[arungovindneelan]

Originally Posted by FMDenaro

I am not able to understand your final goal, you get a field distribution of local Re numbers, that's ok, and then?

let assume we completed one DNS simulation or good experimental result using PIV.
The Re definiton in 1-D would be:



Similarly, we should form other two non-dimensional numbers based on viscous and pressure forces on all the cells. Lets say we are interested in flow separation length (or angle) of flow past cylinder.

We should study the variation of flow separation length with respect to the new non-dimensional number. Every simulation we should vary one new non-dimensional number and we should keep others constant. Then we should do one regression analysis on the data and check any good co-relation exist.

Probably we may need one more non-dimensional number related to the curvature of the geometry to measure the streamline shape.

Then we should try to generalize the solution to flow past rectangular section, triangle sections and other geometries.

[FMDenaro]

Originally Posted by arungovindneelan

let assume we completed one DNS simulation or good experimental result using PIV.
The Re definiton in 1-D would be:



Similarly, we should form other two non-dimensional numbers based on viscous and pressure forces on all the cells. Lets say we are interested in flow separation length (or angle) of flow past cylinder.

We should study the variation of flow separation length with respect to the new non-dimensional number. Every simulation we should vary one new non-dimensional number and we should keep others constant. Then we should do one regression analysis on the data and check any good co-relation exist.

Probably we may need one more non-dimensional number related to the curvature of the geometry to measure the streamline shape.

Then we should try to generalize the solution to flow past rectangular section, triangle sections and other geometries.

The definition you wrote has the eulerian acceleration not the convective term. Clearly that definition fails if you solve a laminar steady flow.
Then, owing to the poitwise character of the definition, you cannot think to take constant the other non-dimensional parameters by changing only one a time.

[sbaffini]

Let me restate it more clearly: Re and other non-dimensional numbers are NOT derived as you would derive a mathematical proof.

You have a physical problem, you list the parameters it depends on, you use PI theorem and you end up with the non-dimensional parameters it depends on.

How do you use them? If you want to explore the problem by varying its parameters, you know that you can just change the non-dimensional ones and not the physical ones. That's it.

If, say, your inflow has a specific profile, then the profile shape might or not have an influence in your domain. If you want to study the profile shape influence, then you have to find the parameters that define it and the PI theorem will tell you what are the equivalent non-dimensional parameters to change in the analysis. If you are not interested in it, or you feel it is not important, you can simply not introduce those parameters in the analysis and just use the mean velocity.

As I wrote previously, it is up to you to properly list the parameters that define your problem.

These are all well known facts. You are free to stick to some NASA slide or to a wikipedia page for this, but those are just a specific view on a subject that has a more fundamental meaning. Honestly, if you ever applied PI theorem to whatever toy problem you would have understood this very clearly.

Going to the AI stuff, while I have no access to the paper you linked, I see the authors are all from computer science background and the references are just ridicolous. This is all I need to know to conclude that it has no science in it which has something to do with fluid dynamics or CFD.

But please, don't let me start on AI, Big Data or any other fancy stuff of today, otherwise I'm going to become rude.

For what concerns your interest, it is up to you how to manipulate your field data to take proper information out of it. DNS exist since the 80's and people have found a lot of ways to extract information from it. You can find your own, but that's not related to the definition of Re number.

[arungovindneelan]

Originally Posted by sbaffini

Let me restate it more clearly:

Going to the AI stuff, while I have no access to the paper you linked, I see the authors are all from computer science background and the references are just ridicolous. This is all I need to know to conclude that it has no science in it which has something to do with fluid dynamics or CFD.

But please, don't let me start on AI, Big Data or any other fancy stuff of today, otherwise I'm going to become rude.

I like to stop this discussion, probably I'm unable to make my point clear and coherent.

If you wish you can find the paper in the following link. It is open.

This is a back-box kind of analysis. Please keep in mind that we can predict some flow features and do visual validation just from our experience (previous simulation images and plots). Since the computer can process more info than us, we can give that intelligence to AI. If you read the paper in that view it may make sense else it will be rubbish.

This paper may look odd because it is based on a convolution neural network. In my personal opinion that may not suitable for ODE or PDE. You may aware of the neural differential equation which is more related to FEM principles (minimizing error in PDE).

Thank you.

[sbaffini]

Dear Arungovindneelan,

we probably misunderstood each other.

Let me say that your post, including its title, was pretty clear about what you were asking for and, I think, we answered clearly on that.

At some point, out of blue, you mentioned AI and a paper.

As you are not among the authors and that was not the main topic of this post, I just wrote, with some irony, to leave aside AI, which is a controversial topic.

I clearly understand, I think, what you are trying, but that is not about the original topic of Re derivation (which, again, doesn't even exist).

I suggest you to open a new thread along the lines of "what would be the best way of training an AI for CFD" where you clearly explain your doubts, the context and the methods.

At some point, if that thread gets traction, I might decide to be rude there, but that's totally within my rights and you could simply ignore me.

I will, however, get back to you here once I read the paper, and if it is good CFD science I will apologize.

[sbaffini]

Originally Posted by arungovindneelan

I like to stop this discussion, probably I'm unable to make my point clear and coherent.

If you wish you can find the paper in the following link. It is open.

This is a back-box kind of analysis. Please keep in mind that we can predict some flow features and do visual validation just from our experience (previous simulation images and plots). Since the computer can process more info than us, we can give that intelligence to AI. If you read the paper in that view it may make sense else it will be rubbish.

This paper may look odd because it is based on a convolution neural network. In my personal opinion that may not suitable for ODE or PDE. You may aware of the neural differential equation which is more related to FEM principles (minimizing error in PDE).

Thank you.

Ok, I've read the paper. I'm not going to say it is rubbish, because it isn't, but that's not CFD science in my opinion, and I think it wouldn't have been accepted on any major CFD journal before 2005.

As you mention, it has, more or less, the same value of a human visual perception of a case... except that I would be able to identify when a case can't be steady laminar, while it wouldn't because it has not been trained to do so.

But, more than this (flow initialization, for example, would still be a very good point for the method), it is the overall approach that seems to be fundamentally flawed to me. But, as I said, this would require a different thread to be opened.