Moto2 is typical of many contemporary circuit-racing categories in eliminating chassis, engine and tyre development, and imposing as spec the power unit, fuel, oil and rubber. As such, the engineering challenge revolves around vehicle setup. But is there nevertheless a place for artificial intelligence (AI) in this cost- and even software-restricted world, the likes of which are today found across racing and rallying on two, three and four wheels?
Rory Skinner aboard American Racing’s Moto2 Kalex-Triumph at the 2023 Grand Prix of India (All images courtesy of Alex Farinelli & Jesus Robledo)
An answer comes from the experience of the OnlyFans American Racing Team (‘American’), which fields two bikes in Moto2. This year, the team is running American Joe Roberts and Spaniard Marcos Ramirez, who last year replaced American Sean Dylan Kelly from round 10. British rider Rory Skinner was contracted for all 20 races in 2023, but he missed three due to injury.
Brian Kelleher was one of the American data engineers. He explains that unlike MotoGP, Moto2 teams only have one bike per rider. American has two mechanics per bike, and each rider has one crew chief and one data engineer. (Kelleher was Skinner’s data engineer in 2023, and since this article was researched, for the 2024 season he has switched to the Red Bull KTM Ajo Moto2 team’s engineering squad).
Rory Skinner on the bike and Brian Kelleher on the laptop at Valencia’s final round of the 2023 Moto2 series
In addition, American has one tyre guy for both riders and he also drives the truck. Kelleher says: “Our team is probably a little bit different from most in Moto2 because we also have a technical director and a rider coach.” As we shall see, the role of technical director revolves around the introduction of AI.
The American bike
American, like most top Moto2 teams, uses the German-made Kalex chassis. The spec tyre is now a Pirelli; last year it was a Dunlop. Since 2019, the Moto2 engine has been a 765 cc Triumph I3, replacing a 600 cc Honda I4. All of the Moto2 engines are identical, and they are built, serviced and maintained by Spanish company Externpro.
As described in RET 143 (November/ December 2022), the Moto2 engine shares the 78 mm bore and 53.38 mm stroke of the Triumph road bike, from which it is derived. Indeed, the mechanical parts are largely standard, although the engine’s performance has been improved relative to the road bike. This is primarily through increased port sizes, optimised for higher revs and greater power.
The maximum rpm of the Triumph Moto2 engine was 14,000 from 2019-22, and it was raised to 14,400 for 2023. Increasing the rpm limit did have the effect of reducing the amount of time the engine is on the rev limit (at the same time, the limiter was made a little softer). The change for 2023 brought new engine components such as pistons, con rods, crankshaft, valves and springs to coincide with the new 2023 Street Triple engine that the Moto2 engine is derived from.
Kelleher observes: “Compared to the 14,000 rpm engine, it has definitely lost torque from 7000-9000 rpm, and then it’s almost the same power and torque from 9000 to, let’s say, 11,000 rpm. With the newer engine, the riders noticed the engine had lost a bit in mid-range, whereas they could feel it had gained a bit at the top end. It’s about a five brake-horsepower (bhp) peak power increase, so there is more lap time from it.”
It is understood that the current maximum power is a whisker over 140 bhp. Kelleher says: “With the 14,400 rpm engine, the fuel consumption went up maybe 5-7%. Some of that is the power – you can’t get five bhp effectively for free – and some the extra rpm. But another factor is that the AFR [air-fuel ratio] target from 2019 to 2022 was 13:1, whereas it has gone to 12.6:1. No explanation was given as to why.
“My guess is either it is to protect the exhaust or exhaust valves from massively high temperatures. Or maybe it was to do with combustion stability, because with the new 2023 engine they pushed the compression ratio up. It could also be to compensate for a single lambda giving a leaner value at the collector compared to if it was at the primary.”
In terms of the integral gearbox, Kelleher reports: “The first gear on the Triumph road bike is quite short, and then to put it into the Moto2 bike, it was way too short, so they learned from that. “But from 2022, while lengthening, first they shortened fifth and sixth gears. That meant the rpm window in which you use fifth became incredibly narrow. So, you went from fourth into fifth then straight into sixth very, very fast. That was initially a problem, but we got the knack of it.
“Then, when the engine speed went up by 400 rpm, the challenge became matching fourth, fifth, sixth around a lap – where you need to shift while not sacrificing your second gear and your third gear. That created its own set of problems, albeit gearing related.” Power is fed via a slipper clutch provided by FCC. Kelleher says: “We have the option of many different friction clutch plates and four or five different stiffnesses of springs.”
