Wow, as both Mechanical Engineer(specialized in energy field) and Software Engineer, this was my dream to I want to work on :) I feel very lucky to came across to your channel.
This linked vid is some kind of weird thing, but it did put the idea of using a steadily rising pure tone soundtrack over a video as a possible embedded time index.... not sure if it makes more sense than a frame count... but it is interesting. Aside from that though... weird. Oughtta prolly delete.
This is exactly the type of video ive been wanting to see. Can we set up a zoom? I would love to talk to you and ask some questions. My name is Isaac Castro I am 21 and about to start my masters program in applied math. In my own time I have been working with machine learning and simple fluid equations. Would love to talk and ask some questions. Let me know!
Beautiful lecture, Steve. One of my dream projects is to combine topology optimization with fluid mechanics to iterate and develop wing profiles and structures. I'll be looking up on the ways to "bake in" known data to help with the code. Thank you!
Thank you for the detailed and informative video. Easily saved me from weeks of literature survey effort by providing a review of selected papers across various ML + CFD applications. All the papers mentioned seem like great starting points to delve into the field.
I would disagree that we have enough data for fluid mechanics unlike other popular problems like image recognition. Fluid problems in general are very complex and unique and as soon as one changes certain conditions, machine learning models become useless. But there is a lot of scope for control problems using reinforcement learning which as such does not solely rely on data.
Machine learning does not only refer to neural networks. Brunton's own method SINDy doi.org/10.1073/pnas.1517384113 has proven useful in identifying the dynamics of fluids solely from data. There is an entire field of research dedicated to developing fluid models strictly from data. I would not write it off so easily.
@@John_Graff Thanks -- I agree, there is such a diversity of methods. Although there is no "silver bullet", there are lots of things to try in different circumstances.
This is a subtlety we think about a lot about actually. Fluids data are vast in some dimensions and sparse in others. Many types of machine learning won't generalize well to new parameter values, but there is a lot of work on methods that do generalize. In fact, fluids is one of the big motivating field to develop new and better techniques.
Nice talk Sir. Still, there is a lot more to be done for depicting the turbulent behavior as it's the last Unresolved Mystery of Classical Physics. It's just the beginning of a new era where we are trying to approach the same old problem but with the new tools at hand. Moreover, sir could you please recommend any book that brings together both the topics i.e AI & CFD?
4:38 proof: Business people are clueless and/or . Then there are those that move from tech roles to business. It is required of them to become clueless, bordering on stupid, to progress further. Rant from a burnt employee :)
could you please tell me more about your studies? i have a Ms in CFD and looking for the next step in my studies...very interested in ML, DL and their application in CFD
Thanks for the great video. It's very fascinating to hear neural nets constrained by physics principles. How much of background knowledge of fluid mechanics is needed to get started with this kind of problems? (if the current background is computer science and machine learning)
Some solid math background (ODEs and PDEs) would definitely be helpful. Getting up to speed on the fluid dynamics is a bit of a steep learning curve, but definitely possible. I should have some basic fluids videos at some point soon.
Is there a way to implement pytorch in OpenFOAM? It would be cool if there is a way to use the output of every step in OpenFOAM sn input for neural network so that it gets better with every simulation. And when it is finally trained you can use the neural network as a solver in OpenFOAM again... :-)
As someone who transitioned from FEM to ML Engineering, it was funny at first to notice similarities and think, ‘Hey, I’ve seen that before!’ I believe the reason AI isn’t widely used in FEM/CFD today is because, in fact, it was already being utilized early on, just under a different name.
Hi Steve, I was wondering if you (or anyone who reads this) could recommend me a paper that I could replicate with an accessible database as a first try to get into this field. I have a good background in fluid mechanics but just getting started in machine learning, so any comment would be appreciated! Great content!
Thank you Steve, if you could please elaborate on this and show us some practical examples/applications that we can do using python/matlab it would be great.
Sounds promising but, Is it really possible to construct generalized models for wide range of fluid flows? given that the physics of fluids is very complicated unlike other fields where ML is being applied. But ML can definitely help us observe some generalized patterns in the fluids (like Kolmogrove's 5/3rd law) that we previously didn't saw or know about. ML algorithms can also help us quickly build custom turbulence models for specific flows that we are interested to learn about with a benefit of lower simulation times. But we need a lot of data.
Very good question. We already have a generalized model for a wide range of fluid flows (the Navier-Stokes equations). I think that one promising area is using ML to build better closure models so we can simulate turbulent systems accurately but with much less compute power.
Stumbled across your channel looking to brush up basics of ML in Fluids, coincidentally after reading the review article you co-authored. Super clear and easy to understand lecture! One thing I would appreciate is a references list of the papers you reference in the video in the video description. Thanks a lot for the video!
Thanks this is a very inspiring video. Could ML for Fluid Mechanics be used to train neonatal artificial respiration systems for best parameterization of air flow intensity with respect to the very small sensible patients lungs - to avoid too strong air flow caused by inaccurate manually parametrization ? I learned that manually parametrizion seems to quit difficult with the risk to be not exactly enough so that early born babies lungs have much higher risk to be negative effected by artificial respiration.
