That's more or less the title of my PhD project, but I'm doing particle physics. Lots of amazing work from AI resesrchers happening in that area, but there aren't many pretty visualisations, so you probably won't see it turn up on UA-cam.
@@jacanchaplais8083 Any recommendations I can google or just look up in general? I think that this is literally some of the coolest shit out there, AI learning physics and then us learning from it.
@@prumchhangsreng979 I think by computing time, he meant the time to predict those simulations. And going through the model to predict, while being much less costly than training the model, is still costly. And it's interesting to know how much computing power it needs to do that compared to traditional methods.
Aren't processing and computing meant synonymously 😉 Well, I like the idea of offsetting simulation time to training time and having the predicted simulation to be faster. There were already some physics ML things that actually outperformed standard methods even reported here! I'd not be surprised if this perfoms well here but also would love to see the numbers. 🤓
I think he said 10 - 100x faster. Which is quite the range and I agree more would be good to see in the video without having to follow through to read the paper
When I first met this channel I was like.... Hmmm lil bit odd but... wtvr. Now Im like "OMG YESSSSSS IM HOLDING MY PAPER TELL ME BOUT IT ALREADY" ahahaha
@@KnakuanaRka You can see that for a couple of frames every time when the flag "waves out"/uncurls from the left to the right, e.g. at 2:45 or at 5:03. Use the keys `.` and `,` to step through the video frame-by-frame, if you need it :)
According to the paper (linked under the video), performance was between 11x to 289x faster when running on the GPU versus the standard simulator. Just note that this GPU is an NVidia V100, a $8000 card focusing on AI performance.
Thanks for the great video Two Minute Papers! It's really cool to see that the same neural network models that I'm researching as part of my PhD (graph neural networks) can be used so effectively for faster physics simulations. What a time to be alive!
@@Joe-nq6hy The paper says that this algo runs at 50fps on a (very) high end GPU so I guess it is totally possible :) and Unreal Engine 5 has now a clever solution for high polycount meshes, maybe it can help in these hard cases
What I want to see is this mixed with adversarial machine learning. Train two more AIs: one that 'referees' by trying to guess which between two simulations is the ground truth; and the adversarial AI which takes as input a set of simulation pairs (AI generated and ground truth), and outputs a guess for the input parameters to a simulation that the referee will predict correctly. This adversarial AI will get more encouragement for simpler and smaller simulations. In this way, the adversarial AI should actually attempt to optimize for creating training data that is most efficient (at least in terms of challenge) for the amount of time spent running hand-crafted simulations. Obviously, we take these input parameters and run them through our handcrafted algorithm, then train our physics-simulating AI to accurately predict, up to some level of accuracy.
I had to watch this twice to wrap my head around this. This is absolutely insane. The time it would take for hardware to catch up to some of these legacy simulations would be many years from now. This makes many real-time physics possible in simulations now (or at least, once the algorithms become more available). Crazy.
I often wonder, when comparing the speed of the physics-based and ML simulations, are we assuming the same hardware (CPU, GPU and memory)? And apart from the training procedure, is there any other overhead to using a neural network? Maybe for instance the binary of the physics-based simulation is quite small compared to the ML-based engine. In any case, it's quite incredible to see the efficiency of these neural networks. What a time to be alive!
I found a partial answer on page 17 of the publication. The physics-based simulation doesn't seem to run on GPU. When both run on CPU, the ML-based simulation is already significantly faster (4x to 22x). And the ML-based simulation is dramatically faster when it runs on the GPU (11x to 289x). As for memory consumption, it only says that the ML-based simulation needs 1~2.5 GB of memory but there's no mention of the reference simulation. It's so impressive to see how an AI figured out such an efficient way to simulate physical phenomenons when compared to the hyper optimized mathematical formulas used in those traditional physics engines.
i think i can answer your question if you phrase it a bit more clearly / technically.
3 роки тому+3
@@yarma22 Physics simulations are very fast on GPU indeed, however, softwares use CPU because of the tradition and good fit for all computers. I was spending quite amount of time on solid-fluid simulation paper for ETH, where they spent tremendous amount of time to parallelize computations on GPU. Result, they can simulate a simulation that it takes 4 months on commercial very known software in 3-4 days. If precision is halfly killed, they can simulate same by 9-10 hours. And there was no deep learning at all. The problem of physic based simulations are they are easy to exploit due to n^2 complexity in smoothed particle hydrodynamics. Deep learning is faster, because all it does is an approximation of 3-4 step physical formulation(AFAIK), and doing it via matrix multiplication in the inference makes it a lot faster.
