Predict setup changes before you wrench
Generated from
content/lms/vehicle-dynamics-ii-theory/08-from-theory-to-data-and-setup/02-using-the-bicycle-model-to-predict-setup-changes.md; edit the source file, not this page.
Source path: content/lms/vehicle-dynamics-ii-theory/08-from-theory-to-data-and-setup/02-using-the-bicycle-model-to-predict-setup-changes.md
Course: Read the forces that steer the car
Module: Connect the math to the garage and the track
Estimated duration: 55 minutes
The skill in this lesson is not building a perfect race-car simulation. The skill is walking up to a proposed setup change and predicting the direction of its effect before anyone touches a spanner. You are trying to answer a practical question: if I change this spring, bar, damper, pressure, tire, or geometry setting, should the car ask less of the front axle, more of the rear axle, follow the road better, move its body too much, or become harder to drive when I add power?
That prediction is valuable even when it is approximate. A simple bicycle-style model reduces the car to a front axle and a rear axle, then asks how much cornering work each axle can do for the load it carries. That is a useful lie. It ignores many details, but it makes you state the expected sign of the change. More front cornering capability relative to the rear should move the car away from understeer. Less front capability relative to the rear should move it toward understeer. More rear cornering capability relative to the front should make the car safer at exit. Less rear capability can make the car rotate more easily, sometimes usefully, sometimes too much.
The honest part is that the simple model is not the final judge. The bonded sources are blunt about this. Very large race-car models can become too input-hungry to be practical, while the quickest reliable answers still come from actual road testing. A steady-state model can also ignore damping, response rates, and time-based behavior. Logged data is limited by the sensors fitted, by logger resolution and frequency, by the circuit, by the lap, and by the weather. A four-post rig can measure things the track cannot, but its car-balance result can still be unreliable because static tire behavior, aero-load simulation, and real circuit dependency are not the same as a rolling car on a live track.
So this is a lesson in disciplined prediction, not prediction-as-proof. You use the bicycle model to say what should happen, where it should show up, and what evidence would convince you. Then you test enough to find out whether the car agreed.
The principle: balance is an axle problem before it is a parts problem.
In a corner, the car's oversteer or understeer balance is tied to the load distribution between the front and rear axles. Suspension tuning changes that balance by changing how load is distributed and how the tires are worked. Springs, anti-roll bars, damping, tire pressures, tire sizes, tire compounds, suspension geometry, and mass distribution all matter because they change either the tire loads, the tire behavior at those loads, or the body and wheel motions that decide whether the tire can do its job.
The first mechanism is lateral load transfer. In a steady turn, load moves across the car. If the front axle takes a larger share of the lateral load transfer, the outside front tire carries more load while the inside front contributes less. Because tire grip is nonlinear with load, the total useful front axle grip does not rise in proportion to the load added to the outside tire. The combined front axle can lose effective cornering capability, and the car tends toward understeer. The same logic applies at the rear: if the rear axle is made to give up more effective cornering capability, the car tends toward oversteer.
The second mechanism is tire load fluctuation. Grip is not just an average load number. The suspension must help the tire follow the terrain and avoid big changes in tire load. Softer settings can reduce tire load fluctuation because the tire follows road irregularities with smaller force changes. But the suspension also has to control pitch and roll because body movement can hurt geometry, contact patch quality, and, in some cars, height-sensitive aero behavior. That is the core suspension contradiction: stiff components reduce body movement from driving inputs, while soft components isolate road input better.
The third mechanism is phase. A car can have a tidy steady-state balance on a skidpad and still be poor when you brake, release, turn, or accelerate away. The mechanical setup determines the basic steady-state oversteer or understeer at the limit, but the driver still makes constant steering corrections because the track surface is never perfectly consistent. Braking can upset cornering balance. A car that is balanced at maximum steady-state cornering speed still has to accelerate out of that state without becoming unstable. If you use a steady bicycle model to answer an exit-throttle problem without naming that limitation, you are already outside the model's cleanest territory.
The working model you should carry in your notebook.
Draw the car as two tires: one front axle tire and one rear axle tire. That is the bicycle picture. You are not pretending the inside and outside tires are equal. You are deliberately compressing the four contact patches into axle-level cornering capability so you can reason quickly. The front model tire stands for everything that changes front axle cornering power: front tire load, front load transfer share, front tire compound and size, front camber behavior, front pressure, and front contact-patch stability. The rear model tire stands for the same set of questions at the rear.
