How accurate is our "cheap" CFD ?

[Uwe]

How accurate is our "cheap" CFD ,I am talking about ours personal computer power..not super comupters from F1,Boeing,NASA,American cup,car industry etc...
Can we rely on this data and how much men who working with CFD can make wrong job?
Does results form CFD depend on men knowledge about aerodynamics and CFD-job or this is just depend at softwear and computer power?

[Dynamicflow]

It all really depends on what you are trying to do, there is no right or wrong answer!

[wesley123]

The tool only is as good as the person using it. If you use it and have no clue what you are actually doing, it won't yield desirable results.

I think a home computer(albeit a high-end one) is very much capable of doing CFD, and providing correct results. However, you'll have to do concessions on the detail of the model.

[Zynerji]

The tool only is as good as the person using it. If you use it and have no clue what you are actually doing, it won't yield desirable results.

I think a home computer(albeit a high-end one) is very much capable of doing CFD, and providing correct results. However, you'll have to do concessions on the detail of the model.

Workstations now for serious computational horsepower can be had at home for <$8000. That is not outrageous for someone that would actually use that much processor/RAM/GPU. Home CFD seems like a great use for said workstation... Aren't the MVRC guys doing this now???

[LM10]

Talking about CFD at home, a bit offtopic: How is the FIA managing the ban on CFD work for the 2022 car for instance? What if engineers simply start their home computer and do some CFD?

[Zynerji]

Talking about CFD at home, a bit offtopic: How is the FIA managing the ban on CFD work for the 2022 car for instance? What if engineers simply start their home computer and do some CFD?

I'm absolutely sure there is some of this going on, and if F1/FIA doesn't know, it is ridiculous. Especially with blockchain tech. A team could have given each engineer a 64core Threadripper with 512GB RAM, multiple GPU desktop to take home, while having a distributed blockchain CFD cruncher running in the background. Get 300-400 of these rigs up and running, and you could be doing some serious crunching in a 100% invisible, un-auditable way...

[nzjrs]

Isn't it more correct to ask how patient one is?

[Uwe]

Talking about CFD at home, a bit offtopic: How is the FIA managing the ban on CFD work for the 2022 car for instance? What if engineers simply start their home computer and do some CFD?

Obviusly they cant do good job with PC ..

[Hoffman900]

Like any model, it comes down to correlation. CFD is cheaper than the instruments and testing it takes to collect data to correlate it to the real world. Even with the resources of F1 teams and airplane manufacturers, they still struggle with this.

I do engine simulation work. The simulation is a lot cheaper to do than building the engine and running in cylinder / port pressure sensors to help develop the burn model, which we have done.

In both cases, you need to have real world measurements to know how close your model is, and then there could be fault in those measurements, and when it is, you then need it to calibrate your model.

It does provide a good starting point, and makes R&D cheaper, but at the end of the day, you still need to build something and do the testing.

From a Boeing friend, I don't think they would put much stock into low end CFDs on desktop computers, except for some very course measurements (maybe). That said, airplane aerodynamicists have to work with a greater factor of safety and confidence than even F1 designers.

[jjn9128]

Heck anyone can buy time on an HPC relatively cheaply nowadays, don't even have to simulate it on your home PC. Really depends how accurate you want to go, the sort of mesh you can run now on a home PC in a day is within 1% of real world data. Or you can run on a Chinese HPC with 80 billion cells and get within 0.5%

[Zynerji]

Heck anyone can buy time on an HPC relatively cheaply nowadays, don't even have to simulate it on your home PC. Really depends how accurate you want to go, the sort of mesh you can run now on a home PC in a day is within 1% of real world data. Or you can run on a Chinese HPC with 80 billion cells and get within 0.5%

You would estimate a good, solid, 2019+ Desktop PC on a 24hr crunch to be within 1% correlation? I'd say I'm very impressed if that is the case!

[Hoffman900]

Heck anyone can buy time on an HPC relatively cheaply nowadays, don't even have to simulate it on your home PC. Really depends how accurate you want to go, the sort of mesh you can run now on a home PC in a day is within 1% of real world data. Or you can run on a Chinese HPC with 80 billion cells and get within 0.5%

You would estimate a good, solid, 2019+ Desktop PC on a 24hr crunch to be within 1% correlation? I'd say I'm very impressed if that is the case!

That seems optimistic. 1% is for what kind of design? A glider or a F1 car? How do you know what 1% is if you have nothing to go off of?

From an engineering side, just throwing a general value of 1% out there is kind of reckless without qualifying it.

[Greg Locock]

Couple of things. Mr Edgar is perhaps not a great theortician in the world of aero. His most famous article in Silicon Chip was on how to run a garage sale.

Cost of correlation data is a brilliant question which I've never thought about. My off topic response, as it is vehicle dynamics, suggests we are talking about $86000 per vehicle. This is one reason why 'they' want to get rid of prototypes. Admittedly most of that would have to be done anyway if we were just developing the car directly rather than as a sim.

[Hoffman900]

Couple of things. Mr Edgar is perhaps not a great theortician in the world of aero. His most famous article in Silicon Chip was on how to run a garage sale.

Cost of correlation data is a brilliant question which I've never thought about. My off topic response, as it is vehicle dynamics, suggests we are talking about $86000 per vehicle. This is one reason why 'they' want to get rid of prototypes. Admittedly most of that would have to be done anyway if we were just developing the car directly rather than as a sim.

It adds up. To even do 1D simulation right (like within 1%) , you need to have a test engine and equipment to record and tweak the burn models, dyno data. A flowbench that can flow up to 100 inH2O can be used to help develop the Cd tables, use Spintron data to measure valvetrain flex and changes with rpm, as well as use it to develop the friction models, etc. All that alone could be $60-80k in work, and you still have to go back and forth... say you changed piston ring design... that may require going back and measuring the change to tweak your friction model, etc.

I have one now that is about 5%, but we have dyno data, flowbench data, in cylinder pressure data, etc. The shape of the curve is spot on though, so not too worried about exact numbers (quality of real world measurements), and just looking at overall trends and changes to the curve. If you take the cost of development of the engine to the point we had quality enough data, you’re almost in $40k. And this is for an amateur effort with a vintage race application, and people directly involved in this aren’t even charging for their time.

I could start from scratch without any of this and produce a curve that looks good and is close-ish (10% or so), but I would have no idea where I was at without real world data.

For aero, look at the resources Ferrari has and they still struggle with it.

[Uwe]

Heck anyone can buy time on an HPC relatively cheaply nowadays, don't even have to simulate it on your home PC. Really depends how accurate you want to go, the sort of mesh you can run now on a home PC in a day is within 1% of real world data. Or you can run on a Chinese HPC with 80 billion cells and get within 0.5%

For fully attached flow maybe, for separated turbulent flow forget about it..not even close