Machine Learning for Computational Fluid Dynamics
Вставка
- Опубліковано 28 лип 2024
- Machine learning is rapidly becoming a core technology for scientific computing, with numerous opportunities to advance the field of computational fluid dynamics. This paper highlights some of the areas of highest potential impact, including to accelerate direct numerical simulations, to improve turbulence closure modeling, and to develop enhanced reduced-order models. In each of these areas, it is possible to improve machine learning capabilities by incorporating physics into the process, and in turn, to improve the simulation of fluids to uncover new physical understanding. Despite the promise of machine learning described here, we also note that classical methods are often more efficient for many tasks. We also emphasize that in order to harness the full potential of machine learning to improve computational fluid dynamics, it is essential for the community to continue to establish benchmark systems and best practices for open-source software, data sharing, and reproducible research.
The Potential of Machine Learning to Enhance Computational Fluid Dynamics
Ricardo Vinuesa, Steven L. Brunton
arxiv.org/abs/2110.02085
Citable link for this video: doi.org/10.52843/cassyni.nn3m2c
Link to Rose Yu's seminar on incorporating physics into turbulent flow solvers: • Rose Yu - Incorporatin...
This video was produced at the University of Washington - Наука та технологія
Just brilliant! Started a month ago my PhD and this video along with your ML Ann. Rev. have just made my background reading a lot easier to get started with. Thank you!
I find this video really giving me the information I was trying to collect these days. Thank you so much! Very beautiful.
Thanks for another great video! As a CFD engineer this is very wholesome :)
Amazing talk, thank you very much for spending the time and for the great delivery!
A great video, thanks very much for your sharing! As a PhD. in fluid dynamics.
I think it's thrilling seeing how ML can be applied to different fields of science, in particular, physics! I'm really interested in learning ML albeit slightly for more hedonistic purposes like high income careers with Data Science, but I always grin and get excited when I see how this booming field is being applied to solving open problems like fluid computation, quantum, and even biology like protein folding. :D
I love watching these videos. Thank you Prof. Steve!
I am really interrested to this field, I work on turbulence modeling with ML in my PhD. thesis. Thank 's Prof. Steve.
Interviewed Steve a while ago, maybe this helps: ua-cam.com/video/gzggdJ4HB38/v-deo.html
The reproducibility and sharing the training data is the most important message of this talk
Amazing video. Thank you so much both of you
Finally! The video which I was particularly looking for ❤️
really fascinating...we're exploring the use of ML in micro weather applications (i.e. winds and turbulence in urban canyons)
Thank Steve and Ricardo, so impressive to see how ML is applied in fluid dynamics in a systematic way. This is the one area I really want to dig into in my following career (in Ph.D. if possible). Can't wait to read the paper.
Thank you!! :)
Great video. He accelerates a lot my understanding.
U'r audio is ''low''!! you always blow me away with these! thank's! love this!! so helpful !! 🍌...I don't need 2 use ansys!! good luck!
I am so excited your topic that I use cfd to predict chemical process.
Good topic and I love your channel
This would be great using as a predictor for a higher resolution simulation.
Thanks for great video
Very interesting! What are the tools you are using for your presentation?
As always very nice and inspiring lecture.
GREAT VIDEO!!!
Very nicely presented. One of the best I've seen. I am very interested in learning more about ML for CFD. I have seen some interesting and very promising work on FEA. I have to add a disclaimer here in that I am a CFD software provider for a developer that has integrated a lot of in intelligence in their product, which makes it much faster, easier, while being very accurate. I love what they have done and I am very patiently waiting for AI/ML based CFD to come of age to even further decrease the computing power and provide extremely fast analyses. Keep up the amazing work!
Hi Steve, can you please recommend the essential videos (in a systematic way) of yours in this channel that are a prerequisite to watch prior to understand this paper in full. Thanks a lot
I suggest ''Machine Learning for Fluid Mechanics" by Prof. Steve et al. It's verry useful to understand ML, even if you are a computer science engineer.
