How Well Can DeepMind's AI Learn Physics? ⚛
HTML-код
- Опубликовано: 26 сен 2024
- ❤️ Check out Lambda here and sign up for their GPU Cloud: lambdalabs.com...
📝 The paper "Learning to Simulate Complex Physics with Graph Networks" is available here:
arxiv.org/abs/...
sites.google.c...
🌊 The thesis on fluids is available here:
users.cg.tuwie...
❤️ Watch these videos in early access on our Patreon page or join us here on RUclips:
- / twominutepapers
- / @twominutepapers
🙏 We would like to thank our generous Patreon supporters who make Two Minute Papers possible:
Aleksandr Mashrabov, Alex Haro, Alex Paden, Andrew Melnychuk, Angelos Evripiotis, Benji Rabhan, Bruno Mikuš, Bryan Learn, Christian Ahlin, Daniel Hasegan, Eric Haddad, Eric Martel, Javier Bustamante, Lorin Atzberger, Lukas Biewald, Michael Albrecht, Nader S., Owen Campbell-Moore, Owen Skarpness, Robin Graham, Steef, Sunil Kim, Taras Bobrovytsky, Thomas Krcmar, Torsten Reil, Tybie Fitzhugh
If you wish to support the series, click here: / twominutepapers
Meet and discuss your ideas with other Fellow Scholars on the Two Minute Papers Discord: discordapp.com...
Károly Zsolnai-Fehér's links:
Instagram: / twominutepapers
Twitter: / twominutepapers
Web: cg.tuwien.ac.a... 
Наука
"We cannot write down the mathematical definition of a cat."
I've found my life's mission
But first of all, you need to define what's a cat
@@sephypantsu it's easy. A cat is something that transforms like a cat. If you know what I'm saying
Nadiya Yasmeen Ironic that most “tensors” in machine learning don’t actually transform like tensors
@@Virsconte can you elaborate on this? To my knowledge the tensors in machine learning are tensors, so they transform like tensors. A lot of them transform like normal matrices, but afaik normal matrices are still tensors.
Non-linear transformations are still transformations.
As a simulation engineer, I'm not even mad that I will be made redundant by neural network. What a time to be alive.
Soon we will miss the days when video games used to be run on real physics simulations, rather than the more efficient approach of a neural network hallucinating something that looks plausible to the players :D
(...which brings up some scary metaphysical philosophic questions, if you think about it...)
Amy Kukleva which questions for example?
@@rubenzuidgeest5373 Well, for instance:
"What if the universe is running on a simulation only as precise as needed to satisfy humans? Then, distant stars and planets wouldn't *really* consist of a huge number of atoms, just a simple approximation of what a fully-simulated planet like Earth would look like. In a lot of theistic philosophies/religions, the world exists as a sandbox for humans. And some have speculated that we might be in a simulation ourselves - as the Simulation Argument goes, the first intelligent beings will be much fewer than all the beings they end up simulating as experiments. If that were the case, why go to expense of simulating the entire universe, if you could get away with a whole lot less? This would make the world a whole lot "smaller" and less complex than we think."
Someone needs to write the physics simulation that the neural network will train on...
@@eaglgenes101 Until
artificial consciousness is achieved.
- Even more ramps! More!
- Sir, we're reaching critical ma--
- I SAID MORE! MORE!!!
Kylo Ren: More! ... ... ... MORE!!!
@@Unethical.Dodgson, he is simp
@@ноуноу-и3ф haha you said the funny simp word funny joke
A resonance cascade is inevitable!!!
Throw more dots, throw more dots, more dots, more dots, more dots! C'mon more dots!
K stop dots.
This is the best channel for being intuitively engaged in AI advancements
What a time to be alive xD I love it
Check out Yannic here on RUclips, you'll probably like him too :)
@@rbain16 I already follow him... And watch all his latest content ... It's really helpfull 😊
I subscribed after reading this 😆
Can't wait to see this tech done in real-time in videogames. Videogames doesn't need 100% accuracy, just "good enough" accuracy and this seems like a great candidate to do it. I'm sure it can run very well on today's GPUs with AI acceleration
Yeah! or any 3D artist for that matter, VFX and 3D illustration or motiongraphics don't need 100% accuracy either, having an interactive simulation to iterate over the final product is way more valuable, the newly introduced optix architecture and realtime denoising features are praised alot in the blender community.
