Using AI to support vehicle tuning - Grok vs Gemini vs ChatGPT

[jvanzyl]

Right I'll save you the effort of trying. The winner so far in my experience is ChatGPT. It doesn't need coaching to read data, it remembers your conversation. It's faster than the others and provides outputs like this:

✅ Charge-cooler performance did improve
Your post-4 July changes reduced IAT rise under boost by ~1.7 °C, which is:

• Real
• Measurable
• In line with a system that was previously borderline and is now improved

This means:

• The charge-cooler is functioning more effectively post-fix
• It's now holding intake temps to within ~15 °C above ambient, which is a solid result for a front-mounted radiator and ~13 psi boost
• Further improvements (ducting, fan, water capacity) can be pursued, but the system is no longer underperforming

💬 Revised summary of your setup (post-changes)
System                 Status   Notes
Charge cooling         ✅        Functional, modest improvement   IAT delta now within normal range for this hardware
Fueling under boost   ⚠️ Still lean (~6–8 % lean vs. target)   Needs ~8 % more fuel in VE table (or confirm fuel pressure tracking)
Injector headroom     ✅ Safe (≤ 53 % duty)           Siemens 630s have 30–40 % reserve
Ignition timing         🛑 Limited by lean AFR/IAT   Conservative (~10° BTDC), rightly so for now
Boost level         ✅ Stable (~13 psi peak)         No spike or oscillation visible

🔧 Final recommendations (adjusted for updated IAT context)
Fuel table adjustment:
Add +8 % fuel in the 150–190 kPa, 3 000–5 000 rpm cells—your logs consistently show you're lean by that amount. Expect actual lambda to settle around 0.84–0.85 if VE is matched correctly.

Fuel pressure check:
Use a mechanical gauge to confirm 1:1 pressure rise with boost (i.e., 3 bar base → 4.3 bar at 1.3 bar boost). If it doesn't rise, your regulator or pump might be underdelivering despite low injector duty.

Leave timing alone until lambda is correct.
10° BTDC under 190 kPa is safe only because you're lean and at 46 °C IAT. Once mixture is corrected, you'll have margin to add +2–3° safely and gain 10–15 whp.

Optional cooling tweaks:
If you're chasing perfection:

Better shrouding around the charge-cooler radiator

Higher-capacity pump or fan

Bigger heat-exchanger if frontal area allows

If you make those changes and re-log, I can help tune the ignition table safely once lambda and IAT are fully under control. You're very close to a safe and strong tune now.


Coming from a place where this is not my world, it's rather nice to have this kind of feedback and analysis available to you. Obviously the caveat is that this is a "machine analysis" and doesn't replace decades of experience and skill.. but it's certainly fun!

[Alex Knight]

If it's accurate, that's incredible feedback. Wow.

I'm currently using it to create and track a 3-month marathon training plan, and I have to say that it's been pretty bloody good to be fair.

[jvanzyl]

If it's accurate, that's incredible feedback. Wow.

I'm currently using it to create and track a 3-month marathon training plan, and I have to say that it's been pretty bloody good to be fair.

It is! The fueling stuff is really cool, obviously it doesn't know that my charge cooling piping is possibly narrowing somewhere so it's recommendations don't include that.. but yeah. Really cool.

[Gaz2405]

So it has the ability to read logs etc. Assume you've given it your full build spec?

[jvanzyl]

So it has the ability to read logs etc. Assume you've given it your full build spec?

It cannot read the native file format, so you just build a log window with the relative inputs and then export that to CSV format. I find it easier to then convert that csv format to normal excel workbook and give that to chatgpt.
And yes I provided it with the build spec and it looks up the values of the turbo and injectors itself.

It would be good to be able to validate it's assertions of what should be changed with an actual "tuner" to see if it's talking rubbish or not..

[Kurt]

I've also had an overall positive experience using GPT to help with tuning, tho I've found the o3 model to be much better than the default 4o, which tends to hallucinate and get itself into a bit of a hole of incorrect information. I've been feeding it Tunerstudio logs exported to CSV, similar to as mentioned above. It has been extremely useful lately for lazily fine tuning base boost duty cycle values. Just give it some log data, a screenshot of my base duty + boost target table and it's done!

As a side note, it's interesting to see you're running 10° BTDC around 190kpa (presumably at peak torque?), as I'm running 17°, tapering up to 20° by redline . Of course it's all setup dependant, I'm running a small turbo, lower comp, low IAT etc etc...

[jvanzyl]

I've also had an overall positive experience using GPT to help with tuning, tho I've found the o3 model to be much better than the default 4o, which tends to hallucinate and get itself into a bit of a hole of incorrect information. I've been feeding it Tunerstudio logs exported to CSV, similar to as mentioned above. It has been extremely useful lately for lazily fine tuning base boost duty cycle values. Just give it some log data, a screenshot of my base duty + boost target table and it's done!

As a side note, it's interesting to see you're running 10° BTDC around 190kpa (presumably at peak torque?), as I'm running 17°, tapering up to 20° by redline . Of course it's all setup dependant, I'm running a small turbo, lower comp, low IAT etc etc...

Yes I'm looking forward to working on the car using this method more and more.
When I had the car tuned at RRR the emphasis was on that this should be an everyday driver.
He tuned it initially for 205 and then put in a bunch of safety margins and settled round 198 ISH I think.

[Midshipguru]

Love this, got to try it! 

[Midshipguru]

Tried it. Loved it

Chatgpt coached me into overcoming hunting idle successfully. On to more complex layers!  Thanks for inspiring @jvanzyl

[jvanzyl]

Tried it. Loved it

Chatgpt coached me into overcoming hunting idle successfully. On to more complex layers!  Thanks for inspiring @jvanzyl

Good grief have I actually helped someone? !

[jvanzyl]

But in all seriousness please post the details?

[Midshipguru]

Well I gave it the user guide and csv export of the log, it digested these and I asked it the question to solve. It came back with a lot of detail, probably too much, on what steps to try

I'll be going again on the weekend to try the suggested steps relating to engine cutting lifting off coming to a stop which looks a bit more involved.

Very powerful, there is no way I could get there without these tools. But remember, there is no such thing as AI ^^