On 12/3/2024 at 9:19 PM, blingthinger said:

No, no I got you. It sounds like exactly the perfect type of test that somehow still seems to have gone untried. I'm not strictly opposed to getting their opinions, but if that's the only judging metric then there are still gaping holes in the conclusions. There's already opinions a-plenty from well (and not so well) intentioned seasoned pilots out there.

As stated in my test concept, subjective "feel" is one, but not the only judging metric. The others are time required to set up flight model, amount of target parameters (like speeds, RPM, etc. from POH or flight test data) met to 99% and CPU time required per flight loop calculation.

4 hours ago, brinx said:

best approach to take in 2024 if building a new sim

And a question on this point: why are you so eager to discuss AI/ML in your own CFD-focused thread? Seems odd when they are entirely different concepts and tools.

19 minutes ago, Bjoern said:

"feel" is one, but not the only judging metric

I neither thought nor indicated that it was. That's why I only quoted that one "feel" bullet when I commented. The rest of your points would indeed be interesting to observe.

10 hours ago, blingthinger said:

It's not exactly 'only'.

Its "only" in the context of flight simulation the same way glass windows "only" let the light in if you are building an apartment block. there are several other forces and components required for flight simulation with equal detail, like the wheels moving on the ground, hydraulic and electrical systems, all the failures associated with those systems.

And that is before you get onto everything else - the developer tools to work with everything, the scenery - especially airports. The actual hardware integrations required.

If I was pointing to one aspect that makes developing flight simulation extremely hard it wouldnt actually be the fluid simulation that seems to be the hardest part, rather than a hard part - the hardest part is the users lack of a real time operating system, because you are trying to run a simulation on an operating system that fundamentally lacks the technical requirements to actually run that simulation.

This is why aircraft run on kit like

https://www.aviationtoday.com/2013/04/01/real-time-operating-systems/

rather than

(linux and mac dont meet those requirements either, although they are significantly better, and linux did very recently - this year- get RT functionality merged into the kernel space)

Edited December 4, 20241 yr by mSparks

Even if it's better, I simply do not care as flying in MSFS feels so generic. I enjoy every second of flying in XP11/12, while it's completely meh in MSFS20/24. I fly in the most popular sim only when I want to enjoy scenery from up close, It's a scenery simulator for me, sadly or not.

Generic != better in this context. Though the "meh" fits in that former shoe perfectly.

2 hours ago, Pe11e said:

Even if it's better, I simply do not care as flying in MSFS feels so generic. I enjoy every second of flying in XP11/12, while it's completely meh in MSFS20/24. I fly in the most popular sim only when I want to enjoy scenery from up close, It's a scenery simulator for me, sadly or not.

Lots of differing opinions on this from both real-world pilots and simmers alike. For me, it comes down to the developers and the individual planes. At the end of the day, use the sims that makes you happy whether it is msfs, xplane, dcs or any other.

This topic was basically about the differences between CFD and BET, and the role I believe CFD is likely to play in flight simulation going forward.

Things are developing quickly:

Quote

Realtime CFD: the advent of AI/ML and Operational CFD Digital Twins!

Very rarely in one’s career does something happen which is both a paradigm shift and a game changer at the same time, related to the industry that you’ve devoted most of your career to – in my case Computational Fluid Dynamics (CFD). But, recently, I have seen something I thought I would never see in my 30 plus years career - a viable approach to Realtime CFD (and indeed real-time CAE) that delivers repeatable engineering level fidelity suitable for CFD analysts.

 

24 minutes ago, brinx said:

Very rarely in one’s career....

Oh your post was perfect on so many levels. Not only did you quote someone without a source/link (says much) but it's a jackpot for visually showing what I've described before.

Link

That writeup is a product advertisement by the VP of marketing of a software firm. The white papers describing their algorithms are locked behind account creation (not free and easy for public consumption and review) and no surprise, there is no description of any technical aspects (be it CFD or AI/ML) that in that article. That said, the software they are selling is indeed very intriguing.

Here's one of the use cases presented near the end of the article:

3 CFD cases. 1 hr run time for each. The ONLY variation is wind velocity. No details on the accuracy of the CFD, but that's less important than brinx's underlying confusion that AI/ML is the same as CFD. They then use the resulting AI/ML ROM to predict the wind at 6 m/s. Not 10 m/s. Not -4 m/s (reversed in the other direction). Not a solar panel doubled in chord. They pick a value that is very much within the:

"series of bounding 3D CFD simulations were then run and used to train ODYSSEE CAE for this application." (quote from the next use case example using the same techniques)

Why do they pick 6 m/s? Because it's within the bounds of the CFD training set. What happens if you look at 10m/s? The devs of that software would deny all liability. Why? Because you run the risk of the 7-toes-on-the-puppy's-paw scenario. The AI/ML model is not deterministic outside of the training set. It's garbage. As for their "...any weather condition..." statement in that pic:

"We then carried out a series of meteorological simulations for the solar plant, and solar panels oriented at different angles to the vertical on a rotating single axis to represent their typical movement."

Or in other words: it took significantly more than just 3 CFD simulations to build the final ROM.

So you have to train for all possible combinations of wing span, aileron deflection, power setting, etc. etc. etc. There is no database like that suitable for a general purpose consumer flight sim. There might be one some day, decade(s) in the future. But the final 0d ROM is not CFD. CFD is involved for sure; brinx is right on at least one thing here. But you'll run so much of it that you'll end up with a ton of fantastic BET models that will not only run orders of magnitude faster than the resulting AI/ML model, but will leave the GPU open for gfx goodies.