@@yarma22 Holy... I can't wait for these types of methods to be usable in 3D software like Blender. It would be amazing to be able to simulate fluids and cloth even just 22 times faster.
@@anywallsocket Sorry for the lack of clarity, I'm a total noob. I was basically wondering if it was fair to compare the situations based only on the speed factor.
This week I was on a boat. Looking down to the river and seeing the waterflow and the airbubbles underneath and all I could think was “what a perfect simulation, what a time to be alive”! (This is a true story)
I can sense your passion, it’s beautiful lol. This is so fascinating! Pretty sure if I was left alone one weekend I would be glued to your channel and obsessively watch all your content! XxDD
the fluid simulations remind me of those pendulums with multiple hinges. The results start up the same, and they diverge over time. This might means that, even if the final results looks different, you can probably trust the neural network's conclusion.
That's so crazy! Like the thing with FEM is that it's so hard to parallelize. A 16 core CPU is only like 2x the speed of a 4 core CPU. Running that on GPU clusters in tensor core is just so much more compute and would allow you to solve super problems you just couldn't feasible do with FEM.
i know “what a time to be alive” has become something of a meme, but... seriously, what a time to be alive! future AI engineers would kill to experience the pioneering times we are!
I'm writing my master's thesis now on simplified models that run aeroelastic calculations on wind turbines, since CFD is generally too computationally expensive. It's wild to think this entire branch of work will be obsolete in a few years due to innovations like this.
Dear Károly, Thank you very much! Your channel is always surprising to watch ;-) I do not know whether I saw it on your channel, but several years ago, there was a paper about a rigged character that was able of picking up a shirt (cloth simulation) and putting it on by itself. Do you happen to remember this paper? I cannot find it anymore. Thanks a lot ;-)
The thing with the laws of physics is that they are the same in all of spacetime so all you need is 2 frames of a few particles and the network should be able to learn them just as well as with any other data set. There is nothing surprising about a network learning a few basic maths equations.
I know Károly’s voice about a thousand times better than I know his name Love the videos, keep it up Woah cool the channel is nearly at a million subscribers, wow!
I don't know if I'm the only one thinking about that but I can't wait to see the new video games with insane graphics while using the same computers that we have.
This seems like yet another confirmation that neural nets are a good functional approximation of brains. Because when we dream (or imagine stuff), we also run a "physics simulation" in our brain, and it is usually quite accurate and efficient. Edit: in fact for the complex motor skills we have, we must be pretty damn accurate with our "physics simulations"... and "we" even applies to tiny creatures like insects.
I wonder if machine learning could be implemented for light rays as well. Seeing how resource intensive ray tracing is currently. It wouldn't be perfect, just like these physics simulations. But if it can pass the eye test it is good enough for a ton of applications.
What i'm seeing is actually the building blocks for a true human-like AGI one day. We know when a flag is blowing unnaturally, or when honey doesn't drip with the right viscosity, simply by virtue of having learnt the physics of the real world through observation. And here's an AI that learnt and predicted it the exact same way!
Something very important that is rarely mentioned, but should be, is that AI might lead to faster simulations, but the results will always be different from the ground truth. This is fine if you're wanting to use AI for purely graphical problems, and maybe coarse approximations. But to truly test solutions to real world problems, physics simulations will be required. E.G. Your AI might state that a bridge will hold up, but you will want to prove that in a proper simulation before you build a bridge humans will be walking on. If you use AI, and the bridge breaks, it is your responsibility for relying on an unproven black box to make real-world decisions.
I probably wouldn't trust it for real world stuff like testing structures etc but it would be interesting if this could be applied to games in the same way DLSS to speed up games physics.
I'd love to see this applied to movies and TV shows that take place "inside a video game". The problem I always see is that the game world is treated like real life, but the audience wants to see that it's fabricated. Maybe using a Physics-taught AI would strike the appropriate balance.
I have a feeling that the concept of AI build accelerators build into the chip, will make more and more sense in future, with all this AI tech, begin to be used inside games, graphic app, and others, maybe we will have someday, a CPU, GPU, and a special AI Card xD in desktop computers and workstation. I am kinda surprised so many cool AI algorithm exist and they are so underused in consumer software in general.