The next gen Xbox Series X and PS5 use AI to achieve real-time raytracing. It really is quite mind-blowing!
Yeah this is perfect for videogames. I'm sure those neural networks will be used intensively for physics and intelligent NPC behaviour in the future in games. So those AI accelerators of modern GPUs like the tensor cores should come in very handy.
@@Greedygoblingames from what I understand the real-time raytracing is the real deal done in the gpu hardware, not AI based. Still amazing
@@9a3eedi RTX does use AI denoising, aswell as new GPU hardware.
the new UE5 uses something different and still achieves photorealistic quality, maybe we'll see a mix of techniques.
"What a time to be alive" - you must be the only person left who uses it unironically!
Well, he is happy to be alive while it possible in current year, lol.
@@minskghoul There's a good way of thinking about it! lol
U ppl underestimate things a lot. A for pandemic goes, COVID-19 could claim fewer ppl's life than any other pandemic of such magnitude (say 1918 or 1968 flu). And all this is because of our understanding of Medical Science and lots of public health measures taken by various govt. Not to mention we've made quite a few potent vaccine candidates in 1/10th the normal time. There are many more inventions like new molecular techniques such as IAMP that is faster than PCR and might make life easier for all molecular biologists.
Look don't watch the news only. News channels are professional scaremongers. Take a breath and watch this:
ruclips.net/video/SJUhlRoBL8M/видео.html
@Laughing Out Loud For COVID-19 u can read WHO situation reports and see worldometer.com graphs. If that feels inadequate u can always scan BioArxiv headlines.
@@aniksamiurrahman6365 True, people always see that bad in the world but it's actually getting way better than it ever has been
"I heard you like neural networks and fluid simulations, so i put a neural network in your fluid simulation"
is it the same Neural network than neuronal network?¿
@Maya :3 I do not know what a "bad pickup line" is 🤷🏾♂️
smh you missed the yo dog
"It's an older meme sir, but it checks out"
... so you can simulate fluids while you neural your network!
everybody at 3 AM: sleeps
me at the same time: mm how AI can learn physics?
by comparing multiple exemple behaviours
It's 1 am and I wonder how I ended up here
:D ikr
The possibilities of this with various science R&D has me feeling electrified. Amazing!
Pretentious comment
I think its basic research, not RnD.
I agree, this is more than just cool graphics, it’s helping understanding physics- our own reality- at a fundamental level.
@Kyler M If you don't know the difference then there's no point in talking to u about it. But I doubt ppl like u r the reason why basic science research is so underfunded.
as its mentionned the aplications are for generalisation purpose not for research or experimentations
I would love to see a capital city getting liquefied like a version of a city getting blown to bits in a movie.
Same.
Or a nuke turning everything to dust
soldier 617 me want big boom boom
Kargadan ^ this comment wants to start some silly beef
@Kargadan cant tell if your joking or not
Here we go again, I didn't search this, however clicked and watched the whole thing....
no way, you've discovered youtube recommendations, and you say you watched it the whole way through? I commend you sir. what a time to be alive
@@Jude- its the youtube AI that recommended
@@omnianti0 how incredible! I've never heard of something so intriguing and curious as this! how did you acquire said knowledge, pray tell
@@Jude- here on youtube and wikipedia
@@omnianti0 so you're telling me that *here* on RUclips *AND* on Wikipedia????? This goes too deep, I've uncovered a conspiracy
"And we are even getting paid to do this ! I cannot believe this " 😂😂😂
are we ?
WHERE I SIGN?
-Video gets recommended to me
-I watch the video
-Me the whole video: What?
-Me after the video: What kind of PC do you need for those simulations?
Which ones? You actually need quite a bit less hardware for the AI simulation then for the actual simulation.
@@donovan6320 in another video its mentionned the pc have 4gpu i guess some of the most expensives
@@omnianti0 No that just heavily depends on the workload. Not to mention depending on the software used they might be utilising the GPU or the CPU... If it's the GPU well then it usually goes a lot faster but the CPU the GPU is mostly there for rendering and that is the bottleneck.