For any proposed change, ask four questions in order.
First, does this change move lateral load transfer toward the front axle or toward the rear axle? If it loads the front axle more unevenly in roll, expect the front to give up more effective cornering power and the car to move toward understeer. If it loads the rear axle more unevenly in roll, expect the rear to give up more effective cornering power and the car to move toward oversteer. This is why roll stiffness distribution is such a powerful balance lever.
Second, does this change help the tires maintain load over the real road, or does it increase tire load fluctuation? A setting that improves balance on smooth pavement may cost grip over bumps if it makes tire load fluctuate. A softer setting may let the tire follow the terrain better, but it may also allow body motion that moves camber, caster, ride height, or contact patch state out of the useful window.
Third, does this change improve or degrade body-motion control for the phase you are trying to fix? Pitch movement under braking and roll movement in cornering can hurt suspension geometry and the tire contact patch. But controlling them by simply stiffening the car is not automatically better, because road input and tire load fluctuations also matter. The right prediction is not stiff good or soft good. The right prediction names which problem is dominant in the corner phase you are fixing.
Fourth, what does the model not know? It may not know the current tire parameters. It may not know contact patch load on a rolling tire because there is no ordinary on-track sensor for that. It may not know road actual position accurately. It may not include damper time response. It may not include the aero interaction with suspension movement. When the missing item is central to the claim, the model should demote the change from expected fix to test candidate.
The technique: make a setup prediction card.
Before a setup change, write a short prediction card. Do not write a wish. Write a falsifiable expectation. The card has six lines.
Line one is the symptom and phase. Examples: steady-state mid-corner understeer, entry instability while braking, exit oversteer when accelerating out of a loaded turn, excessive body movement over a rough section, or poor response after a quick steering input. Naming the phase keeps you from using a steady-state tool to solve a transient problem without admitting the uncertainty.
Line two is the proposed lever. Use the actual lever, not a vague handling word. Front anti-roll stiffness, rear anti-roll stiffness, spring rate, damper setting, tire pressure, tire size, tire compound, geometry adjustment, or mass distribution are all different levers. The data-acquisition source groups setup work around roll stiffness distribution, spring rates, rollbar rates, damping, tire pressures, and related adjustments because those are the knobs that influence the balance.
Line three is the axle prediction. Say which axle should gain or lose effective cornering capability. If the front axle should gain relative to the rear, write front effective cornering power up. If the rear should lose relative to the front, write rear effective cornering power down. This line keeps the model honest. If you cannot name the axle-level effect, you do not yet have a prediction.
Line four is the load-fluctuation prediction. Say whether the change should make tire loads calmer or more variable over the section that matters. A softening change may calm tire load on rough pavement. A stiffening change may control body movement but can increase force changes at the tire. A damper change may affect response and fluctuation more than steady-state axle balance. A pressure or tire change may alter the tire parameters themselves. Keep this separate from the axle-balance line.
Line five is the evidence plan. State what would count as support. Driver feel is allowed, but it should be specific: less steering correction needed to hold radius, cleaner acceleration out of the corner, less pitch or roll upsetting the contact patch, or a balance change that appears in the phase you named. Logged evidence is useful when it exists: suspension position, load sensors in key elements, body or hub accelerometers, and derived channels can characterize movement and response. But the evidence plan must also name limitations: no direct rolling contact-patch load channel, logger resolution limits, and lap, circuit, and weather dependency.
Line six is the rejection condition. Decide ahead of time what would make you back out or change the diagnosis. If the car improves steady-state but becomes unstable when accelerating out, the prediction was incomplete. If the rig result says the tire load is calmer but the driver reports worse balance and the circuit data agrees, the rig was not enough for that decision. If weather, tire state, or circuit condition changes enough to move the reference, do not pretend the setup alone caused the result.
Sub-skill one: predict sign before magnitude.
At the intermediate level, your first win is not a perfect number. Your first win is the correct direction. Will the change move the car toward understeer, toward oversteer, toward calmer tire load, toward more body control, or toward a side effect? Magnitude comes later, and the sibling lesson on identifying tire model parameters belongs there. Here, you are building the habit of not wrenching blind.
