There is a fundamental key in understandin the LES solution. Consider for example the filter to be a volume averaging. Computing a value for a volume averaging can be obtained (in principle) by a DNS solution. But the DNS solution is not unique as different fields can produce the same averaging.

Could that be a reason for different attractors?
I have no exact answers to all these doubts, as I wrote before I would test a unique flow problem performing a DNS, then I would try both LES and RANS using a perfect model. But, again, I don't think that AI/ML technique can add more answers...

yes, of course, but what is done in the perfect LES / RANS publications is taking a single DNS, then computing the closure terms through filtering and then computing the resulting coarse scale solution - since the fine scale solution is considered "given", there are no different instances / samples. The fine scale solution is given, a coarse scale solution is defined by a filter (any shape), and the coarse scale equations (which are derived from the fine scale ones by this filter) are then solved for this given closure / source term.

That indeed sounds like an interesting research project!

No, they cannot. But the problem instability when closing the RANS equations with the exact terms is a separate issue - one that should be investigated strongly in my opinion, because if it turns out that the exact closure terms blow up the solution, then all the models, ML or not are in trouble.

Since this thread has derailed to mainly LES discussions, I think it is ok to post a talk of a somewhat dystopic future. So what do you think our children will say about LES being impossible to improve with AI? Fiction or not, a very interesting talk.



https://www.ted.com/talks/sam_harris...ontrol_over_it

Human intelligence was able to find, that the answer is 42. Now we need AI to find the right question :-)

Pretty distopian, but the old spiel. It is funny how on the one hand people herein thisthread doubt MLs usefulness to LES and on the other hand, people are afraid of superAIs that will see and treat humans as ants... which one is it then? I believe both extreme views are unlikely.

Fundamentally, the more one has experience in a field, the less will trust in it :D

We all know here that CFD is total bull...t [emoji12]

Can you please share the title of the paper you are referring to?

Here it is the link to the original article. Actually, it isn't even the oldest one I've heard of, as there is stuff done even in the 90s.

https://www.sciencedirect.com/scienc...45793001000986

Sorry, I forgot to add the link. I added it now.

Why artificial intelligence is the next big thing
 

Artificial Intelligence is one of the developing equipment which attempts to mimic human intellectual skills in AI systems. Artificial intelligence systems are the aptitude of a computer software package to acquire and contemplate. All can be well-thought-out Artificial intelligence if it includes a program doing something that we would generally think would depend on the brainpower of a human. The rewards of machine learning artificial intelligence are vast and can transfigure any specialized sector. Here are a few:
1) Lessening of Human mistakes
The catchphrase “human error” was conceived as humans make errors once in a while. Processors, on the other hand, do not make these slip-ups if they are automated correctly. With artificial intelligence online, the judgments are taken from the formerly congregated data relating to a particular set of processes. So inaccuracies are abridged and the opportunity of reaching accurateness with a superior amount of meticulousness is an opportunity.
2) Takes chances instead of Humans
This is one of the largest compensations of Artificial intelligence. We can surmount many dangerous boundaries of humans by making an AI Robot that consecutively can do hazardous things for us. Need to send someone to space? Placate a bomb? Find out what’s at the bottom of the ocean? Machine learning solutions can be used successfully in any sort of environmental adversities.
3) Works all the time
A Run-of-the-mill human will work for 4–6 hours a day exclusive of the breathers. Humans are put together in such a method to get some breaks for uplifting themselves and get prepared for a fresh day of work. But using AI we can make machinery work all hours of a week short of any interruptions and they don’t even get jaded, not like humans.
4) Does all sorts of mundane tasks
In our everyday work, we will be executing numerous monotonous tasks like sending messages to clients, authenticating particular documents for mistakes, and numerous additional tasks. Utilizing artificial intelligence we can efficiently mechanize these routine tasks and can even eradicate “tiresome” responsibilities for humans.
5) Makes quicker and better decisions
Using AI together with additional expertise we can make machinery take choices quicker than a human and perform actions rapidly. While taking a verdict humans will investigate countless factors both expressively and sensibly but AI-powered machine works on what it is planned and conveys the consequences in a quicker manner.

There's a reason it earned the monkier 'colourful fluid dynamics'.

Well, at least not as bull...t as AI [emoji1]

That sounds like you are late to the party. Just check out how many profs with good names from good unis do AI + CFD.

This is wrong on so many levels, and I tend to take things very personally during this pandemic, so:

1) I was kind of joking, as I called for bull...t also in CFD, yet...

2) A "party" is the opposite of what I have learned about science. And differently from those at that party, I am free to do whatever I consider relevant without bothering about what masses or employers consider relevant when they wake up at the morning

3) This mindset, and the fact that most people can't study references beyond what they already know, are the exact reasons for which most of the current production in the field is mostly irrelevant for CFD. Peer review is also completely fucked up, so things are not going to change. Thus no, having a published paper on a subject doesn't make it automatically relevant, except maybe for the author whose paycheck depends on some stupid metric based on that

4) AI is not new, in the same way DNS and LES aren't. The general availability of compute power and tools is. You say I'm late, I studied machine learning 15 years ago, and its use case in CFD is changed almost 0. Also, flows in videogames are not CFD

5) There are exceptions, of course, but I count them on two hands, maybe one

6) Few questions: have you studied machine learning at least? How many of the techniques in it you already knew from other areas? For those you didn't, have you managed to understand what they actually do?