@@donovan6320 this channel released a video of this kind with the framerate and the particule count and they mentionned the new software optimisation for gpu
@@omnianti0 Ok? There are many approaches to running a simulation? Some run on the GPU and are highly parallelised, Some run on the CPU but are multithreaded. It depends on the approach used to simulate it.
Guys, game’s with physics systems like this are going to be sooo insane
Half life source engine 3?
If valve isn't thinking of this, someone should just give them the Two Minute Paper channel and they should start from here :)
@@creeperhack1293 I'm fairly certain that they absolutely are thinking of this. What I understand they're looking way into the future of tech and gaming.
Also did you see the newly updated bottles in Half Life: Alyx? The liquids sloshing around inside them are AMAZING and I'm halfway convinced that they use some AI to do the simulation in real time. So maybe no need to wait for Source 3, maybe we already have a testbed for this tech in Source 2.
Yes, but only if it's cost effective to implement. Right now physics-based simulations are extremely capable, fairly optimized, and very well understood by game engine developers, and they're getting more powerful each day because graphics cards are getting more powerful. The cost of implementing a machine learning solution (or even hiring engineers that are familiar with machine learning techniques) is probably fairly high, so for most game developers that happen to develop engines (Valve, Ubisoft, EA/DICE, etc.) I'm not sure it would be worth it to start baking this into their engines unless they just have money to burn. Crytek has always been on the cutting edge though so if anyone would do it, it would be them. Nvidia is probably looking into it too.
@@creeperhack1293 Valve has implemented a shader effect that imitates water in a bottle *without any simulation.*
They're definitely into awesome stuff like this.
It's wonderful to hear that companies are looking into this stuff. I can't wait to see what comes up in the next decade. Just imagine. Also I've taken a look at the bottle shader from Half-Life Alyx and it does look really good but I can't help myself to imagine the possibilities beyond water; like fire, wind and putting all of that together within a ray traced
environment then have all of it react physically correct while you manipulate everything. And crazy enough, we're probably pretty close to being able to do that. Really exciting times.
It can even simulate turbulent flow! That's amazing.
no that's a paradox
@@GrueneVanilleWaffel What's the paradox
@@GrueneVanilleWaffel what paradox dude
@@darshandev1754 our current computers can't even simulate real entropy.. therefore also not turbulent flow
When we will have real answers for QM, then maybe we will get answers about turbulence.. until then , good luck.
You are absolute amazing for sharing this. For some of us, we can't fathom understanding it at the most basic of itself. But with you showing the 3d dimension test gives us at most understanding of what you are doing and how it actually will help.
This would be really great in gaming. Neural networks could generate and adapt moderately-accurate liquid simulations on demand, and correctly respond to changes (i.e when a player’s character enters a body of water)
that 4:54 Configuration applied in that lifecycle is so beautiful. Thank You for bringing the beautiful equations to life.
Please don't ever stop your channel is an amazing niche
I remember how 8 years ago we were fascinated how you could super scale images with NNs. Incredible how it advanced.
This guy is so enthusiastic you can almost hear him self-high-fiving in the background.
"Absolute witchcraft"
This is also very interesting for the game industrie. They struggle because good simulations can take minutes for one frame to compute. But with AI, we can get simular results in milliseconds or less
I was just thinking this. We may find that real-time fluid simulations end up being too computational expensive with traditional GPUs. We may end up offloading the bulk of the calculations with dedicated neural network chips.
At this point I won't be surprised if AI is awarded the mathematics prize for solution of navier-stroke's and turbulence.🤔
possible.
I don't think AI would have an issue finding a numeric solution. The analytic solution is the larger issues. Also, without any assumptions would be very difficult to quantify.
@@benlev3375 Its a very good point!
@@benlev3375 It's also a problem for numeric solutions as we did not prove yet that Naviers stokes in 3d always have smooth solutions, so an iterative algorithm would return a solution but there's no way to know if that's correct or just related to the iterative algorithm used.
Dev: What kind of simulation you can copy?
AI: Yes.
imagine an AI that get so good at physics that it starts simulating real physical laws that we didnt know about.