A sign prediction can still be rigorous. If you stiffen the front roll resistance relative to the rear, the simple axle model expects more front lateral load transfer share and less effective front axle cornering capability at the limit. That predicts more understeer, all else equal. If you stiffen the rear relative to the front, the model expects the rear axle to give up more effective cornering capability and the car to rotate more. That may help a car that will not turn, but it may also make exit power application less stable.
Sub-skill two: separate average balance from fluctuation.
Two cars can have the same average balance and different grip because one keeps the tire load steadier. The suspension is asked to minimize body movements and tire load fluctuations at the same time, even though those goals can pull in different directions. This matters because a setup change that improves the balance number can still make the car worse over bumps, curbs, or surface changes. If you only ask understeer or oversteer, you miss the load-fluctuation question.
When you predict a change, always include both the average balance and the fluctuation effect. A front bar adjustment is not just a front-versus-rear balance change; it can also affect how independently the front tires follow the road. A damper setting is not just a comfort knob; it can change response, tire-load fluctuation, and driver feel. A tire pressure change is not just a grip request; it changes the tire parameter set you are asking the model to use.
Sub-skill three: keep steady-state and transient claims apart.
A steady model is strongest when the car is in a steady turn at the limit. It is weaker when the problem depends on time: brake release, pitch rate, roll rate, damping response, quick steering, road input, or acceleration out of the corner. Carroll Smith's steady-state modeling discussion explicitly leaves out time-based factors such as damping, rates of change, and response. That does not make the model useless. It tells you what kind of claim it can make.
For a steady mid-corner push, the two-axle model is a good first lens. For a car that snaps when you add throttle, the same model may still help, but only after you state that acceleration and transient stability are part of the problem. For a rough braking zone, the model must be paired with a load-fluctuation and body-motion check. For an aero-sensitive car, ride height and body movement may become constraints before the simple balance prediction is accepted.
Sub-skill four: translate driver feel into model language.
You do not need to make the driver sound like an engineer. You do need to translate useful comments into the model. A report that the car will not hold radius unless you add steering is a front-cornering-power problem until proven otherwise. A report that the car rotates the moment you add power is an exit-stability problem, not just a victory over understeer. A report that the car is fine on smooth corners but poor on a rough section points toward tire load fluctuation, body movement, or response rather than pure steady-state balance.
This translation is where instructor discipline matters. Do not ask whether the car was good. Ask where it changed, when it changed, and what the driver had to do to keep it on line. The source on mechanical setup notes that even a finely balanced car needs constant instantaneous steering motions to correct for normal track-surface inconsistencies. A setup prediction should therefore include the steering-correction burden, not just the final corner speed.
Sub-skill five: treat data as evidence, not authority.
Data acquisition helps characterize suspension behavior and build references for setup, but the channels are not the car itself. Some race cars have position and load sensors in key suspension elements. Some put accelerometers on the body and wheel hubs. A four-post rig can provide tire contact patch load and tire deflection information that cannot be recorded on track. Those are powerful tools.
They are also bounded tools. On-circuit logged data is constrained by sensor availability, logging capability, lap, circuit, and weather. A real circuit is needed to optimize suspension settings, but the most important tire contact patch load is not directly measurable on a rolling tire in ordinary track logging. Four-post testing removes lap and circuit dependency, but balance assessment can be unreliable and aero-load simulation is incomplete. The driver who understands these boundaries can use data without worshipping it.
Calibration cues: what improving looks like.
You are improving when your prediction card becomes more specific and less emotional. Early on, you may write that a change should reduce understeer. That is too broad, but it is a start. A better prediction says the change should reduce steady mid-corner understeer by increasing front effective cornering capability relative to the rear, but it may cost exit stability if rear capability becomes the limiting side. A still better prediction names the section, the phase, the expected driver correction, and the evidence channel you will trust.
You are also improving when your wrong predictions are useful. A bad setup result is not a failed lesson if it tells you which assumption was wrong. Maybe the tire parameter assumption was bad because the tires were not in the same operating condition. Maybe the model treated a transient as steady. Maybe the road input dominated the balance change. Maybe the four-post result did not transfer to circuit balance. The purpose of the bicycle model is not to protect your ego. It is to make your assumption visible enough to be corrected.
