[aerosayan]

AI based CFD is getting pretty popular these days. They are claimed to be 10X faster than conventional CFD. I can see why that would make it very lucrative in gaming or movie industry.


What do you all think about the future of conventional CFD and AI based CFD in engineering field?


Regards
~sayan

[sbaffini]

For how it stands today, machine learning (AI is a much larger container that has stuff in it that wouldn't play any role at all in CFD) is just a very sophisticated, non linear, interpolation tool. There are efforts to make them physics aware, but the notion that a limited set of examples can be representative of any possible physical scenario, I think, is ridiculous. Also, I am not aware of any possible mechanism by which the error in a "prediction" can be bounded by a norm nor reduced by acting on some parameters (but I would be glad to be contradicted).

As an engineer that is paid to give technical answers, I can't give those anwers in this way. Or, at least, I don't want to.

It certainly has bazillions of use cases in CFD but not, in my opinion, replacement of actual CFD tools. This is in the same vein that RANS will never be totally replaced by LES which will never be totally replaced by DNS. They are different tools, with different input, output, accuracy and costs.

The most promising, practical and honest proposal I have seen so far is as replacement for lumped parameter models in 1D codes. If they are up to the promises that they make, they will be a clear winner against any other heuristic model. Another interesting use case is in solution initialization and inflow generation (both RANS and LES/DNS).

But, even for other modeling purposes (say turbulence modeling) I see theoretical limitations (and most of the works produced in the field show that most of those working on it are not sufficiently equipped to even bother about certain things).

Unfortunately, the discussion today around AI and related techniques is strongly polluted by: lack of understanding of the math behind it and other techniques, hype, greed. One example that really hitted me hard was a paper about LES SGS modeling and machine learning. Not only the paper made a lot of confusion on fundamental aspects of LES and the role of structural and functional modeling, but at some point it showed the spectra of some machine learned velocity reconstructions, kind of the key of the whole work, and it was, obviously, pure noise, but none of the people involved in the published work (authors and, apparently, reviewers) had any concern about it.

So, I think it is still very difficult to predict today what will come and would distinguish between what actual CFDers will do and what the market will do. People doing actual CFD (as opposed to those that have barely any idea of how an actual CFD case is handled) should have all the instruments to easily compare the two tools and find their trade off (because there will certainly be some use cases for them as well). The market, those who can't tell apart two codes or even two schemes, won't probably do more harm with these new tools than what they already do with classical ones.

[sbaffini]

However, there is another point of view here, which is related to the market, the software vendors.

One thing that I find, to say the least, peculiar of the CFD software business (and probably most engineering software as well) is that, while it could be among the most scalable businesses, turns out it is indeed not, because providing support is a cost, and it is not exactly like your favourite app support, because it is expert support. Still, the software development part is quite slow and easy.

Now, consider any possible commercial tool based on machine learning, and you have still the same support costs (because if you lower the level so will the users, while still pretending to solve the same complex problems), but now your software development costs are potentially much more, because in the battle for the best trained surrogate, the training never ends.

In addition to this, even maintaining such models is probably more delicate and costly than traditional software, both because you probably have several of them and because, without particular insight on how they give their answers, you have to be more careful in changing an old model with a new one. Or, at least, this is what a CFD user would expect from the company, to be able to replicate results.

So, I expect the costs of a related business to be relatively high with respect to the traditional approach. Add greed and the fact that most will try to sell the thing as simply instant CFD (or something along these lines), and there are chances that the cost will not differ significantly from traditional tools, if not higher.

All this to say that, probably, it is not even granted that pushing these tools as replacement for traditional ones makes business sense at this stage. It might not, despite the progress in their development.

[aerosayan]

I don't really trust AI either.
Main reason : We don't know what kind of bias we're teaching the network
There was this dog vs wolf image classifier, which learned that images that have snow in them, are of wolves. It was quite a scandal at that time.