Machine Learning for Fluid Dynamics: Patterns

[@UTElistan]

This is exactly how authors should present their papers. I wish every paper on arxiv was presented this way

[@luorisluo3634]

Being a student to Steve must be one of the most amazing thing in the world.

[@shoutash]

I’m just getting into ML and seeing these videos as a CFD researcher gets me absolutely excited! Thanks for the great content! :)

[@redouanehemi9693]

i want also

[@superuser8636]

Wow. What I love about watching one of these videos for the 1000th time is that I'm just as blown away learning both the history and the math as I was the first time I saw one years ago. Thanks

[@amriteshsinha437]

I use CFD quite heavily and I can't tell you how much I appreciate this . Wonderful channel and nice explanations.

[@Default_089]

Just a month ago I was watching his lecture on deriving the Fourier transformation. This is beautifully done

[@deiling5034]

I am so gratefull this is for free and available nowadays

[@__--JY-Moe--__]

laser scan in a vacuum, on a vibration table..so now nn's can read right off the table.....great presentation..easily understood..thanks...looking 4 a class now!!

[@kubafrank96]

I'm about to start my PhD studying fluid flows around wind and tidal turbines. Never done ML before but this looks like it could be relevant for my future work (I know for a fact that the lab uses PIV measurements) so I'll try to get as much under my belt from this series.

[@mcmormus]

Dear Dr. Brunton i have been watching your videos for some time now for fun and it fills me with joy. Even though I have nothing to do with Deep Learning and Neural Networks (YET), I just wanted to say thank you! Greetings from Germany.

[@vikaskushwaha9619]

I really appreciate the work you have done; the best part is that you came forward and explained your part of the work via visual (i.e. presenting) as it gives a much better idea even for a student like me without going into high-level details. Of course one can read the papers, but as there is a lot of work going on in many groups, in many sectors sometimes it just hard to find the right one for the right person. However, from your paper/visual explanation, I have got a better understanding of how ML can be used in Fluid mechanics. ~ Just thought of a master student 😊

[@RobertKost]

I learned quickly, just to keep up. Excellent presentation!!

[@sjh7782]

I love your contents and delivery, Prof. Brunton. But I am wondering could you maybe talk about how might machine learning (or data-driven approach) help with the development of turbulence theory?

[@arvindnatarajan6853]

That will open doors to a wider field and it would be awesome

[@nikhilm4418]

Thank you Dr.Brunton for yet another informative video. As mentioned at 16:20, extrapolation in such cases is extremely challenging to do well. In this context I was wondering what your thoughts were regarding ways to incorporate physics models directly into the machine learning / inference pipeline (as the process becomes costlier). A video about this topic would also be very interesting.

[@ryanli4689]

There may be an error in the line chart on 18:31, the horizontal axis represents the mode number instead of the amount of training data. I went to the original article and I think the right figure should be Figure 12 (a) of the original article, instead of Figure 4 that is used now.

[@zildjiandrummer1]

I'm subscribed. This is great work, and the currently-bloated ML field needs way more distillation to re-contextualize research such as this.

[@JulioDiaz614]

Dr. Brunton, I am performing measurements of two-phase flow with x-ray radiography which by nature suffers from the exact problem of parasitic noise. We'd be interested to separate "noise" content from the Low-rank matrix of the flow patterns. Would love to get your opinion on this and perhaps refer us to the paper pertaining this application.

[@bluecpp]

This is the true feeling of love! Thank you for this feeling. ☺️🙏🏻

[@crackyflipside]

This is fantastic work.

[@wodemamaya1889]

Awesome! Steve, you have the talent to make the AI stuff human friendly. lol

[@raviprakash5987]

Sir, Please don't stop making videos.

[@mattkafker8400]

Excellent video. Thank you, Professor.

[@saurabhtalele1537]

Really there is no words for u r work Sir How to start in this field, please give some suggestions, i think may be one day , computer will give some advance equation i.e may called improved or optimised navier strokes equation by computer or may be computer will reduce time drastically by ML/AI implementation in CFD, sciml.jl also working in this direction.. Really awesome work sir...

[@jacopobilotto9221]

Thanks for the content you upload. It is highly appreciated

[@JBHACKSAW]

Gotta find time to try this at any cost.

[@anilsharma-ev2my]

By applying fluids mechanical theory so we found next date of earthquake easily Since we compared it's with burnouli theory and continuity equation and tangent

[@severnsevern1445]

Great tutorial!!! Powerful presentation!

[@gustavoexel3341]

In the autoenconders section (5:50) we see the U and V matrices from the SVD being used to compress and reconstruct the vector x, in what seems like z = Ux x_hat = Vz But from what I remember from the SVD you would “compress” your data multiplying by the transpose of U, and “decompress” multiplying again by U z = U^Tx x_hat = Uz And where U could be the economy version, or even a truncated version. Can someone explain what's wrong with my reasoning or what he meant by those matrices U and V?

[@Eigensteve]

This is a really good point, and you are totally right! (I am giving this set of lectures this week and I just realized the same thing, so funny timing!). Yes, the "encoder" is "U^T" and the decoder is "U" where U is the first matrix in the SVD.

[@AdityaChaudhary-oo7pr]

Amazing explaination ... Thank you

[@swapnilbhabal5289]

Would love to see Machine learning / Deep learning course for audio by you

[@cookiemoises101]

Are these methods limited at all by temporal resolution? In a given plasma physics problem, various phenomena may occur in varying temporal regimes (some varying orders of magnitude) and a kinetic approach like PIC might produce too much noise. I’m not sure if it’s being involved but it seems like these methods would be great for particle methods?