Not that I have tried a lot, but nearly everything in machine learning that has merit is behind locked doors We will probably never see the programs shown in these videos, the code never released, and all features of them will be limited to the point they are almost no different to conventional programs One would think the world of AI is full of those willing to share their progress and mistakes, but nope not at all Money eventually is what decides what is and isn't produced, and by extent what we are "allowed" to know about those products
@@rustycobalt5072 The code is often open source but the trained data results are not. I can understand that as it may take huge amounts of computing power plus expensive input data. But I agree with you, much of this stuff is behind pay wall's and outside of our reach. I for example would love to see this tech in software such as Blender and Unreal Engine but I doubt that is going to happen any time soon sadly.
@@rustycobalt5072 Learn how to do it and train your own model if you are too cheap to pay for someone elses work or if you are too lazy then wait a bit longer. The stuff that is shown here is cutting edge research.
@@maythesciencebewithyou I've followed this channel for 2 years now but I still don't know what is this field of study called and where to learn it. Some directions would be very appreciated 😊😊
I would like to see real life comparisons. All we have to to is use smoke so the a computer can learn the wind currents in a scene then add in a simulated cloth.
It's amazing how they can sort of condense computer power like this. My mind boggles that one day I might be able to open up blender or some other app and run a rigid body + high res fluid + cloth sim where things interact and do so without processing for a solid year and cooking my GPU.
I want to see some molecular or better yet quantum physics simulated. Imagine, we could have proper proton torpedoes in the new star wars games! Mikor jön a mezonágyú doktor úr? :)
Let's be honest besides all the excitement: the flags are still quite a bit off from the actual simulated output - which can be a big deal when exact simulations are needed or in composite effect scenarios, which should have been mentioned in the video.
in a couple years or so, games will be indistinguishable from reality. just combining all the algorithms on this channel into one engine would provide the means to it.
As a computational physicist I have to say this is impressive. I was wondering if the model is trained on low viscosity, will it be able to perform simulation with high viscosity in the future? Or, does the model support changing physical parameters (gravity, viscosity, temperature etc.) at all? Thanks.
Is this what Nvidia is doing with DLSS? Taking a simulation or game and then creating an AI that recreates via prediction? If not, has this being applied to game play? Or is it not fast enough yet?
*Demonstration of speed gains of the final AI result (after training)?* We just see similiar results (which is impressive) but no speed comparison, which is the essence of the whole concept: sacrifice some accuracy, for tons of speed.
There some strange effect where somehow the ground truth sims seemed far more uncanny than the ai prediction. Some of the ai predictions seemed to have more weight.
Super impressive, but did I miss something ore we're missing benchmarks to see how fast is the AI simulation running compared to the traditional method ?
Honestly with these algorithms and ML models now adays, I don't think it will be long till game engine animations, physics, graphics and characters are all generated by AI and use a fraction of computer resources
I wonder how close we are to getting real time physics simulation in video games that are handled by AI on the compute cores, running alongside the 3D engine and raytracing running on their own cores as well. We're getting close to the point where we could utilize the full power of graphics cards to really push the bounds of what's been possible.
I always like the idea to use machine learning to optimize execute times of animations, physics etc. I would also like if somebody did this for pathfinding algorithm
I would love to see a difference map of those simulation vs ground truth to see how accurate the algorithm really is because seeing just two images makes it hard to spot the errors.
I'd imagine being able to use it as an iterative design confirmation step, for one. It's already fairly common practice to have multiple levels of verification in Systems Engineering, leaving the final test with a physical product (an airliner for example) to the very last, after many tests with simulators at varying lower levels of fidelity. This technique could perhaps be a version of that for Mechanical Engineering, among others.
Real time feedback > Real time simulation in most fields not concerned with safety. Being able to get an imperfect prediction is far more informative than an empty bounding box , mostly empty point cloud.
I am looking forward to the AI process that takes visual/auditory data from real video and converts them into simulation data. This will allow us to automatically scan the real world and treat it as manageable data.
In addition, it would be awesome if that could recognize each element and give some change with a language directive, as in the example shown with Codex.
There are several applications where this result is more than enough! Where it isn't needed to have a perfectly accurate simulation. If it feels real is just enough. Game dev and VFX are the very first two broad areas that I can think of.
