5 hours ago, blingthinger said:

Didn't you once say that Flight Unlimited used a CFD model yrs ago?

Yes, maybe the first flight sim to use other than a look-up table for its flight models, but of course it was probably something very different from what is commonly interpreted as "CFD" as used by academics and engineers, or as understood by ChatGPT.

Anyway, I've read that when Seamus Blackley left Looking Glass Studios, they've had to trash the flight model and start from scratch for Flight Unlimited II, because his CFD code was spaghetti code: it worked but they didn't know how and why. 🙂

Edited November 30, 20241 yr by Murmur

The main issue of today's desktop flight sim model is garbage in, garbage out.

Level-D simulator runs on old look-up tables similar to old FSX/P3D, but with data from test flight exclusively for making such simulator, it's good enough to input FDR control/wind data and get 99% accurate flight trajectory output compare to same FDR, it's been used for accident/incident analyses for decades.

The only thing I know have similar fidelity you can easily download is Falcon BMS, it got the flight model data from NASA and FBW code from real aircraft.

22 hours ago, brinx said:

Can capture complex flow phenomena like turbulence, wake vortices, and dynamic stall.

A bit of clarification on this part. If the point cloud around the airframe isn't high enough density, none of these phenomena will be captured correctly, if at all. Even the hypothetical 50 million point Cessna model will technically be missing turbulence generation in the final output. Getting that whole turbulence phenomenon bit correct requires the billion-point 3D clouds, so take that chatGPT statement with a grain of salt. Also, for what it's worth in their SDK, 2020/4 ignores the CFD model when it thinks it's at or near stall. No dynamics to be had there.

Now...One bit of fidelity that BET is inherently missing is the propagation of wakes in 3D space. Austin's applying additional models outside of the traditional BET concept to calculate wing wake trajectory and strength. I do recall him recently saying that further improvements are coming there in the next FM update, based on his work with the ALIA VTOL flight test program.

As for the rotating blades mentioned in your first post? Remember that 2020/4 CFD has no rotating blades. There's no way to specify rotor airfoil geometry to feed into the CFD. No surprise there given that even the fixed wings don't have that option either. 

And when I said the GPU would run real-time CFD in ~15 yrs, I totally forgot that we will all probably want gfx rendering too. So tack on another 5-10 yrs to be able to run the CFD model in addition to powering the 8k VR headsets we'll all own.

Edited November 30, 20241 yr by blingthinger

16 hours ago, brinx said:

Where did you get that ridiculous number? With all the software out their capable of CFD simulation, you believe only 20 people understand the subject? In that case, the entire world should unite to protect these 20 people at all costs.

To quote a teacher of mine: "Exaggeration is illustraitive".

Writing a toolbox around the CFD equations isn't much of a problem (hence the abundance of software), validating the results it spits out is, especially for corner cases. And just because you can simulate a C172's aerodynamic behavior within 1 % of actual flight test data under perfect ISA conditions does not mean that you can automatically use it to simulate each and every other aircraft or each and every atmospheric condition and still obtain the same accuracy compared to real world data. As soon as you move outside the boundaries of validated cases, the obtained results are a liability when not thoroughly scrutinized and simply used for further development. But that's for actual engineering. Nobody can die from a subpar CFD application in a desktop flight simulator, so developers can do what they want. Unless they want their software certified for real world training, at which point the regulatory authorities apply standards regarding validation and code quality and which is when "thoroughly understanding CFD" comes in.

 

2 hours ago, C2615 said:

The main issue of today's desktop flight sim model is garbage in, garbage out.

Yes.

27 minutes ago, blingthinger said:

And when I said the GPU would run real-time CFD in ~15 yrs, I totally forgot that we will all probably want gfx rendering too. So tack on another 5-10 yrs to be able to run the CFD model in addition to powering the 8k VR headsets we'll all own.

We don't even have CFD toolboxes running reliably on present day CPUs with acceptable speed nowadays, so "real time CFD on GPUs in 15 years" is the "fusion reactors in the next 50 years" of computing to me.

I mean heck, the software I have to use at work is written in FORTRAN and still looks and works as if I had taken a time machine back into the late 90s!