7) Have I mentioned that I was joking?

my apologies, I was using colloquial language that obviously has been misunderstood. The term „late to the party“ has nothing to do with a societal or political party, it refers to the party with drinks and dance. The saying just means that I have the impression that you are squarely against AI for possibly two reasons a) factual ones or b) you are „late“ in a sense that you have not picked them up / worked with / publisher something. I cannot answer if it is a) or b), only you can do that for yourself. That should take care of points 1,2 and 7.

W.r.t. point 3: You are one of the brave people who publish under their real name. That is commendable, but it also allows people to „google scholar“
you. I just did to get a feel for what you are working on / known for. Either the system failed you and treated you unfairly (which is possible), or your scientific output is moderate. In any case, at first glance, your opinion on AI in CFD seems much less founded than e.g. the works of Duraisamy, Ling, Karniadakis,... so you can blame the system and say its broken, that is another issue. But within the current system, many ppl with tons of publications, reputation and tenure disagree with you. That is my answer to 3.

w.r.t. 4, see comment above. Claiming that AI has been around for a long time and not changed CFD disregards the explosive development in the last 5 years. That will take some time to manifest, for sure.

w.r.t. 5: care to name them?

regarding 6: Yes, I have.

AI or more precisely ML is changing all the other simulation sciences right now. Just pulli g up the drawbridge and hiding behind the moat will not make it go away for CFD.

Apologies as well, because I might have answered too harshly but, yes, I actually understood what you meant by "party". Still, you skipped entirely the joke part and argued about a possible personal reason, a childish one indeed, for me to have that opinion, instead of debating it.

But let me dissect your reasoning further.

I am not in academia and I don't have to publish anything for living. Also, I pursued the Ph.D. path because I had the chance of working on what I wanted, not because I wanted an academic career (and I actually secured a job the day after my defense). But even if that wasn't the case, there is an infinite amount of CFD areas where, quoting you, I would be late to the party. If that was a possible reason for me to blame AI in CFD, that would have applied to all of them. I consider myself an LES expert, I landed on it when the party was way over and didn't have any relevant publication. While I would jokingly call bullshit on LES more than anything else, I am certainly not against it. So, I really don't understand your argument.

Also I am not "squarely against" AI in CFD, I just jokingly called bullshit on it in the same way I did for CFD (and would for LES).

Certainly, I am, indeed, kind of skeptical about the hype for something that has been there for years without anyone caring at all, and all of a sudden now has billion publications. This inevitably implies low average quality, but what is most appalling to me is the fact that quality standards for certain things like verification and validation dropped everywhere, at each level of publication. Which wouldn't even be that worrysome (considering that the tool indeed has certain limitations) if it wasn't for the fact that it is happening in an era of already low students committment, with the risk that such low standards will become the new normal (if they haven't already). Also, honestly, I miss the value of a numerical solution whose accuracy can't be tuned and whose author cannot put a bound on the error. That's not how I do CFD . So, yes, I feel like I have factual reasons.

I won't make names, but let's consider Karniadakis, which would certainly be one of those I would suggest reading. Karniadakis is an authority in almost every field he worked in. And as mathematician was certainly aware of the field way before the internet. But he just jumped on it when everyone did, because that's how most of these parties work, you don't go alone if you want to be paid.

Before AI it was UQ, before that it was GPU, etc. etc. I don't need to be myself a publisher at all to argue about the fact that it is statistically curious that everyone shoots at the same target at the same time.

In any case, the fact that you consider my scientific output as a surrogate for my scientific reasoning, and thus making any of my conclusions second to those of Karniadakis, or anyone with more publications than me, trust me, is a fundamental aberration of how science works, and needs correction immediately. What then if you had to review a work of someone with more publications than you? Simply accept whatever he throws at you because he is smarter? Have you actually made any review at all?

Finally, let's get to the core argument, the explosive development of AI and CFD. Could you, please, name a significant advancement of practical relevance you are aware of? Why you think ML is changing all the simulation sciences and why this is happening now and couldn't have happened 20 years ago?

Actually, I have no problems at all with ML/AI, but with people arguing about them as if they were magical boxes whose marvels just need to be discovered. They aren't. Neural networks are just a sophisticated interpolation/extrapolation technique, and most of the other non supervised methods already existed previously with various names, even in CFD. If you understand what interpolation/extrapolation are, and their limits, then magic just fades out.

Deepfakes? Awesome. Blind surrogate for a well established, deterministic, mathematical technique, just because is fancy and you can? Not for an engineer that has to give factual answers

I share sbaffini opinion. What is your intuition here? It seems to me that there is a bit of frustration in your comment.

I have a feeling that you are one of those new python tensorflow guys who has never done any other real programming language, nor has any experience in approximation theory or applied mathematics. Otherwise I can't explain your naive statements. Where are the fundamental developments?

Expert in using Tensorflow? What else did you learn?

I think you know how science works. Please provide evidence for your statements.

Regards

I just happened to think about this few days ago, and I tought to leave this here before we start again on this thread in the future. Imagine sampling the following 1D function:



with on a grid with and being the domain size. So, basically, one component is fully represented on the grid while the other is aliased. Is there any work in machine learning where, taking an arbitrary number of aliased samples as input, maybe together with and , the values of and are both "deduced"? I would be glad to know if such a work exists and to retract my previous skepticism in case it actually does that.

This is not even close to turbulence but, still, reflects the main issue I see in using ML for certain tasks.

playing games with people is not helpful