9 hours ago, Pe11e said:

Even if it's better, I simply do not care as flying in MSFS feels so generic. I enjoy every second of flying in XP11/12, while it's completely meh in MSFS20/24. I fly in the most popular sim only when I want to enjoy scenery from up close, It's a scenery simulator for me, sadly or not.

Fully agree, although I am really enjoying every bit of MSFS 2024 too...

Flight dynamics wise, it lags behind XP unless you only fly FBW aircraft, in full mode (i.e. laws not reverting to non-normal modes where asymmetries and other types of protection aren't available).

Strangely, sideslipping on most aircraft feels better, more realistic, i MSFS than in XP (???). A good example are the gliders.

Edited December 11, 20241 yr by jcomm

8 hours ago, blingthinger said:

Why do they pick 6 m/s? Because it's within the bounds of the CFD training set. What happens if you look at 10m/s? The devs of that software would deny all liability. Why? Because you run the risk of the 7-toes-on-the-puppy's-paw scenario. The AI/ML model is not deterministic outside of the training set. It's garbage. As for their "...any weather condition..." statement in that pic:

hmmmm, not exactly.... its really not that simple.

For example:

 

8 hours ago, blingthinger said:

BET models that will not only run orders of magnitude faster than the resulting AI/ML model

Running an AI model (generally meaning markov models of Andrey Markov fame) is insanely fast, like there is nothing that beats them, a full answer set in three or four clock cycles - training an AI model requires insane amounts of compute and electricity.

Claims running them takes a lot of compute is fundamentally hyperbole to justify the high prices of AI SaaS. If it really took the compute many are claiming you would need a data center per user, in reality they are charging $100+ a month to access the compute of little more than a single raspberry pi....

7 hours ago, blingthinger said:

brinx's underlying confusion that AI/ML is the same as CFD

@blingthinger, my second hobby outside of flight simulation is robotics where I'm currently using AI/ML. This can easily be done using services from AWS. AWS even offers the certification for those interested in working in the field. It is impossible for me to confuse AI/ML and CFD.

I believe we both already agreed on the advantages of CFD over BET in an earlier post. Where we are disagreeing are on the following:

1. Do we need a full Realtime CFD simulation for home flight simulation.
2. The timeframe of full CFD simulations in Realtime 

To the first, my answer is "no", I do NOT believe a full on Realtime CFD simulation is required to achieve the fidelity required for home flight simulation. Hybrid solutions are possible here.
To the second point, the Realtime CFD simulations that are being ran right now were impossible not too long ago. As the author pointed out, what he and many others did not anticipate is the role that Artificial Intelligence (AI) and Machine Learning (ML) would play in making this a reality. Numerous experts in the field are saying real-time CFD is becoming viable and starting to happen right now while you are saying this is decades away.

So, this timeframe is where we are going to disagree. In any case, I find the breakthroughs to be exciting stuff.

 

5 minutes ago, brinx said:

Realtime CFD

Again. No. You obviously don't understand. AI/ML is not solving the Navier Stokes equations. Period. End of story. There is no such thing as Realtime CFD. This isn't even CFD 'assisted' by AI/M nor is it a 'hybrid'. CFD does not enter the picture in the real time aspect. At all. Ever. It's an interpolation ROM. Not CFD.

 

7 minutes ago, brinx said:

To the first, my answer is "no",

Excellent! Good to see you finally admit that asoboCFD is a fools errand! 

3 minutes ago, blingthinger said:

Again. No. You obviously don't understand. AI/ML is not solving the Navier Stokes equations.

@blingthinger I am working with AI/ML in some of my hobby projects, you are not teaching me anything new here.  Please feel free to show me where I said AI/ML is solving the Navier Stokes equations. You can't because I made no such claim.

2 hours ago, blingthinger said:

There is no such thing as Realtime CFD. This isn't even CFD 'assisted' by AI/M

But yet:

Quote

Real-time CFD refers to computational fluid dynamics simulations that run in real-time, powered by artificial intelligence like machine learning and deep learning. 

 

10 minutes ago, blingthinger said:

Excellent! Good to see you finally admit that asoboCFD is a fools errand! 

Uh???? 🤔

Again, I am not comparing Asobo's implementation of CFD and xplane's implementation of BET. For me, it comes down to the developers and the individual planes. But if you want to see my conclusion solely on CFD vs BET as it relates to flight simulation going forward, simply read the thread again. 

2 hours ago, mSparks said:

For example:

This example is funny after you realize how often Sabine rains down on the disturbingly-common snake oil mathematics of modern particle physics studies. Maybe they're all using LLMs!

You'll also notice that every single example she provided is an interpolation. They can indeed interpolate very well sometimes. If you train them with enough good-enough data.

 

2 hours ago, mSparks said:

a full answer set in three or four clock cycles

Fair point. Certainly an infinite number of tensor cores on the GPU would do anything pretty quickly. I honestly have no idea which AI/ML model would be needed for a generalized flight model. Computational requirements depend significantly on what details you feed as inputs. Do you only send wing sweep angle? Or do you also give it actual geometry surface definition of the swept wing? If the latter, and if e.g. you have to consider a custom 3D model created by a 3rd party dev, you might be looking at some big matricies which would clock a few more cycles depending on available hardware.

Just now, brinx said:

Please feel free to show me where I said AI/ML is solving the Navier Stokes equations

CFD == Navier-Stokes equations. It's that simple. AI/ML is not CFD. It is not Realtime CFD. It is an interpolation model built from CFD results. It is not realtime CFD. It is not solving the Navier Stokes equations. CFD == Navier-Stokes equations.