Full Transcript

Transcript

Nathan Labenz (0:00) Hello and welcome to the Cognitive Revolution, where we interview visionary researchers, entrepreneurs, and builders working on the frontier of artificial intelligence. Each week, we'll explore their revolutionary ideas and together we'll build a picture of how AI technology will transform work, life, and society in the coming years. I'm Nathan Labenz joined by my cohost Erik Torenberg. Hello and welcome back to the Cognitive Revolution. Today I'm thrilled to share my conversation with Tim Duignan, an applied mathematician at the University of Queensland, Australia, who is using cutting edge AI techniques for computational chemistry with the goal of deepening and perhaps 1 day revolutionizing our understanding of electrolyte solutions from the humble and familiar saltwater to the latest lithium batteries. Tim recently shared a short video of salt crystallizing in a water solution. He called this the most exciting result of his career, and this mesmerizing visualization reached nearly 2,000,000 people on Twitter, sparking widespread fascination. In this conversation, we unpack all the techniques that went into making that magic moment. We explore how Tim simulates fundamental physical processes and how by training neural networks on quantum mechanical simulation data, he's created models that can accurately predict the behavior of systems of atoms and molecules several orders of magnitude faster than conventional approaches. This breakthrough unlocks the potential for much larger and longer lived simulations at various levels of scale, all with compute requirements that are remarkably modest by the standards of the foundation models that we're most familiar with today. Throughout our discussion, Tim offers a master class crash course on computational chemistry. He explains concepts like coarse graining. That's a method to simplify complex systems by abstracting away detail that ultimately average out, and also equivariance, a way of representing data that abstracts away from any particular coordinate system and thus helps models generalize better, all with remarkable clarity. We dive into the technical details of how these neural network potentials work, the surprising behaviors that they can capture, and the exciting potential for integrating these techniques into larger, more general AI systems in the future. This conversation is a powerful reminder of how AI is accelerating scientific discovery in ways that we're only beginning to grasp. And for me, it's a compelling reason to believe that AI systems could become meaningfully superhuman in all sorts of weird and unexpected ways over the next couple of years. As always, if you're finding value in the show, we'd appreciate it if you could take a moment and share it with a friend. You can always contact us via our website, cognitiverevolution.ai, where I'm now accepting resumes via a Google form, and I always welcome DMs on any social network. Now I hope you enjoy this deep dive into the frontiers of computational chemistry and its intersection with AI with Tim Duignan.

Nathan Labenz (2:47) Tim Duignan, welcome to the Cognitive Revolution.

Tim Duignan (2:50) Hi. Thanks for having me. I'm a big fan

Nathan Labenz (2:52) of the podcast. That's kind to say. Thank you for getting up super early. We're 14 hours apart or 10, I guess, if we count the other way. And appreciate you accommodating, and I know it's early there for you. I was attracted to your work in the first place as I so often discover things on Twitter, And you posted a tweet that probably is the most viral tweet you've ever posted, I imagine. It certainly went probably traveled a lot farther than you expected, which basically showed a short video of salt crystallizing in water solution. So basically saltwater with a little salt crystal. And you said that this is the most exciting result of your career, and that inspired me to dig in and try to learn more. Maybe for starters, give us a little bit of the sort of context. I think people obviously who tune into the Cognitive Revolution are paying attention to AI. They're probably not paying attention to the sorts of computational chemistry work that you're doing. So, you know, they may not even have any sense for like, we're still studying saltwater. Like that may be a sort of revelation to some. So can you give us the background of what exactly it is you're studying? What are the sort of fundamental questions you're trying to answer? And then we can get into how you're using some cool AI techniques to approach those questions.