Edited November 30, 20241 yr by Bjoern

All I know is X-Plane feels right and, to me at least, MSFS feels weird. As did P3D and FSX, which is why I preferred older versions of X-Plane to those too.

43 minutes ago, Bjoern said:

We don't even have CFD toolboxes running reliably on present day CPUs with acceptable speed nowadays, so "real time CFD on GPUs in 15 years" is the "fusion reactors in the next 50 years" of computing to me.

I mean heck, the software I have to use at work is written in FORTRAN and still looks and works as if I had taken a time machine back into the late 90s!

heh heh heh. I may be super optimistic there yes. But there's no doubt that GPUs have changed the equation(s) (pun intended). Sometimes it's an order of magnitude faster to converge a solution, and we are indeed still talking about multiple GPUs to get there. Many 4090ish GPUs in the case of those big billion-cell meshes. So 0.5 hr instead of 5 hrs. 30-60hz is few orders more from that. There are a few tricks to play too. E.g. you aren't needing to converge the full way down at each time step because the previous time step is reused as the initial condition for the next step. That's an order's worth of time reduction at the least. AI/ML might also help a bit but still plenty of ?? there. I'm also assuming Moore's Law doesn't flatten out too bad in the meantime.

Fortran? 🙂 You'd be surprised (or not) at how many modern research labs still write new CFD code in that one. Thankfully nobody's still stuck on F77, but those droppings are still often littered around for good measure.

13 hours ago, blingthinger said:

At each operating point you'd sum up the forces at all the points on the airframe surface

Thats a very, very oversimplified explanation - imho, too much so.

Because those forces depend on the (chaotic/laminar) nature of the flow over and under the surface creating differential pressures, (which also then cause the surface to deform), which is what the simulation is actually trying to estimate.

tbh, I dont think either have anything to do with any other sims problems, perf tables are just as effective at giving those forces, and for most existing aircraft they are well characterised (and pretty easy to measure if you have access to a real airframe), afaict various other sims issues arise from what they do with those forces to create a decent free body physics simulation. Which is a whole other (very) hard problem set in its own right.

See also

https://en.m.wikipedia.org/wiki/PhysX

Edited November 30, 20241 yr by mSparks

Yep! There's a ton of secondary and tertiary effects that normally go entirely ignored. If you want to do a real-time, unsteady, coupled fluid/structural simulation, put it on that 50 yr list for sure.

5 hours ago, blingthinger said:

heh heh heh. I may be super optimistic there yes. But there's no doubt that GPUs have changed the equation(s) (pun intended). Sometimes it's an order of magnitude faster to converge a solution, and we are indeed still talking about multiple GPUs to get there. Many 4090ish GPUs in the case of those big billion-cell meshes. So 0.5 hr instead of 5 hrs. 30-60hz is few orders more from that. There are a few tricks to play too. E.g. you aren't needing to converge the full way down at each time step because the previous time step is reused as the initial condition for the next step. That's an order's worth of time reduction at the least. AI/ML might also help a bit but still plenty of ?? there. I'm also assuming Moore's Law doesn't flatten out too bad in the meantime.

It's not about the available computing power (plenty nowadays), but rather writing a toolbox from scratch and years of validating the results it produces.

 

5 hours ago, blingthinger said:

Fortran? 🙂 You'd be surprised (or not) at how many modern research labs still write new CFD code in that one. Thankfully nobody's still stuck on F77, but those droppings are still often littered around for good measure.

I'm not surprised at all as the software I was talking about is for CFD (and is written in F90).

1 hour ago, Bjoern said:

It's not about the available computing power (plenty nowadays), but rather writing a toolbox from scratch and years of validating the results it produces.

What I'm getting at is time-to-convergence. It will converge. You'll get an answer. And it will arrive in ~1/30 of a second.

Whomever picks this up won't be writing a toolbox of models for the adoring simming public to chose from. There are already a few out there that are pretty decent with tons of validation studies to look at. Wouldn't take years. It will be a matter of picking a model that has done well in literature for the previous decade and go with it. And maybe add in some user-defined correction factors to go along with it.

