2 hours ago, mSparks said:

training an AI model requires insane amounts of compute and electricity.

Facebook/Meta wants to use nuclear power, and Microsoft just did a 20-year deal to reopen one of the nuclear plants here in the U.S.

20 minutes ago, blingthinger said:

 

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.

 

if you are thinking in terms on interpolation then I can see why you would think that the "bounds" matter, but that is not what markov models "learn".

Instead what they extract is more along the lines of cause->effect relationships, it is important that the training set covers enough of that spectrum to produce good answers, but even when presented with "new" causes they can still produce good estimates as long as the new "cause" is similar to one they "learnt". Thats why they are so good at things like image and voice recognition, they "learn" the input patterns that "cause" the output patterns.

20 minutes ago, blingthinger said:

Certainly an infinite number of tensor cores on the GPU would do anything pretty quickly.

all these models are just big matrices under the hood, one run through the matrix gives you the output. They love to talk about "x billion" parameters, but they are actually just talking about a matrix multiplication on 32bit numbers taking up a few dozen gigabytes for the largest models.

Just now, brinx said:

Facebook/Meta wants to use nuclear power, and Microsoft just did a 20-year deal to reopen one of the nuclear plants here in the U.S.

MS is reopening 3 mile island...

14 hours ago, blingthinger said:

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.

...

I bet it's just a repacked OpenFOAM with a customized ParaVIEW UI.

 

2 hours ago, brinx said:

and Microsoft just did a 20-year deal to reopen one of the nuclear plants here in the U.S.

Hope they won't run it on Windows, otherwise, R.I.P. region surrounding said NPP.

 

2 hours ago, mSparks said:

MS is reopening 3 mile island...

Oh yeah, that's a definite "R.I.P. region surrounding said NPP" then.

Edited December 11, 20241 yr by Bjoern

2 hours ago, mSparks said:

cause->effect relationships

Different model type. That's for the copilot or wingman that sees the human roll slightly right and makes a projection that he/she is going to head towards one of 3 possible targets or maybe avoid weather.

Flight model is a deterministic input/output. It's a precise replay function. It's not learning beyond the training set. If you don't include forward swept wings in the training set, the x-29 model will (be) barf.

 

21 minutes ago, Bjoern said:

I bet it's just a repacked OpenFOAM with a customized ParaVIEW

Openfoam wouldn't hurt my head. The gui better be more shiny than default paraview (which I don't mind). It's the grid gen I'm wondering about. That's not trivial unless it's a chimera-type grid which gets sketchy for boundary layer resolution and even then it's still really hands-on.

1 hour ago, blingthinger said:

Different model type.

They are all the same, "Neural Networks", which have their math rooted in the work down by Andrey Markov dating back to 1856. "language models" generally strap on googles transformers

https://en.m.wikipedia.org/wiki/Transformer_(deep_learning_architecture)

They are one (good) way to turn text into a form you can feed into a markov chain, but not the only way, they were just the one that made it into the public domain as the hardware needed to train them became affordable enough to run in a home office.

Best public talk Ive seen on it before the marketing speak became the dominant voice was the machine dreams at CCC back in 2016

Seen a lot more in the 25 odd years I've been working with them.

Edited December 11, 20241 yr by mSparks

9 minutes ago, mSparks said:

They are all the same

Not exactly. Lots of big matricies, but that's where the 100% commonality stops. 

47 minutes ago, blingthinger said:

Not exactly. Lots of big matricies, but that's where the 100% commonality stops. 

There is a vast array of ways to convert the input and output, and an even wider array of ways to train them.

but the core - what turns input into output of the "AI" models are all neural networks, NN.

mathematical name a weighted node graph.

mathematics options all documented by Andrey Markov way back when, generally referred to in the literature as hidden markov models

https://en.m.wikipedia.org/wiki/Neural_network_(machine_learning)

https://www.sciencedirect.com/topics/medicine-and-dentistry/hidden-markov-model

philosophically what they do is map causes (inputs) to effects (outputs). the "training" is developing "good" outputs for a given input.

Edited December 11, 20241 yr by mSparks

11 hours ago, mSparks said:

vast array of ways to convert the input

Ya and this is where my brain has been wandering ever since brinx brought this concept up. It's an interesting thought experiment: how to describe/discretize a 3D model of an airframe and the surrounding air state and power/control inputs such that you get a pressure field on the airframe surface as the output? You have to deal with arbitrary 3rd party modeling efforts so I imagine there'd have to be a convolution or 2 somewhere in there. Say starting with the full asoboCFD grid in 2020: that's 8000 points on the input just to get the balsa glider geometry and the surrounding air. What do you do to it from there? It's a good puzzle, this one.