True, but generally speaking it would be possible. There is nothing a PC cannot computer that a team of scientists could, theoretically speaking.
I really think a computer will have a real hard time, with abstract thinking and that sort of intuition us humans have. I don't think we will have a AI that can develop a solid theory, that might or might nor explain reality at its roots (think string theory or relativity). As of now i think an AI would try to recognize patterns in the irregularities and try to recreate this as best as it can via trial and error and would develop some sort of general rules and lots of exceptions. Unlike humans don't mind dealing with 200 rules for 200 variations of a certain event. Just as the AI in the video hasn't found a general solution for navier-stokes but rather guesses via trial and error. I think its probably similar to how you can easily predict the path of a thrown ball quite accurately, without doing any calculations. You just develop a general experience based "feeling" for how the starting conditions and the results of certain scenario are connected.
AI fundamentally operates on a set of parameters (input data), historical results/data (neural network), and an algorithm to calculate/tweak/optimize through repetition (e.g., trial and error). Any results you see from AI/machine learning is done through an astronomical amount of repetition and not much else. It's fun to think about AI inventing/discovering abstract concepts and theories, but the exercise is futile. We're quite a long ways from that possibility.
@@vothka205 The mathematical model isn't literally given to the AI, hence you don't need quotes. You're giving it the input/outputs of a mathematical model. The AI, without ever seeing the inner workings of said model, can end up replicating its effective behavior anyways. It'd be an engineering feat for sure, but you could create an AI feeding it real-world observations and see how well the AI can model the physics.
2:19
I can actually see a skull in there.
@@kicka55 Rorschach?
0:51 man just casually references something that he did *300* episodes before. Weird flex but ok.
can you make a video for discuss this
- how many ramps you want?
Dr.Káloly ZsoInai-Fehér - *yes*
"When you know this series very well you know that this video wont take 2 minutes"
only 2minute per frame
This is the type of channel that fades in and out of your recommendations every few months. I'll never subscribe, but boy am I happy when it fades back in to my recommendations
"Finaly, the million dollars reward for solving Navier-Stockes equations will be given to a computer !"
@Abhishek Patel You may consider that I was joking.
The MLFLIP vs FLIP comparison (@2:04) shows the learning algorithm to be only about 13% faster because it is calculating one eighth the amount of raw data. The comparison is 160x150x50 vs 320x300x100 in volume size. If you do the math with the different step sizes the MLFLIP only beats the physics simulation by a small margin.
Mind you, I still think it's a good idea if it saves you hours of calculation and tuning.
When you calculate a physics math in order to generate an image you take a lot of time, however when you imagine something you pick up data that is already computed in previously random simulations and is stored in a memory then the algorithm instead of calculating will just put the pieces in their places where they are needed
That is a simple yet awesome explanation. However, do the Neural network have a memory?
@@_sky_3123 training neural networks works by altering the neurons' weights, so they dont inherently have/need a memory.
I loved, how he said we are paid to do this. Witchcraft.
He loves his job and I love watching his job.
Please keep going.
So if I understood properly,
Essentially, what we see is an AI that "Animate" each frames "by hand" to make an entire scene that looks like a physics simulation?
And it can do it somewhat reliably and realistically in different environments?
That's actually insane
This is the approach I was proposing ever since nVidia released RTX (which is basically the same thing applied to solving ray tracing problems instead of computational fluid dynamics ones). And yes, this is insane. It could also be used in the reverse manner: to determine whether the physical simulation is good enough or not.
I think the AI is learning, from training data, the forces (gravity, friction, impact) applied to each particle, and deriving its own application of those forces. Physics-based simulation differs from frame-animation in that the result mimics real physics using real forces.
@@ideallyyours Most likely it becomes very proficient at understanding patterns. In other words, it observes how the particles tend to behave in certain situations and environments, then it tries to picture similar behavior for the NN approximated particles.
Just like how neural networks learn what people's photos look like and then put together faces of people that never existed.
An advantage of ai physics would be next to zero physics bugs and other interactions in any physics based simulation. So when Ai playing hide and seek start manipulating physics problems that we wouldn’t normally be aware of, it would likely be far less of a problem with ai physics.