Interesting application. I have heard about a lot of work on computational fluid dynamics (CFD) to simulate these various flows. With enough data and "expert" designed solutions, ML might be able to learn some of these solutions.
I’m an undergraduate major in fluid mechanics and my instructor studies both machine learning and turbulent flows. It’s good to know this topic has such good prospects
Would There Then Be A Better Way, Utilizing High Dimensional Modalities, To Fine Tune A Surface To Produce Specific Fluid Dynamic Superstuctures Which Could Be Leveraged To Reduce Drag In Specific Fluids And Conditions?
Thank You For The Brief Summary Of Field. I Enjoyed Your Overview Perspective Of The Basic Subject Matter. The Presented Information Flowed Well Towards Summation And Sufficiently Outlined The General Scope Of Functional Applications. In My Sincere Opinion, Well Done.
Idea - If you define differential equations as loop functions. Ex. y[0] = ...; loop (dydt = y ; y: = y +dydt*dt) then we there should exist loop(loop(loop(...))) functions since that belongs to xyz space. Three loops to initialize a 3D volume.
Hi, first of all I would like to thank you for that great content you are offering for free on RU-vid! I have two questions regarding any kind of reduced order models (built with AE or POD): 1) if you are decomposing the problem into time and space and you say you use "snapshots" of different moments in time, I wonder how do you get the different snapshots at different points in time WITHOUT doing the the high fidelity simulation for the whole timespan? Because in any numerical simulation I know you use one timestep after another and can't skip any timestep... And with doing the whole high fidelity simulation I wonder why would like to built a reduced order model at all? 2) Assuming one can do simulations of different snapshots: How should one space the snapshots values of the high fidelity simulations across the interval of the time/parameter space? For example if I have an intervall from 0 to 1000 and I want 10 modes, do I want modes at 0, 100, 200, ... or unequally spaced points? Is there a literature for this kind of problem? I would be very thankful for an answer and looking forward hearing from you! Kind regards, Daniel
Absolutely no useful information on how one would actually implement any of this! It would be useful to actually show what kind of training data was used from CFD simulations - what time-slices, what ML technique ... etc with an example - even with the shallow bowl example.
Great lecture. Can someone please give insight on how industrial CFD will be changed with ML progressing in fluid mechanics. How prepared an CFD engineer should be in future?. Thank you
In the example on extrapolation at 23:10 we clearly see a tendency from snapshots a to b, as from the extrapolations b to c. In e) we see clearly a sudden increase in the flow field. My understanding is that this could not be predicted because the snapshots represent only a part of the flow, so unseen fluctuation affect what is visible in this limited windows (more so if these are 2d images of a 3d flow field). This makes sense?
Is there a structured way of approach for learning this It seems really interesting Although i am more of solid mechanics person simply becauseit is more predictable than fluid mechanics But this video makes me learn both machine learning and fluid mechanics
Excellent video. Can you please make videos solving problems using ML? I'm asking for Live programming recording. This will be a great help as it will focus on "how" simulations are performed. Cheers
Great work. What is your opinion about the Material Point Method(MPM) to model high-level behavior of fluid dynamics? Ppl are modeling fluid motion behavior using this method in recent papers.
Thanks a lot for the nice lecture, really appreciate it. Could you please let me know how can we try to solve parabolic PDEs or try to predict if the system can show the Turing patterns or not? Could you please let me know from where can I start to know how to solve PDEs in ML? I have so many questions that it is embarrassing for me to type all of them. Once again, love your lectures. Thank you.
This is indeed the tip of the iceberg. This topic has been discussed for the last twenty years in the field of turbulent combustion. I hope you mention them in your new review paper.
Really amaaaazing Lecture! I am a newcomer in this interesting field where fluid mechanics meets ML. Any suggestions where I can get more materials in this field?
This was wonderful to help me in describing my current work challenges with ignorant executives. Beautifully done. The best by far I've ever seen. Kudo's to you!
Thank you for this video. This helped equip me with at least a rudimentary knowledge of the topic that enabled me to ask the questions I needed to ask. You helped me get a job.
Steve, can you tell me what is the system you are using to do your presentation where it looks like you are writing on a transparent board on the other side of the board but we can see it the right way? It looks so cool.
When I first took a lesson of feedback/trained neural network a quarter century ago, I didn't think that it could incorporate existing knowledge. Thanks to this presentation, I learned that it has actually been done.
Thank you so much for providing such an understandable integration of the concepts. I am a Clinical Psychology Ph.D. student (with a career history as a commercial pilot). I am fascinated by the combination of fluid mechanics and psychology (the invisible biology) and what machine learning can provide to the field of psychology in the near future. The psychology field is behind... way behind, on organizing the quant and qual data into usable models--but this gives me a little hope. :)
Yeah you get a result but any fluid dynamic person would say that's useless. Generating pictures with ML is great but for engineering you have to rely on the results