Well, they’re trying to replicate the simulation with an AI, so the simulation would make sense as a ground truth. Working off real systems should come two papers down the line. ;-)
Ground Truth when referring to AI-accelerated structures can be assumed to refer to a (previous) non-AI established purely calculated result. Like a high sample ray traced "ground truth" image versus a low sample rate AI filtered one.
I will have to read the paper because for now I have zero idea how this learning algorithm even works. What data do we input except for functions values at mesh points at different times? How would it predict the further movement if we vary some parameters like wind speed? I don't believe it's possible without some implementation of physics equations
idea: teach multiple AIs complex simulations, then have an AI learn how to instantly map arbitrary complex simulations into the eventual AI. that is to say, train an AI to take educated guesses as to what an AI derived solution would probably look like.
when you deconstruct the AI that has learned physics and look at the individual nodes, do they contain the equations of the laws of nature? can you find fundamental laws and/or rules in there? if so, can we use such an AI to obtain a theory of everything?
AI is used extensively in games now. That's what DLSS is. It's how RTX works without being super noisy. And AI driven techniques are used extensively in the graphics apps now that are used to make games and other things. I'm sure that as they develop cool new things like this, it's being implemented as quickly as possible, but it is still new. It takes time to turn a paper into an app. I think the reason this particular thing isn't yet implemented is because it just would take too much computer power. To do this probably requires the entire GPU, which would leave nothing left over for your game, however if used by artists to help create the content for games, it is a game changer.
What a time to be alive!
truly fascinating.
Should be a tshirt.
livin in the future blingin on my hot line
You have to hold on to your paper first
Actually ♥️😁👍🏻
I'm squeezing my papers so tight they are fused to my fingers!
This is the opposite of a problem!
I think you still had the old physics AI on, the new AI fixed paper-finger fusing
same, and i still dropped them all over the place in shock and amazement of the results
I'm squeezing my paper because someone is holding my pen
I couldn't be happier to see a new two minute physics video!
2 papers down the road the title will be, "Can a Machine teach us Physics?".
Well that’s the goal. Make reality easy
That's more or less the title of my PhD project, but I'm doing particle physics. Lots of amazing work from AI resesrchers happening in that area, but there aren't many pretty visualisations, so you probably won't see it turn up on UA-cam.
@@jacanchaplais8083 share papper links
@@jacanchaplais8083 Any recommendations I can google or just look up in general? I think that this is literally some of the coolest shit out there, AI learning physics and then us learning from it.
@@jacanchaplais8083 is that in a similar vein to the work Taoli Cheng has been doing?
Would be interesting to see computing time comparisons with standard methods
@@prumchhangsreng979 I think by computing time, he meant the time to predict those simulations. And going through the model to predict, while being much less costly than training the model, is still costly. And it's interesting to know how much computing power it needs to do that compared to traditional methods.
@@AvastarBin ah so its the processing time. Im interested too xD
Aren't processing and computing meant synonymously 😉
Well, I like the idea of offsetting simulation time to training time and having the predicted simulation to be faster. There were already some physics ML things that actually outperformed standard methods even reported here! I'd not be surprised if this perfoms well here but also would love to see the numbers. 🤓
@@ewerybody i check dictionary and i mixed up the word compute with program(verb). :/ i was wrong hm
I think he said 10 - 100x faster. Which is quite the range and I agree more would be good to see in the video without having to follow through to read the paper
Love the little "hold on to your papers" emblem 😂
When I first met this channel I was like.... Hmmm lil bit odd but... wtvr. Now Im like "OMG YESSSSSS IM HOLDING MY PAPER TELL ME BOUT IT ALREADY" ahahaha
I can’t wait to tell my grandkids that “back in my day, physics ran on the cpu and fluid sims could only do 3 frames a second.”
3 frames a second? What kind of supercomputer do you have? XD
Don't worry. You won't have grandchildren. Lorde KIaus Schwab of the 4IR shall not permit it.
@@shin-ishikiri-no What?
I love how the "ground truth" flag is clipping through the pole and the AI is replicating that exact bug.
Where does that happen?
@@KnakuanaRka You can see that for a couple of frames every time when the flag "waves out"/uncurls from the left to the right, e.g. at 2:45 or at 5:03. Use the keys `.` and `,` to step through the video frame-by-frame, if you need it :)
Now this AI could literally "hold on to it's papers" since it has the physics to do so!