On 11/29/2024 at 7:52 PM, Murmur said:

You do not know the details of how MSFS has implemented the so-called CFD, nor the details of how XP has implemented BET, so you can't judge whether one implementation is better than the other. The only judgment could be the final results possible in either platform in terms of flight model.

This is why I didn't say Asobo's implementation of CFD is better than xplane's implementation of BET, or vice versa. I was simply comparing CFD vs BET to see which is superior and the drawbacks, when applied to flight simulation.

@blingthinger I agree with some of what you said. Particularly as it relates to the speed of which it can be computed in real time. To be clear, I'm talking about for home simulation, not for engineering a new plane. Are there any advantage to BET besides speed when compared to CFD? I do not see any.

21 hours ago, blingthinger said:

We're at least 10-15 yrs away from a legit real-time CFD model and it will run on the GPU when that does happen. BET is orders of magnitude faster and less resource intensive than CFD and for the next decade or 2 offers more real-time fidelity than CFD can.

I think your estimation is way off if what I'm reading is correct.  The elephant in the room is the AI. It is already reducing the time to compute. While we are not quite there yet for real-time computations, we are likely not far off. See the image below of an AI assisted simulation taking 20 milliseconds vs 12 hours for a regular CFD simulation. I would put my money on CFD being incorporated into more sims going forward.

Quote

This quantum leap is achieved by training AI models on vast datasets of fluid dynamics problems, allowing them to predict fluid flow and other phenomena with remarkable accuracy.

Edited December 1, 20241 yr by brinx

2 hours ago, brinx said:

Are there any advantage to BET besides speed when compared to CFD?

I wasn't talking about engineering a new plane either. I'm referring to CFD in a general sense. Assuming infinite available time, resources, engineer skill, and no need for real-time output, no, BET has no particular benefit over CFD.

2 hours ago, brinx said:

The elephant in the room is the AI

First, to state the obvious: AI/ML is not CFD. Though it is certainly trained with CFD. The ????s I was referring to earlier when I mentioned this particular beast is that even though they claim to have vast datasets of fluid dynamics problems at hand, it's only going to be as good as that dataset. This is statistics. Not fluid dynamics. 

Over the next decade or so, they'll continue to flesh out these statistical models with canonical flow fields and geometries built from high resolution CFD. Will it be able to handle a random  airframe geometry? That's the biggest question I've got right now. Depends on how "canonical" they go? In my mind I'm comparing a C337 vs. a 747 and struggle to conjure up canonical flow fields that link the two. By canonical, we're talking cylinders and spheres and flat plates and curved plates and all manner of airfoils and even donuts, all in 3D, and many, many random combinations of them all blended together, interacting with each other. My bet is that the AI model used for the car pic you've shown was generated from a big dataset of shapes that look like 4-door sedans probably all with short trunks. I'd be 100% willing to bet that there were no F1 cars in the mix.

These statistical models also tend to have problems with smoothness. The chance of getting a bump in drag coefficient between flap settings is not zero. That's what 7-fingers-on-a-hand equates to here. It's fine for a web search.... but in a strong crosswind? I struggle to envision an AI-based model being sufficiently smooth and precise, especially with the randomness of a user's joystick inputs, not to mention real weather. It's difficult to not see traces of Will Smith eating spaghetti as an output.

Maybe if you trained it a ton on 1 specific airframe (all possible control deflections, velocities, etc.)? But by that point, you've almost certainly done enough supercomputer work to build up a really good BET model and then leave the GPU for the gfx stack.

Don't get me wrong. I'm watching this closely too. But based on what I've seen IRL with these statistical models in real engineering contexts, I'm still doubtful that a generalized AI model would be up to snuff for the flight sim crowd. Also, this isn't the only way to mix AI/ML with CFD, though the other ways on my mind are still in line with my original time frame pontifignostication.

Edited December 1, 20241 yr by blingthinger

In my opinion this AI aided computation is not much different from a look up table like GSX was doing… it is a statistical estimation of what will happen not a real calculation.

 

1 hour ago, soaring_penguin said:

In my opinion this AI aided computation is not much different from a look up table like GSX was doing… it is a statistical estimation of what will happen not a real calculation.

 

I guess you meant "FSX"? 😉