Me: oh, this should be quick, it's a two minute paper!
Video: 7 minutes 17 seconds
Me: -_-
can you even remember the time when Deepmind was the most impressive learning generative program? these difficult processes for people to do by hand were solidified years ago, but language is only starting to get as convincing as these simulations
I'm glad that videos by "two minute papers" average on 4x the length advertised. There's plenty of neat stuff you talk about
Normal people: "wow this technology is so cool and has so many potential applications!"
Me: *frantically trying to figure out cat equations*
6:20
every time i
think about this
Shivers...
run...
down...
my ...
spine.
YES!
This NEEDS to be in Blender.
YES
A Simulation of Liquid-Crystals that periodically take varying forms, then create multiple neural networks that are projecting information to one another through these liquid crystals, and boom now we can simulate life
It is really amazing, but still I observe that in the last animation a few blue particles behave somehow nonphysically. They drop too slow. How to make sure the neural networks capture the physics correctly seems a really challenging problem.
The prediction based route might actually be the ideal one here, not just from a processing standpoint but from a random standpoint. Most simulations start with a definitive number on intensity or gravity, meaning that it's expected that two identical simulations would behave the same, if under the same variables.
A prediction would accidentally simulate the minute randomness of real life particles, if trained enough to try and behave like its simulated variant.
"Studying the Navier-Stokes equations in college days"
WOW
In the paper it actually shows that this new method is not faster than the original simulation. Maybe we'll get there in the future, but for now it is just a proof of concept that simulations can be done using AI. By the way, the ground truth simulation is a smoothed-particle hydrodynamics simulation (SPH), not a "real" DNS fluid simulation. These SPH simulations are often used for movie effects for example
The inherent "understanding" the neural network seems to posess of the underlying physical principles is insane. It makes sense that it would be able to replicate known motions but the fact it predicts new situations (additional objects, different sizes,...) accurately is astonishing. I'll have to take a closer look at that paper.
"Absolute witchcraft, and I'm even being paid to do it" made me laugh!
Yt videos during the day: Channels i follow, gaming, music and so on
Yt videos after midnight: ...
Hey, like always a very interesting Concept. Could you predict if there is any solution to use this method in render pipelines, like in Blender, Unity. Unreal?
These AI simulation might be great for Games at some point, especially if the Game physics are totally different from the real world.
I love the Hungarian accent. Reminds of Stephen Fry's stories about his grandfather
"absolute witchcraft and we are even being paid to do this, I can hardly believe this." 😂 that's a passionate man right there.
I don't know much about this but it is interesting. I thought fluids behaved in a chaotic manner which made it difficult/impossible to simulate. I thought the 3-body problem was at the bottom of it all. It's worth looking more into this. Thanks for the video!
When you work with computers most of your life, you apparently end up sounding like one.
Its cool when you realize that something similar is going on in our brain.
We have a pretty good intuitive sense of how the state of a fluid simulation should evolve over time, even if we don't understand the underlying physics. For that same reason our own intuition will only ever remain an approximation, and I'd argue the same is true for the learning based methods described above. There are plenty of domains where this is more than sufficient though.
me : * thinks about domino track *
my brain : * modem noises *
also my brain : hmm, yes, this track won't work, one domino is too far apart
me : * then correct it *
my brain : * concrete on concrete scraping noise *
my brain : ok, i have corrected 256 dominos
me : * okay, run the simulation again *
my brain : * modem noises *
my brain : * shows me a simple 3d representation of what it would look like *
me : * perfection, now i go- *
my brain : * shows me a really cool game that doesn't exist *
me : * i gotta make TH- *
my brain : * shows me water instead of dominos *
me : * modem noises *
my brain : 404, couldn't find W_P_F, shutting down
me : square !
From data analysis to computer vision, area after area is being replaced by AI. Now even simulation... what an amazing time to be a Machine Learning engineer.