Tim Duignan (4:09) Yeah. Definitely. So I've been fascinated with saltwater or the more technical term is electrolyte solutions for over a decade now. I've been studying them since my PhD, and I've kind of fallen in love with them, know? And you talk to people, normal people, you say I'm studying electrolyte solutions and they kind of blaze over a bit, they go like Powerade or Gatorade, they don't quite get it, know? They're like, why would you be studying that? Surely we know everything there is to know about them, but remarkably we really don't. And there's still a ton we need to learn about them and understand them better. And they're actually also incredibly important, They're, I would argue, 1 of the most important substances on the earth, right? The ocean is a salt solution, obviously, but its saltiness plays a critical role in all sorts of phenomena. So climate change, c o 2 dissolves into the ocean to form carbonate ions, right, and that's critical into where our c o 2 is going. In all of the technologies we need to fight climate change as well, electrolyte solutions are critically important. Batteries, they determine how fast a battery can charge and how long they can last. You know, they're thing that connects the positive and the native terminals in the lithium ion battery. So they're really everywhere, and their biology is the other big 1. Right? The flow of ions in and out of your cells is what creates all the electrical signals that make life possible. It's what all of your brain functions on, and the body has developed all of these machines called ion channels and ion pumps to move these ions around. So they're critically important just throughout all of science and technology, really. So we really want to understand them better. And so that's why I've been working on trying to understand them for so long. And the remarkable thing about it is that we really can't even predict their most basic properties from what we call first principles, right, which means, at least in our context means, from fundamental physics, starting from nothing but the equations of physics. And that's the dream I've had really for a decade or now, could we start from nothing but physics, in particular Schrodinger equation and Newton's laws and these kind of things, and predict the properties of these electrolyte solutions, because that would be an incredible achievement if you could do it, I think, and then it would be a really nice demonstration that you could take this technique and then start taking it further and simulating larger and more complex systems as well. So that's kind of what we've been working on for so long, and this is why I'm excited about this latest paper we put up on Twitter, which got this remarkable response, I think. It was great to see other people getting excited about it. Know, for a long time, I thought, 1 cares about this, I was just working on it kind of on my own or with a small team of people, you know, and it's very small number of researchers in the world actually working on these self solutions, right? So it's amazing to see other people can get excited about it as well. So it was really great.

Nathan Labenz (6:36) So I think great job on motivating the problem in terms of all these different domains in which it turns out to be important. If you can cover climate change and batteries and life, you can even touch on all those with a single question. It's clearly an important question. When you say that we can't predict the properties of these solutions from first principles, what are the sort of properties that you would want to predict? And I guess those are all just determined experimentally?

Tim Duignan (7:08) Yes. Exactly. So currently, we measure all these properties experimentally, and the main 1 would be things like solubility. So like how much can you dissolve in water or solution, that's a really important 1, particularly for batteries that determines you want a battery that a liquid an electrolyte that can dissolve a lot of ions in it, to have high concentration ones in many cases. And then there's kind of more technical chemistry things that come along with that, things called activity coefficients and chemical potentials and kind of all of these important quantities that and what you do with those numbers is you actually then plug them into larger scale models that you could then use to model bigger processes. So you might plug those into a climate model, for instance, to predict, you know, the different properties of the ocean and then saltwater, actually, you know, aerosols form as well from these electrolyte solutions. So there's all sorts of different cases where you would really want to know these properties that you could then plug into the models. And at the moment, we get all that data, yeah, from experiment. And the problem with that is that in many cases, know, say you have very high temperatures, which very often occurs in many minerals processing instances, know, where again these electrolyte solutions are up at its very high temperatures, it's very hard to get that data experimentally because you need very expensive equipment that can handle those kind of temperatures, or whenever you have like a new electrolyte solution that isn't well studied, then suddenly it's like, oh we've got to go back to the lab and we've to do these exhaustive experiments to characterize these electrolyte solutions as a function of temperature and pressure and get all that data, and that can be very, expensive, and that means that very often you know people just don't bother, And in particular, know, it's it's very hard as an experimentalist to get funding to just go and do basic measurements of electrolyte solutions. Right? There's not you're not gonna get a nature or science publication out of that, and so there's honestly not a lot of this happening really anymore. Most of the data that we have comes from, you know, a long time ago when experimenters were working on building up these databases, and we're not really building these databases much more anymore. And so this is a real problem when it comes to machine learning and AI because, you know, databases are key here. So in particular for electrolyte solutions, this is a real problem, the limited size of the databases. And there's people doing amazing jobs taking the data that's out there in the literature now and taking models that can allow you to extrapolate that data a little bit, but it's still inherently limited to having good measurements relatively close to the systems that you're interested in. So that's really the issue there.

