Artificial Einstein: Did AI just do the impossible?

hi if you're new here I'm Brian keing the Chancellor's distinguished professor at the University of California San Diego and I'll be your Guru today to unlock one of the most interesting aspects of modern age the Confluence of computers artificial intelligence and physics as AI shapes our world it also has come to my attention that perhaps we can use it to unlock new laws of physics new discoveries that could make it a sort of new Turing test that would replace the ordinary imitation game that Alan Turing proposed some 70 years ago Now using AI is has been shown to produce all different sorts of very amusing and sometimes startling and sometimes not too politically correct images and even new video from textto video creation engines but now we're going to venture into the world of creating new laws of nature physical laws to unravel secrets about the origin and physics of our universe so strap in this is going to be exciting and the next few years in artificial intelligence are going to reshape my field of AC emia perhaps replacing me artificial Brian keing with an AI Avatar so I've thought about a few different trends that I want to pay attention to and I'm going to take you along for the ride on this channel the first one is the role of artificial intelligence Quantum Computing both in generating the a description of how quantum computers behave but also exploiting the power the unique power of quantum computers to do new and novel things in physics so AI algorithms are optimizing the design of quantum circuits accelerating the development of computers that can perform calculations beyond the reach of so-called classical computers this Synergy could soon solve complex problems in seconds ranging from Revolutions in cryptography to Material Science maybe even climate modeling and exotic materials that produce new behaviors that we could use for energy generation and solve the climate crisis things like high temperature superconductors we also can see the role of artificial intelligence in unraveling Cosmic Mysteries to understand how galaxies and stars formed the universe is put into a box and that box is giant simulation right now we're looking at a few billion particles and maybe more but with the Advent of AI perhaps we could scale that up orders of magnitude more more Faithfully imitating and mimicking the behavior of these systems and this could actually make a revolution beyond our planet by looking also at vast data sets that no human being not even my brilliant graduate students could ever pour through these AI physicists could uncover new galaxies black holes Cosmic phenomena and put together pieces of the puzzle that we as human brain physicists with natural intelligence or natural stupidity as some have accused me would then be made obsolete because we just simply can't analyze and process the latent data that is present when you have surveys of billions of stars that are coming out or galaxies or regions of the cosic microwave background that my colleagues and I study another realm where we're overwhelmed with data is presents an opportunity for artificial intelligence to revolutionize particle physics particle accelerators such as the large hron Collider have reached their energy limit meaning that they produce so much energy and they really can't cram any more squeeze any more energy out of the existing tunnels there are no new rivals in terms of size and collision center of mass energy for these super colliders so it's really up to us to glean as much information out of the existing data that we can possibly before even proposing things like a muon collider a collider on the moon or perhaps a solar system collider as some particle physicist dreamers have been proposing maybe it'll happen but it's not altogether likely so artificial intelligence is already transforming particle physics it sifts through data at the petabyte scale that again no human could ever do and a lot of what we do is throw away you know 99.99999 events out of a petabyte worth of data you might only have 10 or 100 events that qualify as being worthy of attention that can actually add to the credulity that we have in the standard model of particle physics and maybe reach Beyond it to test exotic model such as the fundamental physics project of Stephen wilfer these exotic predictions can be only mined by these massive large physics models such as my colleagues and I are trying to develop there's another question is whether or not you can design new materials using artificial intelligence that could be on the macro scale making things like new high temperature superconductors or even at the molecular scale doing things like design new types of drugs or nanor robots and things like that to make stronger materials faster conducting materials lossless power transmission for the next generation of electronics another major Trend that has relevance to physics uh discussed with past guest Tim Palmer on the podcast is climate physics having a CERN but for climate physics was Tim Palmer's proposal we need this kind of enormous scale to simulate the Earth's complex climate environment you almost need a computer the size of the Earth's atmosphere in terms of the number of particles that can track and so again no human no graduate student no matter how brilliant he or she is could ever have a fraction of the ability to assess data that artificial intelligence can and already does so a valuable tool and coupled with advances in Computing speed throughput and design perhaps invoking quantum computers it may be possible to really harness the