The "Thesis on Fluids" needs a way to calibrate the force fields. The results are not identical, not even visually the same. But you might be able to use region or density specific force models. For this video, can you at least start reporting some quantitative measure of the differences between the full simulations and the learned visualizations? Then you will know if you are getting better or not. Eyeballing it won't do. Thanks.
they also use the same Ai for making the video its generalisation as mentionned
@@omnianti0 Thinking about it more, matching to other models is counterproductive. Just go straight to reality, determine error in an appropriate way, and drop all that old baggage. This seems to be a good approach, but it needs to try harder real-world problems. Industrial chemistry, transportation, energy conversion, process fluidics, atmospheric plasmas, combustion, flow acoustics, propulsion controls. Not toy problems. Games are OK, but running refineries and power stations are just more complex games - with measures of value, reliability and efficiency. You learn a lot when you try real problems and tackle them seriously with intent to find workable and sustainable solutions.
@@RichardKCollins ok but Ai never give accuracy it ever about aproximation or mass repetition if you stack refinerys or want a gloabal design ok but not the final
i guess it good for 3D printing since their is tolerance in the process
@@omnianti0 My first AI graduate course was in 1968 when I was an undergraduate. Algorithms for measurement, classification, and control systems of all sorts have to learn and improve, or they are useless.
You have a video referencing Mars. Think of pooling a few hundred million computers and their owners to do a complete model of the planet Mars and all stages of colonization. Linked to all associated technical, scientific, economic and social groups involved. Cheaper to test alternative strategies virtually first, than waste decades with physical experiments. Track and summarize all that is known, outstanding issues, progress; raise funds, develop new technologies, form groups for specialized projects. Not a game, a "User owned global community" for a real purpose. A global project that lasts for many decades and has billions of participants and viewers. Success shared by user owners. This level project needs all that humanity currently knows, and more. Great challenges, real world problems. All existing data and sensor streams, playing ultimately with real people, real tools, real problems, and real outcomes. On earth, plenty of real problems that need algorithms refined by groups of not just a few, but a few million.
@@RichardKCollins im not graded but i understand their is more important problem to fix on the planet before spending lot of resource to make a museum or to colonize remote planets = it can be human security health communication and more than all that the creation of a new language optimised for human physiologiy because it make few sens to relly on multiple empiric diferent language and meusurement standard for coordinate a global project
This technology is actually useful on predicting the global weather. Just input the right variables for it and running it in a very powerful computer. I can only imagine if you want to explode a firework with outputting a drawn square-shaped pattern (not a typical circular explosion pattern) and let the AI dynamics simulator backtrack it for you and tell you how to build such firework for example.
5:33 - this "passing messages" approach is basically how elementary particles interact with each other, or spacetime interact with itself.
AI prediction is so fascinating to me, because there's just that tantalizing promise that with the right breakthroughs in AI, quantum physics, both, or something else, we really could break the seemingly absolute law that the future is unknowable and unpredictable. I hear quantum physicists are breaking causality these days, maybe the solution to the problem of being unable to observe without altering is to observe, then predict, then send the information backwards in time.
So basically you ask the computer to pour water in a glass box with no open ends. THIS IS AMAZING
Learning how to write a code for an AI to do a Physics simulation for a certain scenario is amazing, developing and learning how to make the simulation even better and make it look like real life will one day have a really big use, both for fun(for video games) and for further scientific experiments that otherwise would not have been possible without the incredibly precise animations. These methods and algorithms for learning AI how to do something just like in real life should be i think one of the most focused future programs in science because it will most certainly help science understand universe properties in the most precise way possible!(Imagine how cool would it be that in some point these algorithms are applied to video games! It will basically become like a virtual universe that will contain anything that programmer coding it wants! We will basically make our own universes sandbox!)
yes this channel has 1,6 million subscribers, yet I believe is so underrated.
It'll be cool to see these technologies be implemented in gaming, especially since a lot of companies that design microprocessors want to find ways of adding support for better machine learning processes. Since calculating simulated particles would be nearly impossible to do realtime, a neural net could come much closer to that reality, especially given how much less processing power is required.
I think that realistic-ish fluid simulations that run fast are amazing. This tech could offer some really cool experiences in games and whatnot.
Imagine having a video game where every objects interaction is calculated with AI physics simulation, instead of pre-rendered / pre-calculated physics.
It would need to observe pre-crafted narrative and interactions and then learn how to make new ones that are enjoyable and make sense.
These simulations are so satisfying to watch.