Is that viable for real time simulation? I would have liked some more info on the performance. Great video!
Yes! the whole point of compressing all the nuance into the training phase is that you don't have to run complicated calculations at runtime.
4:59 "And it can do it 10 to 100 times quicker."
According to the paper (linked under the video), performance was between 11x to 289x faster when running on the GPU versus the standard simulator.
Just note that this GPU is an NVidia V100, a $8000 card focusing on AI performance.
But whats the frametime? Is it already f/s or still s/f or m/f
@@Soul-Burn seems in my budget
Had this idea months ago you could even create an entire physics engine built on this premise to make it run certain interactions faster
Same bro
Cant wait until this sort of stuff is real-time
Two more papers down the line
When it gets real time it will be the time to be alive!
If it is 100x faster than simulation it should be faster than realtime already because C4D can run most of those simulations almost real time.
cloth simulation is already mostly real time
@@vladimirtomin8223 neat
Thanks for the great video Two Minute Papers! It's really cool to see that the same neural network models that I'm researching as part of my PhD (graph neural networks) can be used so effectively for faster physics simulations. What a time to be alive!
Can't wait to see these swaying cloths in videogames!
@@Joe-nq6hy The paper says that this algo runs at 50fps on a (very) high end GPU so I guess it is totally possible :) and Unreal Engine 5 has now a clever solution for high polycount meshes, maybe it can help in these hard cases
What I want to see is this mixed with adversarial machine learning. Train two more AIs: one that 'referees' by trying to guess which between two simulations is the ground truth; and the adversarial AI which takes as input a set of simulation pairs (AI generated and ground truth), and outputs a guess for the input parameters to a simulation that the referee will predict correctly.
This adversarial AI will get more encouragement for simpler and smaller simulations. In this way, the adversarial AI should actually attempt to optimize for creating training data that is most efficient (at least in terms of challenge) for the amount of time spent running hand-crafted simulations.
Obviously, we take these input parameters and run them through our handcrafted algorithm, then train our physics-simulating AI to accurately predict, up to some level of accuracy.
I had to watch this twice to wrap my head around this. This is absolutely insane. The time it would take for hardware to catch up to some of these legacy simulations would be many years from now. This makes many real-time physics possible in simulations now (or at least, once the algorithms become more available). Crazy.
Wow, just...wow. This is starting to feel like we're at the horizon of the singularity.
"Can We Teach Physics To A Machine?"
Yes, but the machine will probably ignore you.
6:22 the flag just clips through the pole though. So I guess this would be useful for video games where accuracy isn't life or death.
I often wonder, when comparing the speed of the physics-based and ML simulations, are we assuming the same hardware (CPU, GPU and memory)? And apart from the training procedure, is there any other overhead to using a neural network? Maybe for instance the binary of the physics-based simulation is quite small compared to the ML-based engine. In any case, it's quite incredible to see the efficiency of these neural networks. What a time to be alive!
I found a partial answer on page 17 of the publication. The physics-based simulation doesn't seem to run on GPU. When both run on CPU, the ML-based simulation is already significantly faster (4x to 22x). And the ML-based simulation is dramatically faster when it runs on the GPU (11x to 289x). As for memory consumption, it only says that the ML-based simulation needs 1~2.5 GB of memory but there's no mention of the reference simulation. It's so impressive to see how an AI figured out such an efficient way to simulate physical phenomenons when compared to the hyper optimized mathematical formulas used in those traditional physics engines.
i think i can answer your question if you phrase it a bit more clearly / technically.
@@yarma22 Physics simulations are very fast on GPU indeed, however, softwares use CPU because of the tradition and good fit for all computers. I was spending quite amount of time on solid-fluid simulation paper for ETH, where they spent tremendous amount of time to parallelize computations on GPU. Result, they can simulate a simulation that it takes 4 months on commercial very known software in 3-4 days. If precision is halfly killed, they can simulate same by 9-10 hours. And there was no deep learning at all. The problem of physic based simulations are they are easy to exploit due to n^2 complexity in smoothed particle hydrodynamics. Deep learning is faster, because all it does is an approximation of 3-4 step physical formulation(AFAIK), and doing it via matrix multiplication in the inference makes it a lot faster.