complex physics that we know about which is actually relatively simple we'll be discussing in just a moment the physics that was around since the 17 or 1800s describing particles fluid flow and things like that so the fusion of artificial intelligence in physics is not only expanding our knowledge but reshaping the very methods of scientific inquiry will we be able to learn about new laws of physics perhaps new forces of nature that would we otherwise be blind to if not for the role of the artificially intelligent algorithms that are just coming online right now so let's dive deep into what AI is good for can it just reproduce physics or simulate physics which is very different than actually uh working through a physics engine is it possible for an artificial intelligence to actually Divine or predict new physical laws new forces new Fields New particles is that possible I always ask the question in my audience can an artificial intelligent entity such as a large language Model A large physics model can it reproduce The Sensation that Albert Einstein wondered what would happen to an observer in freef Fall would he experience a gravitational field and of course the answer was no and that led to the Einstein equivalence principle we're not going to be talking about new laws of physics we're going to talk about replicating the physics at the microscale and do you have to do something like CG which doesn't really solve physics problems or do you actually need to solve the equations of motion for particles to create complex phenomena to visualize them like smoke or atmospheric phenomena in uh the Earth's climet system so in this video we're going to discuss a paper that proposes a neural network-based method for simulating how fluid dynamics behaves sounds boring but don't worry tools that I discuss for you today also have an application to fake worlds like video games too so in the words of my new friend Alon musk we could all be living in a simulation but the simulation would have to reproduce things on a scale that humans would otherwise be blind to either means by simulating a huge Universe with stars placed very far apart or by simulating things at the micro scale that we don't have access to with our current technology so today we're going to talk about neural networks and how they were used to train and learn fluid and smoke Dynamics and how can we continue to improve them are they just complex you know cartoons or CGI no they're not uh but there's a limitation when you actually try to simulate the physics versus replicate what physics looks like we're going to talk about how these neural networks drastically reduce the time required for pred predicting simulation and the outcomes of these simulations we'll talk about the neural network itself and how it was able to generalize beyond what it was programmed to learn and handled new and complex simulations beyond its training domain in physics uh graduate school usually you learn an equation called the nav Stokes equation it's actually a set of multiple equations and they're what are known as partial differential equations that are quite complex they describe the motion of viscous fluids they're named after Claude Luis navier and George Gabriel Stokes who independent ly derived them in the 1800s the navier STS equations are fundamental foundational tool in fluid dynamics and they've been used for hundreds of years now to model a wide variety of phenomena including the flow of air around an airplane wing something they couldn't have conceived of in the 1800s the flow of water in a pipe the motion of blood in the human body and even the flow of things like the jet stream the atmosphere the Gulf Stream in the earth's oceans now the navier Stokes equations are beautiful and they can be written in many many ways but the most common way is the incompressible navier Stokes equation that is a fluid like water which is not compressible and modeling how the particles in it behave treating them as a collective the equations are derived from the conservation of mass and momentum two of the foundational principles of all of physics and they can be used to describe the flow of all incompressible fluids The incompressible navier Stokes equations looks like this there's a symbol and the symbol is called Dell or the Divergence of a flow of the Velocity field so the first equation is that the Divergence of the Velocity field U which is symbolized by this upset down triangle the diversion of the Velocity field is zero that means we come to the interpretation of it in a bit and then the density times the time derivative of the Velocity flow plus the product of the Velocity flow times its gradient equals the negative of the gradient and pressure plus the What's called the Lum of the uh velocity flow times the dynamic viscosity symbolized by you so these are what's called a nonlinear system of equations they're very difficult to solve analytically but there's many many ways you can solve them computationally using what are called numerical methods and we learn those in graduate school they're very complex they're very interesting and you discretize the system and and use them to simulate on the microscopic scale things that otherwise we attribute to macroscopic properties like density and pressure so these methods are used to simulate the flow of fluids in many many applications as I said before and they are used in weather forecasting climate modeling and as I mentioned aircraft design and even the flight of say golf ball through uh the atmosphere when I hit it it doesn't get up that high known to hit the worm burner uh from time to time but