[@danawen555]

fantastic video

[@bryancovell2031]

How do you encode physics such as Navier-Stokes into ML models to augment image-based extrapolation problems?

[@mikesmusicmeddlings1366]

How do you decide what the eigenflow fields should look like geometrically? I understand you can get an accurate enough representation by summing weighted eigenflow fields, but it seems really difficult and arbitrary to choose what those fields should be

[@abdallahshaat2296]

Dr. Brunton, why do you use orthogonal decomposition?

[@motbus3]

Hey professor, one question. would you do a series about creating visualizations for simulations ? would be really cool to have Matlab and python version.

[@danielpetka446]

I would love that too

[@milansekularac6196]

Fluid flows are very nonlinear, with scales ranging from Kolmogorov to integral scales, plus the time scale span, let alone the added physics like compressibility, heat transfer or reactive flows where all these phenomena interact. How does the computational effort for these approaches compare to the numerical solution of PDEs governing flow? Thanks

[@SAINIVEDH]

Sir, can you give insights on current ML research in the field of structural engineering. Thanks

[@HawkinsOkeyoEng]

Hi Sai, am also trying to brainstorm on this, but so far, am kind of stuck on how to approach it....I would really like some insight if something pop-up on your side

[@mrunaldharbathula9872]

Great work sir

[@davidhasin6258]

Super interesting and applicative!

[@prashantdahiya711]

What does Eigen values have to do with fluid mechanics ?

[@m0nzderr1]

Great and beautiful ML applications! Although the "image-based" approach looks very promising, IMHO it is still too far from being applicable in the real CFD world. Especially, when one has to make investment decisions relying on simulation data (often from previously unseen scenarios). Usually CFD simulations are carried out to obtain the detailed 3D fields of variables that satisfy conservation laws and represent certain physical and chemical phenomena. Such fields generated from compact representations definitely will look realistic but not necessarily realizable. Without additional constraints neural autoencoders have too much freedom to choose "what is under mustache". On the other hand, despite being computationally hard, LES simulation, for instance, will always guarantee a realizable picture within given level of details. I don't believe in ML alone being able to pull the CDF off, but ML tied to CDF models could be the future.

[@姚命宏]

I think how to measure the fluid field at each point also matters, because that is where the reliable “data” for ML models come from. The smoke and laser thing he talked about in the video is a great example of how to measure a fluid field with great details instead of a few points in it. Each particle, when combined with the laser shining on it, is a sensor that can measure the local velocity. All such sensors combined together can be a precise “data description” of the field, and possibly also of the solution to corresponding N-S equations.

[@m.ai.chi.22.]

Hi Proff, is the code for thr Erichson's super resolution provided ? I was wondering if it could improve PIV result's time resolution.

[@zrmsraggot]

Hi At 6:41, is z vector ' interpretable ' ? Thanxs

[@stephenoni2019]

Hello Dr. Brunton, what software would you recommend for geologists with no background in fluid mechanics to model fluid dynamics of magma chambers? Thank you!

[@1olp1]

why not outsource this task to cfd engineers?

[@tassioleno808]

I need a copy of your book!

[@paulleveque2624]

Hi awesome stuff, il curently studying mechanics in France, and I would love to learn and study more about deep learning and mechanics but my teachers tell me that it’s new stuff ne not a lot of people have studied that field ... any advices on where to look for learning more about these subjects ? Entreprises or university ?

[@notgabby604]

Why don't you try Fast Transform fixed-filter-bank neural nets? They swap around what is adjusted with fixed dot products and parametric activation functions (A) like fi(x)=ai.x x=0 i=0 to m. The fixed dot products can be enacted with fast transform (T) like FFT WHT etc. To stop the first transform from taking a spectrum you can apply a random fixed pattern of sign flips to the input data. Such a net then is: sign flips, T, A, T, A.......T. The fast Walsh Hadamard transform is nice.

[@daesoolee1083]

Woah I love the topic.

[@amarilloatacama4997]

Thanks for the video What book can I use to learn the very basics of POD/PCA? What area of Mathematics does it belong to?

[@TheMazyProduction]

Steve has videos on them. They belong to data compressions.

[@SAINIVEDH]

kzbin.info/www/bejne/nqfJfmOMqcR0hpo&ab_channel=BobTrenwith

[@amarilloatacama4997]

Thanks for the answers, I was asking for books rather than videos. I edited the question.

[@drscott1]

Brilliant!

[@camiloruizmendez4416]

Links for the papers?

[@姚命宏]

How these ML models are going to help us understand the physics of fluid? Modern ML models are mostly about correlation instead of causality. In another word, if the modern ML models are relying on the “fluid data” that we can sample from reality, it seems impossible for such model to reliably generalize to the scenarios where we can’t sample data, and yet I think these scenarios are where the true challenges present, when it comes to both the understanding and application of fluid dynamics.

[@PaulPukite]

The reason that ML works for fluid mechanics is because all it does is manipulation of non-linear relationships. It's finding these non-linear patterns, yet we don't know how to reverse engineer them -- which is where our current math predicament lies.

[@anandratanamaurya9975]

Please name the software for machine learning.

[@abisarwan20]

Do u have discord's group?

[@leif1075]

He said modern deep dreams popylar today..am i the only obe who has never heard of or seen this libd of painting?

[@judgeomega]

Yannic Kilcher recently reviewed a ML paper which is relevant to this video; kzbin.info/www/bejne/f5K2aGWXfdd9gac

[@rosemariekatemutya2746]

Tunay wala akong maintindihan 🥱

[@jermainedecastro9319]

Ako din bhoiii 🤣

[@tenidomarvin2591]

Really?? It is understandable