Suggested Paper: www.annualreviews.org/doi/abs/10.1146/annurev-fluid-010719-060214
Shameless plug from my side: ua-cam.com/video/gzggdJ4HB38/v-deo.html - interview with Steve :)
Bravo Steve !!
Love the explanation and ways of portraying literature in these area. Excited to read that paper. Last few minutes where you talked about benchmarking, reproducible results, and open source code are the keys. Also, to be critical while comparing with state of the art techniques and finding which to use for your problem statement is first step to go ahead with. Really enjoyed the presentation. Thank for sharing.
insane work
Great video, but the sound is too low. I need to use max volume.
Thanks. It is somewhat frustrating that there are no links in the description to the all the articles mentioned in the video. For example, for 2 articles of Beetham & Capecelatro 2020 i found only 1. Is the 2nd one from 2021?
11:43 With the results presented so far I'm not impressed, because it's always possible to optimize a stencil or WENO scheme for one particular problem. I would be curious to see what these NN based schemes do when presented with new problems. I've yet to see any NN based approach be used as a black box to improve or accelerate CFD calculations. Also, for the interpolation problem, wouldn't any monotonized scheme cure the overshoot issue and be much cheaper to evaluate? How many weights are in that network - how many FLOPs? I guess I need to read the original paper but I don't understand what is so amazing about that.
Bummer!
This is so overhyped! It is easy to fool people without core knowledge of CFD.
@@kesav1985 Indeed
oh no steve brunton turned drumheller fountain into a flying saucer
Also, nice video and exciting new research!
This is the future
good video
What is the performance difference between a direct computation and an RNN DNS?
One video on turbulence model with fourier transform
Hi,
I am relatively new to this. How can you compare a numerical simulation of a PDE to the exact solution, when you can’t solve the equation and hence don’t know what the exact solution is?
People use the method of manufactured solutions for this sometimes. You specify the solution (satisfying IC/BCs) beforehand, compute the differential operators based on this solution, and then include the result as a source term of the PDE. This only tells you that you are solving the PDE correctly, it does not tell you that your PDE + chosen parameters are a proper fit for the physics.
Can you post the link to Rose Yu's seminar at UW?
Thanks for the reminder! Here it is: ua-cam.com/video/h7TfFssBFEs/v-deo.html
@@Eigensteve Wow, quick reply! Thank you for putting in the hard work to make these videos. They are magnificent!
What is the physical meaning of each POD
I came here for my undergrad project. Well it's out of my head 😅
This content displays an impressive depth of insights. A book I read with like-minded themes influenced my path. "The Art of Meaningful Relationships in the 21st Century" by Leo Flint
OMG why I can almost find one of your video on every the topics I'm interested in/stuying
Music cool! Name, please?
Bam, 1000th like! :)
Ummmm. This is interesting, but I highly suspect that the ML model used for one specific set of conditions will not properly predict outcomes for other conditions. So, I’m not super sure how actually useful this is in all reality.
some of the mentioned methods (such as SINDy) are meant to produce models of stable predictions beyond training conditions
👏👏👏👍
I am sorry to raise some criticism, Prof. Brunton, I am an old CFD engineer with some experience in development and industrial applications. As a novice to ML I feel a bit disoriented, I went through the paper of Kochkov, that of Sinai, and honestly, some of the things look to me completely pointless. At 7:21 there is DNS on a coarse mesh, that needs to be trained on the fly, using a DNS for the same test case on a high resolution mesh. Does it make any sense?? Likewise, at 8:55 I can see the Burgers'equation accurately described by the neural interpolator. But can we apply that same learned model for another equation and having the same accuracy? Turbulence modeling also is questionable, and many important CFD groups seem to have already ababndoned the idea. The only part which seems very interesting is the POD, but it is not obvious to me how this could be transferred to industry heavily relying on CFD (steady RANS, URANS). Sorry for the naive comment.
Volume is too low
ML for solving differential equations is totally hype. It can not solve large-scale simulation sizes.
My CPU got scared...
So much computing for almost no relevant result...