A good on-track sign is phase-specific. If you predicted better steady-state front grip, the car should need less corrective steering in the same kind of steady corner and should hold the requested radius more easily. If you predicted calmer tire loading over a rough section, the driver should report less skipping, less sudden grip loss, or less need to catch the car, and any available suspension or acceleration data should not contradict that report. If you predicted safer exit balance, the car should accept acceleration out of the corner without becoming unstable.
A weak sign is a result that appears somewhere other than the named phase. If you changed the car for mid-corner understeer and the driver only liked the first steering response, you may have solved a response problem, not a steady balance problem. If lap time improves but the balance complaint remains, the change may have helped another part of the lap. If the car is faster only as tires come in, the tire-temperature state may be masking the setup effect.
Failure modes: how the prediction process goes wrong.
The first failure mode is the skidpad trap. It is tempting to tune for maximum steady cornering speed and declare victory. The source material warns that if skidpad cornering speed were the only requirement, setup would be easy, but the car must accelerate out of that state without becoming unstable. Do not let a neat steady-state result blind you to exit behavior.
The second failure mode is the stiffness reflex. A stiffer setting can reduce body movement from driving inputs, but soft settings isolate road input better and can reduce tire load fluctuation. If you always stiffen when the car feels lazy, you may control body motion while making the tires less able to follow the road. If you always soften when the car lacks grip, you may improve road following while letting pitch or roll damage geometry and contact patch state. The model has to carry both sides of the contradiction.
The third failure mode is the single-lever myth. Roll stiffness distribution is powerful, but suspension tuning is not only anti-roll bars. Springs, damping, tire pressures, tire sizes, compounds, geometry, and mass distribution also change the effective model. If your prediction ignores the tire and geometry state, the axle sign may be right and the result still disappointing.
The fourth failure mode is rig certainty. Four-post data can identify frequency response, body movement amplitude, contact patch load fluctuation, damping rates, elasticity rates, and modal components. That can be excellent for choosing damper settings after circuit testing. But if you use rig output as a direct balance verdict, you are leaning beyond its cleanest use. Static tire behavior differs from rolling tire behavior, aero simulation can be limited, and circuit balance still has to be checked.
The fifth failure mode is sensor blindness in the opposite direction: dismissing data because it is imperfect. The right response to imperfect data is not to ignore it. It is to use the channels that exist, know what they can and cannot say, and combine them with driver feel and repeatable tests. A model without evidence becomes opinion. Evidence without a model becomes noise.
Cross-references inside this module.
Use this lesson before the sibling lesson on identifying tire model parameters from logged data. Here you decide what direction a setup change should move the car and what assumptions matter. There you work on extracting better tire parameters so the model becomes less generic.
Use the sibling lesson on identifying tire parameters without fooling the model when your setup prediction depends heavily on compound, temperature, pressure, load sensitivity, or degradation. The current bond is clear that tire parameters matter to grip and suspension analysis, but this lesson does not teach parameter fitting.
Use the sibling lesson on translating force-moment theory into driver feel when your prediction is correct on paper but your driver comments are vague. You need the driver to describe radius, steering correction, brake upset, throttle acceptance, pitch, roll, and surface sensitivity in terms you can map back to axle behavior.
Use the sibling lesson on retiring the simple model before it lies when the missing physics becomes central: transient damping response, aero-load interaction, large tire-state changes, poor sensor resolution, changing weather, or any situation where the model's assumptions are no longer a harmless simplification.
The takeaway.
Before you wrench, make the car pass through the model. Which axle gains or loses effective cornering power? What happens to tire load fluctuation? What happens to body motion and contact patch quality? Which phase should change? What evidence would prove you right, and what would prove you incomplete? That is how the bicycle model earns its place in the paddock. It does not replace testing. It makes testing sharper.
Worked example: steady skidpad balance that will not drive off the corner
Start with a car that has been tuned around steady-state cornering. On a skidpad or a long constant-radius corner, the balance looks respectable. The driver can bring it to the limit, hold it there, and the car does not show an obvious terminal push. The temptation is to call the balance fixed.
Now add the missing phase. As the driver tries to accelerate out of that same loaded state, the car becomes unstable. The rear will not accept power cleanly, or the driver has to delay throttle because the car asks for a catch instead of a drive. The simple steady result was not false, but it was incomplete.
