Machine Learning for Fluid Dynamics: Models and Control

This video discusses how machine learning is currently being used to model and control fluid dynamics.

Download paper at the Annual Review of Fluid Mechanics:
Follow updates on Twitter @eigensteve

This video was produced at the University of Washington

[sameerdesai]

I feel like I have a hit a gold mine! Being a master student in mech, few weeks ago I was thinking of entering into this field and here I got a perfect start. I just discovered your channel and have already watched hours of videos, great visual representation and lectures. Thanks a lot professor!

[Turcian]

I think you just became my favorite teacher!

[hugoboulenc]

I discovered your videos recently and each time I’m watching one of them, I find myself pausing and pondering every 2 minutes to try to get the best understanding I can. Your work is so fascinating and inspiring ! I will be starting a PhD in fluid dynamics in a few months, focusing on atmospheric entry of space debris, and I would really like to incorporate some of these Machine Learning for sparse modelization of complex flows concepts in my work ! This feels like the future of fluid dynamics and physics in general.

[omarcavazos]

Thank you Dr. Brunton. I wish you were my fluids professor. Your explanations are wonderful and concise.

[valoraz]

Awesome and inspiring lesson. After listening to your last video I managed to make some progress in my own project, this will push me again in the right direction, thanks for taking the time to do this.

[rostamr]

This is awesome series of lectures. I will be taking this course and master program through my company.

[giulioortali]

I wouldn’t say that we have big data in fluid dynamics. In fact, we don’t have enough data to properly train “standard” deep neural networks, like they do in classical ML tasks (image recognition, NLP etc). In my opinion that’s one of the reason why incorporating physics inside neural networks, either in the structure of the network or as regularization terms like in PINN, can help us.

[DerekWoolverton]

People wax poetically about the dawn of the computer age, but I think right now is as exciting time to be working with technology as ever there was.

[tantzer]

“For ever in future, it is fast, and efficient, an accurate.” Not to detract from the beautiful work presented here so eloquently, but I wonder how such a statement can be proven when one incorporates non-analytical methods as seems to be the case here? (PS Deep gratitude to the professor for presenting these exciting developments in an accessible fashion.)

[BobbyHill]

Thank you, Sir, for explaining so nicely.

[MrTheHazelnut]

Consistent and humble presentation. Thank you.

[delphiburton]

I have modeled bioreactors for several years and how found an elegant way to think about complicated concepts like this, and often I think they are sometimes best done with literally no math

[oigxam]

I love you!!! Your approach for explain this topics is awesome!

[samsonaraya]

Can you do next Deep learning for heat models :)

[zhexu]

super informative, thanks for sharing professor

[elshaub]

First of all, thank you a lot for these videos. As an introduction or starting point for machine learning applied to fluid mechanics, these videos are gold. I have really looked for material in this field in the literature and the web and I happy to have found this channel. I have been having a lot of "aha-moments" in the last days! I now envy your PhD students a bit less ;)
I have a question to 4:10. If you write the Navier-Stokes equations pre-multiplying the viscous term by 1/Re, that means you are writing the equations in a non-dimensional way. In that case, shouldn't you be also multiply the transient term by the Strouhal "St" number as well?

[ramkitty]

makes sense that the later models were more power efficient. by incorporating the nPowers there is a more linear approximation of the harmonics rather than square wave truncation of a declarative value. There is a point in east sooke that gets an occasional ocean level cloudburst that has some sort of resonant wake vertical staking pattern that looks like a couple of pillow mints.

[delphiburton]

the idea that gravity is a perturbation from these other dimensions folded in on itself fits right in here

[keanos]

Hi Steve, first of all thank you SO much for your videos! I'm a biologist trying to learn data analysis to understand and model biological systems. I recently bought your books and while I'm waiting for it to arrive, I'm following your video lectures. I was wondering whether I should follow the older (2016) lectures indicated from the website of the book or your more recent videos on the same subjects ? I'm afraid of missing out something

[tcratius]

I just had this weird idea. I am guessing flow is a multi target scenario where each target class would hold a particular probability of occurring as in they form either a discrete or probability distribution? And I am not so sure the moth follows the big effects felt I the turbulence, it would be smarter energy wise to navigate to the lesser know systems. If this resonates with you than I will share my second thought on this matter. I say this as machine learning is what I have learnt in data science at James Cook Uni and through some of my experience with Kaggle projects. So even though I lack the maths, I can see what you wish to capture, just a thought anyway, and I love your enthusiasm with the topic. :)