are expected to teach. Those are the habits UNDAMENTAL brings to how organisations observe their physical systems, with one emphasis that a longer horizon demands: keeping part of the budget for the questions that have not yet been asked, because an industrial asset will be asked many of them over its life.

The idea is also testable on a small scale, without access to a racing team.

Hypothesis Proposed, not yet run

The system. A well-instrumented thermal-fluid rig, such as a heat exchanger, recorded densely enough to reconstruct its behaviour across operating conditions and introduced faults. That record is the reference.

The constraint. The measurement budget is then cut, and strategies for choosing what to keep or acquire are compared: conventional sensor selection and state estimation, and acquisition designed for questions asked later.

The measures. How well the hidden state is estimated. How early and reliably faults are detected. The net value of the decisions the data supports, after the cost of acquiring it.

The test. Operating regimes held out entirely from design, so that the comparison rewards what carries over to conditions nobody planned for. If the approach holds there, it is a candidate for a bioreactor, a battery pack or a refrigeration network.

9

Sources

Chaloner, K. and Verdinelli, I. Bayesian experimental design: a review. Statistical Science 10(3), 273–304 (1995). DOI ↗
FIA. 2026 Formula 1 Regulations, Section F: Operational, Article F4, Aerodynamic Testing Restrictions, including F4.5.2 on power-unit heat exchangers. Issue 11 (1 October 2026). PDF ↗
Formula 1. Explained: what is the F1 cost cap and why has it gone up? (10 April 2026). Link ↗
Howard, R. A. Information value theory. IEEE Transactions on Systems Science and Cybernetics 2(1), 22–26 (1966). DOI ↗
Kalman, R. E. A new approach to linear filtering and prediction problems. Journal of Basic Engineering 82(1), 35–45 (1960). DOI ↗
Kammer, D. C. Sensor placement for on-orbit modal identification and correlation of large space structures. Journal of Guidance, Control, and Dynamics 14(2), 251–259 (1991). DOI ↗
Kennedy, M. C. and O'Hagan, A. Bayesian calibration of computer models. Journal of the Royal Statistical Society, Series B 63(3), 425–464 (2001). DOI ↗
Krause, A., Singh, A. and Guestrin, C. Near-optimal sensor placements in Gaussian processes: theory, efficient algorithms and empirical studies. Journal of Machine Learning Research 9, 235–284 (2008). Link ↗
Mercedes-AMG PETRONAS F1 Team. Data and electronics in F1, explained. Link ↗
Motion Applied. McLaren Applied becomes Motion Applied (2025). Link ↗
New Atlas. Fortescue looks toward greener mining with 240-tonne electric truck (16 January 2023). Link ↗
PMW Magazine. FIA extends contract for McLaren Applied to supply F1 standard ECU. Link ↗
Rasmussen, C. E. and Williams, C. K. I. Gaussian Processes for Machine Learning. MIT Press (2006). Link ↗
Srinivas, N., Krause, A., Kakade, S. M. and Seeger, M. Gaussian process optimization in the bandit setting: no regret and experimental design. Proceedings of the 27th International Conference on Machine Learning (2010). arXiv:0912.3995. arXiv ↗
UCL News. UCL, UCLH and Formula One develop life-saving breathing aids (17 April 2020). Link ↗

The worlds in Fig. 5.1 are invented, and its results are UNDAMENTAL's own simulation. The reading of the three sensor categories in section 2 is ours.

← All research

What Formula 1 knows about measurement

A field that budgets its learning

Three reasons to measure

So the car can run

So the team can learn

So the car survives