The suspension of future performance cars might be aided by artificial intelligence
We often hear damper engineering talked up as a bit of a dark art, and there’s plenty of truth to that.
This is one corner of the car industry that demands not just mathematical rigour but also world-class seat-of-pants feel from its practitioners – something of a synaptic je ne sais quoi to weave into all the objective, properly measurable stuff.
Perhaps the only area that exceeds damping in terms of intricacy is tyre development, which remains a wild west.
As road testers, it’s good practice to worm our way into conversation with damping engineers whenever the chance presents. Their ultra-nuanced work is never less than fascinating.
I remember a chat with Polestar suspension whisperer Joakim Rydholm when the 2 was still in the works. He was in the midst of 120 gruelling iterations for the Öhlins dampers, each one of them requiring a merry dance.
Car up on lift, wheels off, dampers off, dampers cleaned, opened up, internals removed, piston out, discs changed, rebuilt and remeasured before going back on the car.
All this for each damper, and all in situ at the remote Hällered proving ground in the forests east of Gothenburg. It was 20 weeks of graft to get the dynamic identity of the car right. He assured me there were no shortcuts: “The human is sensitive and you can’t calculate that.”
Five years later, Rydholm is still correct about the human element but, as with everything else, it seems, the world has moved on when it comes to the speed of process.
Manufacturers now rely more on driver-in-loop simulators and less on hard proving ground yards, expediting development.
Instead of messing around with fiddly shims as the sun is going down, a new bit of code can be programmed and, hey presto, the chap in the rig, surrounded by screens, should now be detecting 2% more low-speed rebound damping force on the front axle or somesuch.
The capabilities of the models that these sims use to replicate real suspension movements have advanced at dizzying speed. It’s why Porsche and BMW now have dedicated facilities.
At present, physical prototypes remain the gold standard for subjective assessment, but some makers are already talking about 100% sim-developed cars, which would save them million-euro test mules, oceans of fuel, tyres and the upkeep of sprawling facilities.
And the misery of turning up for your costly week at Nardò only to find that it’s raining.
But there’s a bump in the road: the models that inform the suspension movements on the sim rig are now so complex that they can’t keep up with real time. Which is where Kenny Motte comes in.
“A model contains hundreds of Xs, and they all need to be retrieved within the millisecond of time you have before the next simulation step,” and this is a problem, says Motte, who works on Tenneco’s Monroe sub-brand.
(If you’ve been in a McLaren 750S, AMG C63 or Volkswagen ID 4 GTX, you’ve experienced its handiwork.) There are too many variables and not enough time.
For every input that an imaginary damper receives on an imaginary road, the response is determined not just by your programmed valving, the nature of the impact and road speed but also oil character, how the gas is dissolved in the oil, whether the piston is close to full compression or fully extended, flow speed through certain valves, pressure at precise locations inside the compression chamber, temperatures and even the elastic deformation of the damper tube.
Truly, the dependencies are enough to make your brain shrivel. It’s likewise an issue for the modelling computers, whose instructions to the rig begin to lag behind the forward travel of the car that’s being simulated.
Result: unusable data and misleading feel for the driver in the loop.
Motte has thus been busy training an artificial neural network (ANN) that takes inputs into its AI black box and spits out a reaction output, ‘calculating’ a bunch of matrix multiplications to do so.
The process is rather more instantaneous than ‘solving’ for countless interlinked Xs, as the classic models need to do, although one drawback is that you can’t ‘look under the bonnet’ of an ANN to see what’s happening and why.
If speed is the key benefit, the efficacy of the ANN is down to the rigour and quality of the training, which is undertaken with massive sets of known data on how dampers respond.
As you feed it data, the ANN will begin to generate decision-making nodes to match input with output. Monroe’s people are experts at this stuff.
So, dampers: velocity goes in and force comes out, but now being developed with the help of AI.