He says the clutch setup can affect race strategy, adding: “What can happen – and it’s just the nature of slipper clutches – is that if you have a very free engine-brake strategy, it’s hard to engage the slipper clutch. Some riders like a very free, two-stroke type engine brake versus a strong, four-stroke type, so you have to nail where your clutch needs to be with that type of rider. Once you find that setting, it becomes fixed.”
The Moto2 engine/gearbox is run by a spec Marelli electronic control unit (ECU). This operates a single injector per cylinder (larger than the production spec), a distributor-less ignition system and a fly-by-wire throttle, and it ensures the quality of the manually instigated gearshift to prevent mis-shifts and over-revs. However, the ECU’s capability to provide traction and wheelie control is not employed in Moto2.
Mapping
This ECU is derived from Marelli MotoGP hardware with more restricted software. Moto2 mapping for fuelling, ignition and throttle control is through spec Marelli Sysma software. Kelleher notes: “You can’t get away from it, and what we can change [via calibration] is maybe 5% or 10% of what you could change in a fully open ECU. The other 90%-plus, that’s locked out from us.
“While we have the ability to change the fuelling, that is not per cylinder. It’s a general compensation. So, let’s say at 10,000 rpm and at a given throttle – 2%, 5%, 10% or 100% throttle – we can add 2% of fuel or 3%. It’s just a percentage offset that we can add or take away.”
American Racing 2023 data engineer Brian Kelleher
While he does not see the background fuel map, he says: “We do have the visibility of the ignition map. We can’t advance ignition; we can only retard it, and again that’s done with an offset. So, let’s say you put in minus 10 at a given site. You apply it in Excel. Then you’ll see what your real ignition map is.
“I think that not seeing the fuel map can cause problems. You could, let’s say, at 80% throttle add 3% of fuel. That could be more fuel than a full shot, but you don’t know that. It has been requested for the fuelling to be more visible. “I do have simulation software I have developed for the fuelling. If you make a fuelling change, you want to see what the AFR will do in advance, rather than just hoping for the best. Sometimes, if you just map one site, what you could have done is mess something up somewhere else in the lap,” he points out.
“Before coming to Moto2, I worked for Triumph as a senior calibration engineer on its road bikes. As such, I was putting all the numbers into the ECU with full control, end to end, from fuelling to emissions and also traction control. That’s around 15,000 parameters. There were also the strategies of how the ECU works; for example, how does the bike start cold?
“The biggest difference compared to Moto2 is that it does feel now like my hands are extremely tied. I have done three years in Moto2, and I feel the restrictions are frustrating. There is stuff that I want to change, but I can’t.” When it comes to mapping such that throttle position correlates to a certain torque value, Kelleher reports: “The options we have are 20 maps that were supplied by Marelli. We pick one of those for each gear. You can put one of those maps into all six gears, or you can have a different throttle map for each gear.
“The fact is, everything seems to be a compromise. What one rider likes often another won’t; what’s responsive for one rider is switchy or snatchy for another. But it should never even cross the rider’s mind that the throttle map is a problem. They should be able to concentrate on the riding; the throttle map is the last thing they need to be thinking about.”
In the American Racing garage at the 2023 Grand Prix of India. The four people working on the bike are, from left to right: Joaquin Pardo Parcel (MR24 front mechanic), David Sanchez Malagon (RS33 rear mechanic), Denis Sarazin (tyre guy) and Tsutomu Matsugano (MR24 rear mechanic). In the background are, from left to right: Julien Robert (MR24 crew chief), Brian Kelleher and Eitan Butbul (team co-owner)
In Kelleher’s opinion, “if engine control was a little more open, that would be a big step. The people who are in MotoGP won’t be there forever. Currently, the gap from Moto2 to MotoGP is a large leap engineering-wise for data calibration, strategy, whatever. So how do the newcomers learn how to do the more complicated things?”
On the other hand, when it comes to adding additional functionality such as traction control or even wheelie control, Kelleher doesn’t advocate it for Moto2. “It might be too much work for some people. We might have some teams having to get an additional engineer. Then it’s a case of who has the best engineer to do that stuff.”
Data analysis
Data is, of course, collected from the operation of the ECU, including throttle opening, and from the engine speed, engine temperature and other power-unit sensors that input to the mapping. Data from all the Moto2 engines is collected at each meeting after practice, qualifying and the race by Externpro and Marelli to check that no one has excessive water temperature, is overrevving or mis-shifting, and so forth.
Data from each bike is also downloaded by its data engineer. In addition to the power-unit data, there are chassis sensors measuring wheel speeds, and suspension and brake operation. In addition, the bike carries a six-axis Inertial Measurement Unit (IMU), which is a combination of a three-axis accelerometer and three-axis gyroscope. These measure what are known as the ‘six degrees of freedom’: roll, pitch, yaw, thrust, heave and sway. Kelleher notes the key here is the lean angle, “showing the position of the rider into and out of a corner – how they pick up the bike and so forth. This will be discussed with the rider.”