um I'm working on that now in addition to the incompressible navor Stokes equation there are a number of other forms of the navier Stokes equations that can be used to describe the flow of compressible fluids such as gases gases can be compressed unlike water say in certain incompressible fluids The compressible navier Stokes equations are much more complex than the incompressible navier St Stokes equations and they're more difficult to solve however the compressible navigar Stokes equations are necessary to describe the flow of fluids at high speeds such as the flow of air around a rocket or my blasting serves in tennis these equations are such powerful tools for modeling the flow of fluids it's almost impossible to capture how important they are even though they're computationally uh expensive meaning they take a lot of classical Computing time they can only simulate the flow of fluids on small scales but a new method that we're talking about today learns the Dynamics of fluids from data and then simulates large scale flows much much faster than these old methods it makes possible the ability to simulate fluids in real time so you can do things and have applications in Realms as diverse as virtual reality and video games so imagine playing a golf game and and you want to have the physics in real rendered in real time depending on the weather where you're at here in San Diego and like to model the golf course on a rainy day a cloudy day a sunny day and have actual representation be faithful to what I would experience if I was actually out hacking up the golf course the key to the new method is what's called a graph-based representation of the fluid the graph represents the particles in the fluid and the connections between particles represent the forces that act between them this representation allows the neural network to learn the relationships between the particles and to predict how the fluid will behave over time the new method was trained on a large data set containing many many fluid simulations the data set included simulations of water of smoke and sand the neural network learned to simulate these fluids very accurately and it was also able to generalize to new fluids that it had never seen before during training this new method is a significant advance in the field of fluid simulation it was actually the first method that could simulate large scale flows in real time again opening up new Vistas for rendering Virtual Worlds here are some of the physics discoveries that were made using the new method the neural network was actually for the first time able to learn how to simulate fluids flowing in complex geometries that traditional fluid simulation methods can't do so-called finite element analyses the neural network learned to simulate fluids with multiple phases like liquid and water and ice and water and this is something that traditional fluid simulations are not able to do as well so this new method is a powerful tool for studying fluid dynamics it can be used to simulate flows that are too complex for traditional methods and it can be used to discover new laws of fluid physics now if we could write a simulator that runs the laws of physics and basically reproduces them to create programs why would we need a learning based algorithm well that's a good question the goal is to show how neural network video footage of lots and lots of fluid and smoke simulations and then have the neural network learn how the Dynamics work to the point that it can continue and guess how the behavior of a smoke puff would change over time say in a room depending on the room's properties its geometry Etc normally neural networks are used to solve problems that are otherwise close to impossible to tackle for example it's very hard if not impossible to create an algorithm that detects say the features of a dog and can always reliably be used to discover what a dog is or a chimpanzee is but it's very difficult to write down a mathematical description of a dog or a cat or a or chimpanzee but these days we can teach a neural network to do it and not need a mathematical algorithm to determine how the sensation of a dog or identification of a dog would work but this is a different task the neural network is applied to solve something that we already know how to solve especially since if we use a neural network to perform this task we have to train it which is long and arduous so often times you see how fast or hear about how fast these different simulations or neural networks are but they never include the training data and how hard and timeconsuming it was to produce that so why would you want to do this does it make any sense well it actually does and for the reason that I want to explain right now is that the training step only has to be done once in this in this particular example and then afterwards you just query the original neuron Network which means predicting what happens next in the simulation it goes almost instantaneously this takes way less time and requires way less memory and overhead while retaining the High Fidelity the accuracy as well as the Precision of the simulation what are some of the drawbacks of this approach well one is known as generalization these techniques include a newer variant that you can see in this video segment can give us detail of the simulation in real time or close to real time but if we present them with something that's far outside the envelope of cases that had already seen in the training data domain they'll fail this doesn't happen with our hand designed bspoke training or algorithms but only in these types of AI based methods so the