The prediction card should start by refusing the wrong diagnosis. The symptom is not simply needs more rotation. The symptom is acceptable steady-state balance with poor acceleration-out stability. The phase is exit. The model now has to protect rear effective cornering capability during power application, not just reduce mid-corner understeer.
If the proposed change is to move roll stiffness balance rearward to make the car rotate more, the axle prediction is clear: the rear axle will give up more effective cornering capability relative to the front. That may make the car point better in the steady phase, but it is suspicious for this exit problem because the rear is already the axle that cannot accept the next job. The rejection condition should be written before the change: if the car turns better but accepts throttle later or needs more steering correction at exit, the setup went in the wrong direction for the actual problem.
A more disciplined prediction might ask whether the front can be helped without sacrificing the rear, whether tire state is masking the balance, or whether body movement under acceleration is changing the contact patch. The exact answer depends on the available setup range, but the modeling move is the same: do not optimize the steady circle and forget that the car has to leave it under power.
Worked example: four-post rig result versus circuit balance
A second common situation starts away from the circuit. The car goes on a four-post rig, and the rig result says a softer spring or damper direction reduces tire load fluctuation over the tested road-input frequencies. That is useful information. The rig can supply contact patch load and tire deflection information that ordinary circuit logging cannot directly provide. It can also remove lap, circuit, and weather dependency from that part of the investigation.
The prediction mistake is to treat the rig result as the whole setup decision. The bicycle-style setup card should separate the load-fluctuation prediction from the platform prediction. Softer may help the tire follow road input, so the tire-load fluctuation line can be favorable. But the body-motion line may be unfavorable if the car now pitches more under braking, rolls more in cornering, or moves suspension geometry and contact patch state outside the acceptable range for race driving.
The evidence plan must therefore include a circuit check. On track, you are looking for the section where the rig-supported benefit should appear: rough pavement, repeated bumps, curbing, or a surface that previously caused breakaway. But you are also watching for the side effects: slower response, worse platform control, excessive body movement, or a balance change that the rig could not reliably judge.
A good conclusion might be that the softer direction is useful only within a narrow range. It calms tire loading without letting pitch or roll become excessive. A bad conclusion would be that the rig was wrong because the circuit result was not a simple win. The better lesson is that the rig answered one question well and another question poorly. The model keeps those questions separate so you do not throw away useful measurement or over-trust it.
Common mistakes
Mistake one: changing parts before naming the axle effect. The driver says the car pushes, and the paddock immediately reaches for a bar or pressure change. Good looks different. You first write whether the proposed change should increase or decrease front effective cornering capability relative to the rear, then you decide whether that sign matches the complaint.
Mistake two: using steady-state language for a transient problem. Entry brake upset, quick steering response, damper behavior, and exit power acceptance are time-based problems. A steady model can still guide you, but only if you label the uncertainty. Good looks like this: the card says the change should help the average balance, but the transient response must be verified separately.
Mistake three: treating soft as automatic grip. Softer settings can reduce tire load fluctuation and help the tires follow the terrain, but too much body movement can hurt geometry, contact patch state, and height control. Good looks like a paired prediction: this should calm tire loading over the rough section, and we will reject it if pitch or roll makes the car less consistent.
Mistake four: treating stiff as automatic control. Stiffer settings can reduce body motion from driving inputs, but they can also make road input show up as tire load fluctuation. Good looks like identifying whether the driver complaint is body motion, load fluctuation, or axle balance before increasing stiffness.
Mistake five: trusting a sensor channel without asking what it cannot measure. Logged data is limited by fitted sensors, resolution, frequency, lap, circuit, and weather. Ordinary on-track logging cannot directly measure rolling tire contact patch load. Good looks like using available data as evidence while naming the blind spot.
Mistake six: trusting driver feel without translating it. A vague report that the car feels bad is not yet a model input. Good looks like converting it into phase and correction: the car needs extra steering to hold radius, becomes unstable as throttle is added, skips over surface input, or changes balance as pitch and roll build.
Mistake seven: judging a change by lap time alone. A setup can improve one part of the lap while failing the problem it was meant to solve. Good looks like checking the named phase first, then using lap time as a secondary result.