He reports that for data analysis, American uses the Marelli Wintax4 software. Having worked in the development of road-going Triumphs and their electronics, he observes: “That was the same challenge. The rider comes in and they have X problem, and we have to look at it, figure out the solution and send them back out.
“It’s all just looking at data, to be honest. I think people sometimes feel that if you’ve come from an OEM to go racing, you’ve done less than them. But, working at Triumph, I was looking at data sometimes 50 hours a week, testing 50 weeks a year, so you’re actually looking at data way more than anyone who is in racing.”
Kelleher explains that his job at a Moto2 race meeting involves “an overview of everything”. In terms of key setup parameters, he says: “I think the most important part is the rider. Nowadays, 95% of the grid is on a Kalex bike and everyone has a ballpark setting that’s working for them. It’s not like MotoGP, which is much more of an engineering exercise.”
He notes that American’s rider coach, John Hopkins, goes out during each session and watches all the Moto2 riders. “He is an ex-MotoGP rider, so he’s extremely good at absorbing what the riders are doing; their lines, and how they open the throttle or brake. “So, he will be watching and videoing our riders, and the other riders. Then he will talk our riders through the differences, which does help them a lot, because it’s always difficult to tell them how to change their riding, but it’s easy for them to see what John means.
“His comments are invaluable to our riders, but John doesn’t get involved much in data analysis. We, the data engineers and crew chiefs, will be scrolling through the data, looking at a lean angle [and asking]: is it too much, are they hanging off the bike too much or not leaning enough? That discussion is a regular thing – and the same with how they are braking or getting on the throttle,” says Kelleher.
“If you’re comparing two laps you can see where and how much time a rider is losing. But also, how they are braking; is it beneficial to maybe brake a little earlier, or have more, or even less apex speed so that you can pick the bike up sooner and get the throttle on sooner? “I’d say, for me, it’s 60-70% making changes to engine-brake strategy. And then I’ll get to fuelling adjustment and the chassis setup, and then the last part would be the riding.”
Engine braking
Before a 2023 race weekend, Kelleher says he always looked back at his notes on engine brake from the previous year. “Sometimes, on tracks that are low grip, I will have to open up the engine brake; make it more free. “I can do simulation or the engine brake; calculating rear-wheel torque in Nm or wheel force and also what the throttle is doing, because sometimes if you make a smooth torque curve for engine brake, the corresponding throttle operation is not smooth and the rider will comment on that.”
The engine brake can be closed- or open-loop. Open-loop means the only value used is the negative torque of the engine – the lower it is, the stronger the engine brake. It follows that open-loop must be characterised according to the individual rider, whereas closed-loop adds or takes away the throttle, based on a pre-defined, measured slip target.
Kelleher explains: “Let’s say you want negative 10% slip during engine brake. If the bike was actually doing 15%, closed-loop would open the throttle to try to compensate for that. You do need to be able to simulate that though, to at least make a judgement call.”
Comparing closed- and open-loop, he says: “I’ve tried both. There are downsides and positives to both. The big gain with the closed-loop strategy would be if you know what your target really is, and if your actual engine-brake map is not where you need to be, at least closed-loop might make up the difference.
“Likewise, if the tyre is going off, closed-loop will try to keep engine brake at least consistent, especially over a race distance. There does seem to be more consistency with closed-loop, as a whole, but there are also inconsistencies. If the rear wheel comes off the ground, even just momentarily, closed-loop will whack open the throttle, and when it lands again, the throttle might be too open, so you do have to take the good with the bad.”
It’s worth noting that the way the rider uses the front and rear brakes affects weight transfer, and how the engine brake is going to work. As for the actual engine brake strategy, Kelleher says: “There are three options: there’s no fuel cut or there is one-cylinder fuel cut or two-cylinder fuel cut.
“We do only have one Lambda sensor in the exhaust collector. That’s set by the rules. It would probably be beneficial to have a second Lambda close to the exhaust ports. For any engine, mapping that location gives the best representation of what just happened in the cylinder’s combustion event. There’s quite a long delay down to the collector. It would be good to have three extra Lambdas – one at each exhaust port – to be able to map cylinders individually. The reason they shy away from that is the cost,” he points out.
“The three engine-brake fuelling options; in theory, if you change each setting, it shouldn’t affect the engine brake if the corresponding torque map is correct. The way Marelli has done it, they have one torque map with no fuel cut, and then they have one torque map with all three cylinders cut, which they call a friction map.