@@yarma22 Holy... I can't wait for these types of methods to be usable in 3D software like Blender. It would be amazing to be able to simulate fluids and cloth even just 22 times faster.
@@anywallsocket Sorry for the lack of clarity, I'm a total noob. I was basically wondering if it was fair to compare the situations based only on the speed factor.
Aha! Finally, a paper I have read before seeing your video! I am strangely proud.
This week I was on a boat. Looking down to the river and seeing the waterflow and the airbubbles underneath and all I could think was “what a perfect simulation, what a time to be alive”! (This is a true story)
I can sense your passion, it’s beautiful lol. This is so fascinating! Pretty sure if I was left alone one weekend I would be glued to your channel and obsessively watch all your content! XxDD
the fluid simulations remind me of those pendulums with multiple hinges. The results start up the same, and they diverge over time. This might means that, even if the final results looks different, you can probably trust the neural network's conclusion.
What a time to simulate being alive!
That's so crazy! Like the thing with FEM is that it's so hard to parallelize. A 16 core CPU is only like 2x the speed of a 4 core CPU. Running that on GPU clusters in tensor core is just so much more compute and would allow you to solve super problems you just couldn't feasible do with FEM.
This channel should be named Minute Papers.. like Minute Physics. That would explain this 7-min episode.
getting closer to our ancestor simulation
Holy shit imagine VR in 10 years, we will literally have seperate lives on there.
This is amazing! Btw, you forgot to talk about simulation duration with ai. Is it faster and if so, how much is it faster?
i know “what a time to be alive” has become something of a meme, but... seriously, what a time to be alive! future AI engineers would kill to experience the pioneering times we are!
Thanks a lot, I'm not a big fun of your visualisation videos but this one, FANTASTIC!!!
and let's read a paper!
:-)
Thank you for the amazing videos, having a crap day and these make me feel better!
0:30 this is the most used clip in all of two minute papers
Lmao frrr
What about the one with all AI generated faces?
Cant wait till this gets implemented into scientific research and gives us some groundbreaking discoveries about our universe.
i mean with that smooth voice of yours you could teach machines and humans alike anything.
I'm writing my master's thesis now on simplified models that run aeroelastic calculations on wind turbines, since CFD is generally too computationally expensive.
It's wild to think this entire branch of work will be obsolete in a few years due to innovations like this.
Dear Károly, Thank you very much! Your channel is always surprising to watch ;-)
I do not know whether I saw it on your channel, but several years ago, there was a paper about a rigged character that was able of picking up a shirt (cloth simulation) and putting it on by itself.
Do you happen to remember this paper? I cannot find it anymore. Thanks a lot ;-)
Your enthusiasm is epic as always - loving it 🙂
The thing with the laws of physics is that they are the same in all of spacetime so all you need is 2 frames of a few particles and the network should be able to learn them just as well as with any other data set. There is nothing surprising about a network learning a few basic maths equations.
3:01 AI is now able to simulate the alternative timeline from Day of the Tentacle, where the American flag is swapped for a Tentacle costume.
Would love to know the performance gains with this method
10-100 faster than the training algorithms
I know Károly’s voice about a thousand times better than I know his name
Love the videos, keep it up
Woah cool the channel is nearly at a million subscribers, wow!
This felt like a ASMR of Two Minute Papers.
I am come here see you so early, it make me come here early. Thanks!
I don't know if I'm the only one thinking about that but I can't wait to see the new video games with insane graphics while using the same computers that we have.
I am clutching my papers
This seems like yet another confirmation that neural nets are a good functional approximation of brains. Because when we dream (or imagine stuff), we also run a "physics simulation" in our brain, and it is usually quite accurate and efficient. Edit: in fact for the complex motor skills we have, we must be pretty damn accurate with our "physics simulations"... and "we" even applies to tiny creatures like insects.
I wonder if machine learning could be implemented for light rays as well. Seeing how resource intensive ray tracing is currently. It wouldn't be perfect, just like these physics simulations. But if it can pass the eye test it is good enough for a ton of applications.
What i'm seeing is actually the building blocks for a true human-like AGI one day. We know when a flag is blowing unnaturally, or when honey doesn't drip with the right viscosity, simply by virtue of having learnt the physics of the real world through observation. And here's an AI that learnt and predicted it the exact same way!