Mistake eight: making the model responsible for missing tire information. Tire compound, pressure, temperature state, size, and load sensitivity matter. Good looks like labeling tire-state uncertainty instead of pretending the front and rear model tires are fixed constants.
Drill: the three-card setup prediction
At your next event, run this as a three-session drill. You do not need to change the car three times. The purpose is to train prediction discipline before setup discipline.
Before session one, spend ten minutes writing one prediction card for a hypothetical change you are considering. Choose one real complaint from the car, name the phase, name one lever, predict the axle effect, predict the load-fluctuation effect, name the evidence, and name the rejection condition. Then drive the session with the car unchanged and collect the best driver comments and any available logged channels for that phase. The success criterion is not whether the car improved. The success criterion is whether your complaint became specific enough that the prediction card could be tested.
Before session two, revise the card. Keep the same complaint unless session one proved it was the wrong problem. If you are allowed to make a small, safe setup change within your event rules and support level, make only the change on the card. If you are not changing the car, treat the session as a thought test and ask whether the evidence you saw would have supported the change. The success criterion is that every driver comment after the session can be placed into one of three bins: supports the prediction, contradicts the prediction, or belongs to a different phase.
Before session three, write a second card for the most likely side effect. If the first card predicts less understeer, the second card might predict possible exit instability. If the first card predicts calmer tire loading, the second might predict more body movement. If the first card predicts better platform control, the second might predict harsher road input. The success criterion is that you can state both what the change is supposed to fix and what it might break.
After the event, keep only cards that taught you something. A perfect-looking card with no evidence is less valuable than a wrong card that exposed a bad assumption. Over time, this drill builds a personal reference library for the car: which levers move axle balance, which levers change load fluctuation, which driver comments are reliable, and where the simple model needs help.
When to retire the prediction
You retire the simple prediction when the missing physics becomes the main story. If the car is in a steady corner and the change mostly shifts front-to-rear load transfer, the bicycle model is the right first pass. If the problem is dominated by damping response, rate of change, brake-induced pitch, quick steering, road input frequency, aero-height interaction, or changing tire state, the prediction becomes a hypothesis rather than a decision.
You also retire it when the evidence base is too weak. If the logger cannot see the channels you need, if the weather or circuit condition changes the reference, if tire operating condition is different, or if the rig and track are answering different questions, do not force a confident conclusion. Mark the result as unresolved and ask for better evidence.
Retiring the prediction is not a defeat. It is the point of using a model responsibly. A simple model should make simple decisions faster and make complex decisions clearer. When it starts hiding uncertainty instead, hand the problem to deeper tire modeling, better data acquisition, more controlled testing, or a more complete vehicle model.
Author Review
No quiz questions are attached to this lesson.
Sources
| # | Document | Chunk | Pages | Score | Collection |
|---|---|---|---|---|---|
| 1 | Analysis Techniques for Racecar Data Acquisition | 066cee65-8c68-773f-fe62-1ae30116d1ae | 13 | 1 | uio_books_raw_v1 |
| 2 | Analysis Techniques for Racecar Data Acquisition | 5a451d8b-d8c5-232d-375b-036b33989a5f | 16 | 1 | uio_books_raw_v1 |
| 3 | Analysis Techniques for Racecar Data Acquisition | d83ed6c5-d9cf-36a0-f6a1-fbd8a61efeea | 16 | 1 | uio_books_raw_v1 |
| 4 | Race Car Engineering Mechanics Paul Van Valkenburgh | e5ada18a-331b-8f45-54aa-5ac71c5cc184 | 75 | 1 | uio_books_raw_v1 |
| 5 | Racing Chassis and Suspension Design Carroll Smith | 633f50e8-a0d2-ad92-24b2-3350738b0cd1 | 193 | 1 | uio_books_raw_v1 |
| 6 | Analysis Techniques for Racecar Data Acquisition | f725bafe-8b10-b36a-5b91-3395a519319d | 16 | 1 | uio_books_raw_v1 |
| 7 | Analysis Techniques for Racecar Data Acquisition | d18c0afe-a337-709c-36e1-a544a81e704e | 16 | 1 | uio_books_raw_v1 |
| 8 | Analysis Techniques for Racecar Data Acquisition | 65286bf1-c759-e3f0-846e-0bad762d24ed | 19 | 1 | uio_books_raw_v1 |