Edited December 12, 20241 yr by blingthinger

7 hours ago, blingthinger said:

Ya and this is where my brain has been wandering ever since brinx brought this concept up. It's an interesting thought experiment: how to describe/discretize a 3D model of an airframe and the surrounding air state and power/control inputs such that you get a pressure field on the airframe surface as the output? You have to deal with arbitrary 3rd party modeling efforts so I imagine there'd have to be a convolution or 2 somewhere in there. Say starting with the full asoboCFD grid in 2020: that's 8000 points on the input just to get the balsa glider geometry and the surrounding air. What do you do to it from there? It's a good puzzle, this one.

I guess it would depend how you frame the problem.

e.g. if you gave a NN the object(s) shape(s) and initial fluid flow as inputs and trained it on the real world resultant fluid properties as outputs, effectively what it would "learn" inside the node graph is the laws of fluid dynamics, that might very well give you a fluid flow calculation system that runs significantly faster and more accurately than trying to design and implement the laws of fluid dynamics as computational algorithms with complex math. 

Or it could just fail inexplicably to converge on a reliable solution.

These are the kind of problems having lots of money thrown at them at the moment, some will be hugely successful, most will fail often for reasons completely unrelated to the problem.

My guess is that would probably work well  because its the same kind of problem set as voice and image recognition, which have proven so successful we all now literally have them in our pockets.

1 hour ago, mSparks said:

Or it could just fail inexplicably to converge on a reliable solution.

This is a huge part of the problem. It's not the model usage that will diverge. That's just a few matrix operations after the weights are established. The training convergence process could very well diverge though.

 

1 hour ago, mSparks said:

more accurately

More problems here too. The vast, vast majority of the training info will be from CFD. The final model won't be any better than the CFD. It won't suddenly become more accurate vs. real life or fill in gaps where there was no training data, simply because it's been rolled up into AI/ML. Attempting to anchor to data in the training process is fraught with gotchas.

1 hour ago, blingthinger said:

will be from CFD.

this is a bad idea, it would be far easier and give better results to do it in a wind tunnel with a couple of 3D cameras, they even have open source designs for low cost supersonic wind tunnels now:

https://www.researchgate.net/publication/377845291_Design_and_prototyping_of_low-cost_open_circuit_subsonic_wind_tunnel_for_educational_purposes

If it works in that, would more than justify scaling up to the big ones.

also, I see

https://www.space.com/starship-super-heavy-wind-tunnel-tests

training on CFD software then "best case" it matches the CFD software, which could very well be wrong.

Mind you, if you can train it on CFD software, then that's also a decent step to proving it is worth taking it to a wind tunnel at all.

Edited December 12, 20241 yr by mSparks

1 hour ago, mSparks said:

it would be far easier and give better results to do it in a wind tunnel 

Ohhhhhhh if only that were the case. If only!! 

 

1 hour ago, mSparks said:

Mind you, if you can train it on CFD software, then that's also a decent step to proving it is worth taking it to a wind tunnel at all

There's no doubt you could build the training set with CFD simulations. There's a reason it's called a "numerical windtunnel".

You wouldn't train a model with experimental measurements/data alone, if at all. That's not the question. At most, you would use the data to try to tune the CFD solver: pick turbulence model, tune turbulence model, pick discretization scheme, convergence critera, etc.

The fundamental questions here are: how much CFD would be needed for a generalized flight model and what canonical flows would be chosen to populate the set?

Edited December 12, 20241 yr by blingthinger

8 hours ago, blingthinger said:

Ohhhhhhh if only that were the case. If only!! 

 

There's no doubt you could build the training set with CFD simulations. There's a reason it's called a "numerical windtunnel".

You wouldn't train a model with experimental measurements/data alone, if at all. That's not the question. At most, you would use the data to try to tune the CFD solver: pick turbulence model, tune turbulence model, pick discretization scheme, convergence critera, etc.

The fundamental questions here are: how much CFD would be needed for a generalized flight model and what canonical flows would be chosen to populate the set?

I wouldnt say the dataset is the problem, - the problem is the "problem set", this stuff

getting generalisability with NNs is possible, but it's also very difficult to achieve.

4 hours ago, mSparks said:

but it's also very difficult to achieve

In the case of a generalized flight model, extremely difficult.

1 hour ago, blingthinger said:

In the case of a generalized flight model, extremely difficult.

were are not talking flight model here, just the reynolds numbers and pressures resulting from an airspeed and object shape.