Something very important that is rarely mentioned, but should be, is that AI might lead to faster simulations, but the results will always be different from the ground truth. This is fine if you're wanting to use AI for purely graphical problems, and maybe coarse approximations. But to truly test solutions to real world problems, physics simulations will be required. E.G. Your AI might state that a bridge will hold up, but you will want to prove that in a proper simulation before you build a bridge humans will be walking on. If you use AI, and the bridge breaks, it is your responsibility for relying on an unproven black box to make real-world decisions.
I've been waiting for that to happen for so long
I probably wouldn't trust it for real world stuff like testing structures etc but it would be interesting if this could be applied to games in the same way DLSS to speed up games physics.
All of the videos in this are oddly satisfying
only left to know if it is actually faster, as there is on point in approximating a simulation rather than doing it without saving boat loads of time
Physics Rules!
I'd love to see this applied to movies and TV shows that take place "inside a video game". The problem I always see is that the game world is treated like real life, but the audience wants to see that it's fabricated. Maybe using a Physics-taught AI would strike the appropriate balance.
After seeing all of your other videos, Yes.
I have a feeling that the concept of AI build accelerators build into the chip, will make more and more sense in future, with all this AI tech, begin to be used inside games, graphic app, and others, maybe we will have someday, a CPU, GPU, and a special AI Card xD in desktop computers and workstation. I am kinda surprised so many cool AI algorithm exist and they are so underused in consumer software in general.
GTA 7's early presentation of liquid physics.
will we (normla people) also be able to use this amazing technology? will we see it implemented in some 3d software soon?
Not that I have tried a lot, but nearly everything in machine learning that has merit is behind locked doors
We will probably never see the programs shown in these videos, the code never released, and all features of them will be limited to the point they are almost no different to conventional programs
One would think the world of AI is full of those willing to share their progress and mistakes, but nope not at all
Money eventually is what decides what is and isn't produced, and by extent what we are "allowed" to know about those products
@@rustycobalt5072 The code is often open source but the trained data results are not. I can understand that as it may take huge amounts of computing power plus expensive input data. But I agree with you, much of this stuff is behind pay wall's and outside of our reach. I for example would love to see this tech in software such as Blender and Unreal Engine but I doubt that is going to happen any time soon sadly.
@@rustycobalt5072 Learn how to do it and train your own model if you are too cheap to pay for someone elses work or if you are too lazy then wait a bit longer. The stuff that is shown here is cutting edge research.
@@maythesciencebewithyou hmm yes fuck poor people
@@maythesciencebewithyou I've followed this channel for 2 years now but I still don't know what is this field of study called and where to learn it. Some directions would be very appreciated 😊😊
I would like to see real life comparisons. All we have to to is use smoke so the a computer can learn the wind currents in a scene then add in a simulated cloth.
It's amazing how they can sort of condense computer power like this. My mind boggles that one day I might be able to open up blender or some other app and run a rigid body + high res fluid + cloth sim where things interact and do so without processing for a solid year and cooking my GPU.
I want to see some molecular or better yet quantum physics simulated. Imagine, we could have proper proton torpedoes in the new star wars games! Mikor jön a mezonágyú doktor úr? :)
Let's be honest besides all the excitement: the flags are still quite a bit off from the actual simulated output - which can be a big deal when exact simulations are needed or in composite effect scenarios, which should have been mentioned in the video.
in a couple years or so, games will be indistinguishable from reality.
just combining all the algorithms on this channel into one engine would provide the means to it.
As a computational physicist I have to say this is impressive. I was wondering if the model is trained on low viscosity, will it be able to perform simulation with high viscosity in the future? Or, does the model support changing physical parameters (gravity, viscosity, temperature etc.) at all? Thanks.
Is this what Nvidia is doing with DLSS? Taking a simulation or game and then creating an AI that recreates via prediction?
If not, has this being applied to game play? Or is it not fast enough yet?
*Demonstration of speed gains of the final AI result (after training)?* We just see similiar results (which is impressive) but no speed comparison, which is the essence of the whole concept: sacrifice some accuracy, for tons of speed.
There some strange effect where somehow the ground truth sims seemed far more uncanny than the ai prediction. Some of the ai predictions seemed to have more weight.
Super impressive, but did I miss something ore we're missing benchmarks to see how fast is the AI simulation running compared to the traditional method ?
This on headphones during a cross Atlantic flight into Europe would be extremely soothing.