key differentiator in this paper we're discussing today is that its generalization capabilities are actually astounding much more capable than before the predicted results matches simulations incredibly well if you see it in slow motion you can evaluate it even better well another thing that could be used as a sign of generalization is if you can apply it to something other than smoke or the training data or water and so forth you can actually apply it to say sand or slime my kid's favorite substance and that's a great Step Beyond just say water and smoke and now this scene shows how it's working slowing the evolution of objects with different shapes or in different simulation boxes this is an example of of almost pred internatural new learning techniques so it can deal with these new shapes and new boundary conditions but it also handles the interactions of them really well as well you can train it on a small domain with a few particles and then scale it up once it's learned these general concepts to a much much bigger domain you can also see what would happen in completely unknown situations like if you have water flowing down an incline and then if you remove the incline ramp you can see that it understands what to do with the particles and solving the navigar Stokes equations in more or less in real time now let's try something else an hourglass with Sands running through it such as the Days of Our Lives it's incredible the paper uses this new graph-based method that represents the artificial intelligent physics that's it's learning and it can pass messages among different CPU nodes and it can learn in that way in this neural network way to create a simple and generalizable model that can be a t to force it could be used as a Great Leap Forward to simulate things it's never ever encountered before perhaps using objects that are completely deformed understanding the behavior of fluid flow like air around complex geometries like new stealth bomber wings and so forth these simulations are really really amazing they haven't reached the point yet where they're pred iting The navier Stokes equations imagine if you could go back to the 1800s and show you know sketches of Sands through an hourglass and then some intelligent you know Computing engine like babbage's engine could predict the navier Stokes equation that's the holy gril that's what I'm looking for could we have shown a neural network say the behavior anomalous behavior of the planet Mercury and how its perhelion Advanced year by year accumulating just a fraction of what's called an arcc and this op these observation led Einstein to come up with a way to test his general theory of relativity but what if it went the other way around what if we had never he had never predicted general relativity as an explanation could some computer now predict say the behavior of objects near a black hole and extreme gravity and use that to learn laws of quantum gravity that's what I'm interested in Andrea gz and Reinhard ginzel past guest Reinhard kenzel has discussed on the podcast where they're Imaging in the infrared these black holes near the IR vent Horizons or in combination multimessenger astronomy with things like ligo with things like the infrared Optical astronomy that Andrea gz and Reinhardt gensel do and then Shep dolman who is the leader of the Event Horizon telescope combining those three different Messengers into what's called multimessenger astronomy could we not predict what would happen at the actual boundary between the Event Horizon and the bulk space time that it existed and could we for example predict things like the stretched Horizon as past guest Lenny suskin talked about the event horizon where these Quantum effects of gravity would be manifest if his theory is right so stay tuned for that I'm very interested in this project I use artificial intelligence all the time I use it for constructing new ways to educate my students creating artificially intelligent teaching assistance I'm having one of those on my website my new site coming out soon uh which you'll be able have access to when it's done and you can ask good questions about cosmology as if it's your own personal ta completely for free uh utilizing all my notes and all the course references and papers that I've read over the past two decades as a professor so that's it for now this is an exciting time to be a physicist we may have no need for theor now theorist will always be necessary we'll always need that kind of zeitgeist the spirit of the times will always require brilliant theorists like those that I feature but we need experimentalists too and observers to collect the data so that these super intelligent artificially intelligent overlords can then help us predict analyze and make sense of this incomprehensible volume of data so stay tuned for that and remember any sufficiently advanced technology is indistinguishable from Magic I think these artificially intelligent physicists are quite magical so let me know what you think in the comments below

Channel: Dr Brian Keating

Views: 65,669

Keywords: Dr Brian Keating, into the impossible, brian keating, Experimental cosmology, Brain Keating, Into the impossible Podcast, Science Podcast, Scientific interviews, exploring the universe, Universe Facts, Scientific podcasts, Universe exploration, Brain keating, Space discovery, Physics, Experiment, Eric Weinstein, artificial intelligence, fluid dynamics, chatgpt, max tegmark ai, machine learning, stephen wolfram, computational fluid dynamics, openai sora, game design

Id: VpIyvcVAhAU

Length: 19min 40sec (1180 seconds)

Published: Thu May 02 2024