"We cannot write down the mathimatical definition of a cat." is such a beautiful quote for all mathematicians and physicians.
I think it should go "We cannot *yet* write ...".
can't wait for future papers to simulate an entire universe on a quantum computer
Maybe you're in a simulation running on a quantum computer right now?
Im more suprised by the realistic looks of that rendered glass!
The mentioned limitation, of this technique, seems more like a footnote. If it can't generalise to a specific type of scenario, just train it on that type of scenario. For example, if it can't generalise to solids; just include solids in the tiny amount of training data.
This work is very close to perfect, and I also couldn't be happier with it.
It probably won’t be mathematicians that solve the Navier Stokes equations, but rather, super learning computers.
As a mushroom cloud engineer this allows my team to do a lot of optimization on computing cluster which is a magnitude cheaper compared to our previous process based on real experiments.
I love the amount of progress we have made, that we can label what is essentially the prediction of a mathematical model, as the _ground truth_.
That sediment transport simulation may be the most beautiful thing I've ever seen in my life.
"we already know how to write down the maths for this simulation"
Navier Stokes Equation: Allow me to introduce myself
That's really good, but you skipped over the one where it didn't work properly! The hourglass sand timers. The particle sim was on the left, and the AI was on the right, most likely, and the AI made the sand move in a way I've never seen in an hourglass. @5:00
This is honestly so incredible to see. I'm super interested in AI but i've never seen this kind of AI before and it just blows me away...
I am around 20 years early, but what i take from this is that with AI accurately predicting physics, the processing demand can be cut down to the point of being allowed to be run real time, maybe eventually for video game graphics.
The thing about physics simulations is they're not necessarily very accurate for a wide variety of cases. This can be especially true for detailed properties of fluid flow. Computational error exists and is almost always compounding, and in reality we have finite compute power so most problems are done by verifying a simulation on a known physical test and then simulating nearby cases. In verifying the validity of a case you're basically using the trained neural net (your brain) to discriminate good and bad simulation results.
A sufficiently well trained person can identify a good airfoil shape by eye about as well as CFD. It may be that by training neural networks directly on shapes and their properties you can forgo characterizing the flow in a lot of cases.
I have no idea what the narration is talking about but I got interested in AI physics now.
will you do a video on gpt-3?
I've always wondered why actual physics formulae weren't applied to simulations. Great vid!
imagine this being used in games, sand moving how it would in real life when stepped on/interacted with... digging a trench will cause it to be filled with water if it rains... and just anything you can imagine...
Bruh your channel is so amazing, I look space shit to keep hope in humanity. But you highlights a paper that actually published and can be used a few years in the future which feels humanity actually made progress and not only dreaming
I'm thinking that this could revolutionise physics simulations in video games.
Like DLSS it could use data learned and apply it in real time. The data simulations could be super detailed to make the in-game simulations realistic as well, at the cost of little GPU power.
I can't wait to see what happens a few more papers down the road.
Can’t wait for this kinda stuff to be in video games.
When you said this was useless I just about started yelling at my computer lol.
That was an incredible video, well said! Thank you for the awesome content!
This is amazing. We are probably only a decade or so away from making accurate simulations of reality. The video games of the future are going to be uncannily vivid and immersive.
Imagine video games where instead of rendering geometry and lighting for the scenery and characters they are just put in using an AI algorithm instead, which would cause them to look super realistic and would be less demanding on hardware.
All your examples have (input A => ground truth simulation) and (input A => prediction). But what if we train NN on the ground truth with input A and predict on the another input B?
Input A and Input B are space geometries, initial particle distributions with their velocities, etc.
Half of the benefit of this video is the interesting information, the other half is how much this guy frickin LOVES fluid simulations XD
I didn't understand all the numbers and equations but those simulations look cool so I watched till the very end!
From "check out this great new fluid simulation code" to "check out this AI that makes fluid simulation code"
"As we come to rely on computers to mediate our understanding of the world, it is our own intelligence that flattens into artificial intelligence." Quote from a book called "The shallows". We really need to be attentive to what we stand to lose by delegating tasks that demand wisdom on AI programs.
This opens up an opportunity to understand higher dimensional physics