Honestly with these algorithms and ML models now adays, I don't think it will be long till game engine animations, physics, graphics and characters are all generated by AI and use a fraction of computer resources
I wonder how close we are to getting real time physics simulation in video games that are handled by AI on the compute cores, running alongside the 3D engine and raytracing running on their own cores as well.
We're getting close to the point where we could utilize the full power of graphics cards to really push the bounds of what's been possible.
I always like the idea to use machine learning to optimize execute times of animations, physics etc. I would also like if somebody did this for pathfinding algorithm
hey!! this is a big help!! thank U Dr....
I would love to see a difference map of those simulation vs ground truth to see how accurate the algorithm really is because seeing just two images makes it hard to spot the errors.
Apart from computer graphics, what are the applications of this? Would you ever trust this, for example in the stress test?
I'd imagine being able to use it as an iterative design confirmation step, for one. It's already fairly common practice to have multiple levels of verification in Systems Engineering, leaving the final test with a physical product (an airliner for example) to the very last, after many tests with simulators at varying lower levels of fidelity. This technique could perhaps be a version of that for Mechanical Engineering, among others.
Real time feedback > Real time simulation in most fields not concerned with safety.
Being able to get an imperfect prediction is far more informative than an empty bounding box , mostly empty point cloud.
I am looking forward to the AI process that takes visual/auditory data from real video and converts them into simulation data.
This will allow us to automatically scan the real world and treat it as manageable data.
In addition, it would be awesome if that could recognize each element and give some change with a language directive, as in the example shown with Codex.
In your ad read for weights & biases add that they have a youtube channel as well??
now if only we can get an AI surgeon and dissect this AI to make an even better simulator
Truly what a time to be alive I hope I can make it to see games use AI to simulate in real time
Only thing this NN can't predict could be the interaction between my hand and my papers while watching this video!
There are several applications where this result is more than enough! Where it isn't needed to have a perfectly accurate simulation. If it feels real is just enough. Game dev and VFX are the very first two broad areas that I can think of.
Shouldn't "Ground Truth" actually have a foundation in the world of reality rather than a simulation of reality?
Well, they’re trying to replicate the simulation with an AI, so the simulation would make sense as a ground truth. Working off real systems should come two papers down the line. ;-)
Ground Truth when referring to AI-accelerated structures can be assumed to refer to a (previous) non-AI established purely calculated result. Like a high sample ray traced "ground truth" image versus a low sample rate AI filtered one.
Hmmm this paper is very reminiscent of the work coming out of caltech regarding their graph neural operators
I will have to read the paper because for now I have zero idea how this learning algorithm even works. What data do we input except for functions values at mesh points at different times? How would it predict the further movement if we vary some parameters like wind speed? I don't believe it's possible without some implementation of physics equations
I'm curious if this amazing technology could be applied to videogames with Nvidia Tensor Cores of the RTX cards
I wonder if this can be applied to protein folding problems and make those simulations far more efficient?
If we ever want to create true VR, we will have to get the physics down.
idea: teach multiple AIs complex simulations, then have an AI learn how to instantly map arbitrary complex simulations into the eventual AI. that is to say, train an AI to take educated guesses as to what an AI derived solution would probably look like.
If not already, at this rate, we'd get AI that's able to predict raytracing in a few years (or maybe even months).
Now I'm feeling excited by the new tensor cores on the nvidia 30xx series
when you deconstruct the AI that has learned physics and look at the individual nodes, do they contain the equations of the laws of nature? can you find fundamental laws and/or rules in there? if so, can we use such an AI to obtain a theory of everything?
Wie findest du die ganzen Paper? Hast du alle Zeitschriften abonniert oder schaffst du das?
Indeed . What a time to be alive
why isnt this type of technology used in modern game engines?
AI is used extensively in games now. That's what DLSS is. It's how RTX works without being super noisy. And AI driven techniques are used extensively in the graphics apps now that are used to make games and other things. I'm sure that as they develop cool new things like this, it's being implemented as quickly as possible, but it is still new. It takes time to turn a paper into an app. I think the reason this particular thing isn't yet implemented is because it just would take too much computer power. To do this probably requires the entire GPU, which would leave nothing left over for your game, however if used by artists to help create the content for games, it is a game changer.
It feels like a superluminal AI is indeed simulating the universe
Most solids do deform do to outer influences when you think about it.