“My guess is they then make an interpolation; for example, what a one-cylinder fuel cut would give. So, if we turn off only one cylinder, your desired engine brake should be the same as if all are running. From any testing we’ve done, it’s close, but it’s not the same.
The Moto2 bike is a highly dynamic machine. This is Rory Skinner at the 2023 Texas race
“The downside of no fuel cut is if you get to the stage where your throttle is shut, you’re not going to get any more engine brake, because all cylinders are running. That is probably the most free of the three options. “But the question I have is: shouldn’t you alternate the cut cylinders to keep them all the same temperature [which is not an option we have]?
“The other question is: you have to restart these cylinders, but the port walls on whatever cylinder you’ve cut will have one well and truly dry… if you can imagine that for 30 or 40 revolutions of the engine, you should maybe have 40% extra fuel to compensate for the dry port walls. The same could be said regarding the quick shift strategy on upshifts.
“Looking at the data, you don’t know. This stuff is hidden from us. But, for sure, from what I’ve seen, if you cut one cylinder or two cylinders, the restart of the cylinder does seem to be different. But, again, it’s difficult to tell off a single Lambda in the collector.”
Artificial intelligence
Kelleher notes that the role of technical director was new for American because of a fresh partnership with Dell’s BOSS (Boot Optimised Storage Solution) AI operation. BOSS AI presents itself as a “specialist in the artificial intelligence and machine learning industry, with a focus on helping organisations get more from their data”. It reports that it has worked with the US Government, and in health, finance and other major industry sectors.
“The goal with BOSS AI is to create software to help with chassis setup and calibration, but these guys had no prior involvement with motorbikes at all,” remarks Kelleher. “Matthieu Grodecoeur came onboard as our technical director to deal with them; also, to help on both sides of the team wherever he can.”
Rory Skinner on the grid with three American Racing crew members, from left to right: Brian Kelleher, Stuart Shenton (RS33 crew chief) and Ignasi Casamitjana (RS33 front mechanic)
In terms of the ultimate aim of BOSS AI’s involvement, Kelleher says: “This is to know that for a certain track, a certain rider, if we have this setting and we’re doing this lap time, if we then make that change, it will make this difference to the lap time. That’s the end goal, really. “When BOSS AI came onboard they looked at what we do. The way I look at the data, with fuelling or engine brake, that’s very manual, and they said to me, ‘AI could do what you’re doing a lot faster’.” When asked if that is a bit worrying, in terms of his future role, Kelleher replies: “I suppose that is just the way everything is going.”
Joining the data
For his part, Grodecoeur remarks: “We are working with BOSS AI as part of a partnership with Dell. Our team is very new to this. We started by sending them the previous three seasons’ data from [Marelli] Wintax – all the channels. The big job for us was to put the setup info with it.
“They were given the bike setup, the weather conditions and the data from the bike, and with all this information they could tell us, ‘Okay, this is the condition we have right now’ and then give us the probability of the best setup. “That was based on past data. We then started to use current data, but it’s a very slow process because they need a lot of data points to be accurate, so we have been building this in simple stages.
Marcos Ramirez on track for American Racing at Valencia’s 2023 season-closing Moto2 race
“For example, the front suspension position is linked to the front spring and the preload, and so forth. They call their analysis machine learning. They are able to much more easily point fingers to things we will not see ourselves because there is so much data.
“The way I understand it, in MotoGP, Ducati is working such that they are able to come to the track with a predefined setup and it will be more or less right from lap one.” Grodecoeur admits this is American’s goal, “but, as we don’t have so much resources, it’s going to take a long time. Also, BOSS AI is in the USA and we are in Europe, so the communication is sometimes difficult, but this is what we are working towards.
“At first, in 2022, we were able to do a nice project predicting tyre pressure, because in Moto2 your minimum pressure is 1.4 bar. Depending on the track conditions, and using past season data and so on, most of the time we had a tendency whereby we overshot that. “So the idea was to see how we could make sure we got 1.4 and not too much on top of that. BOSS AI were able to quite precisely define the pressure we should start with to have the running pressure at 1.4. That was good. And it was quite easy for them,” he says.
“That tyre-pressure work was very promising. Then, reaching a collaboration on a larger-scale opened doors to solutions for all other aspects of bike setup. With the motorcycles in Moto2 being so similar, and just 0.1 of a second making all the difference in this class, this is an enormous opportunity to improve our overall package and help our riders reach their full potential.”
Grodecoeur observes that American has pioneered the use of AI in Moto2. “Dell is a partner of McLaren in Formula One, and this is exactly what they’re doing on a much larger scale in Formula One,” he says. “Having this relationship with Dell, they can explain what they’re doing in Formula One. That’s pretty nice for a small team like ours.”