Nathan Labenz (9:20) Interesting. Are there, like, surprising findings anecdotally that could sort of just give us a little bit more intuition for sort of the kinds of things that you may encounter if solution doesn't behave the way you expect them. I guess I'm imagining like battery fire is the sort of extreme failure mode. Right?

Tim Duignan (9:39) Mhmm. Exactly. So that's a big 1. So in a battery, you have the positive and negative terminals, and you got the electrolyte in between. And actually on that surface, that interface between the the positive, the metal terminal, or the crystal terminal, the liquid electrolyte, you can actually get salt precipitating out. And that plays a critical role in determining the lifetime of the battery. And if that salt layer, that crystal salt layer on that surface of that electrode doesn't form properly, you can get real problems. And particularly, you can get things called lithium dendrites forming, which then can cause a battery to short circuit, and, you know, that then destroys the battery and starts these battery fires. So actually, yeah, this process of crystallization can be really important there, for instance, where you want your salts to crystallize out in the right form essentially, because if you get the wrong form, you can get really nasty things. This is something that we still really don't understand at all. So there's 1 thing which is just predicting these kind of what we call thermodynamic properties, which is what I was talking about before, but then another thing is often you just wanna know, like, at the molecular scale, what exactly structures are forming? What are these things, what do these salt crystals look like, how are they arranged, and these kind of things, because that would give you useful information about, you know, what's gonna happen when we have a battery and we put it at a certain temperature and pressure, how it end with a certain charge on it, what's going to happen with that layer of salt on that surface of the electrode is really important question for batteries and yep things can go badly wrong if you don't understand this properly.

Nathan Labenz (11:01) So for something like a battery, are there sort of additives? I guess I'm thinking about like doping sort of strategies because you you might say, okay, here's a naive, you know, lithium battery with sort of the simplest possible recipe. Now I maybe make an array of these if I'm Tesla and I try to operate them under a bunch of different pressure and temperature conditions and characterize, like, what the viable range is and figure out if that's gonna work for me or not. And then presumably, the naive approach doesn't really work, and then I have to go now I start adding in, like, little admixtures, doping type of things. Is that

Tim Duignan (11:34) Yes. Definitely. So Help me

Nathan Labenz (11:36) develop my intuition there as far as it goes.

Tim Duignan (11:38) So, yeah, this is a big field, there's a guy called Jeff Dunn who's really the leader on this who's developed he's doing some really nice work on this, developing the million mile battery, right? Which is the idea that you're gonna make a battery that can go for 1000000 miles, which is very cool. And the way essentially they do that is they're using fairly standard lithium ion batteries, but they're doing exactly what you're saying. Doing different additives to the electrolyte to try and control the formation of what's this layer on that surface of the terminal, the positive terminal, the negative terminal of the battery, which is called the solvent electrolyte interface. So they're trying to control that process by adding different chemicals to the electrolyte before you charge up the battery, because actually the first time you charge up the battery, which is called the formatting step, that's when this layer forms, and it's really critically important for determining how long the battery will last. The critical thing is that the way we do that currently is a huge amount of experimental trial and error, right, so just trying different things, adding them, and this is a real problem if you want 1000000 mile battery, right, because you have to then add put the additives in and then cycle it for the equivalent of 1000000 miles, right, which takes a long time. So there's a huge amount of experimental effort to try and find these right additives to do this. So if we could instead kind of understand it from first principles, get a really deep molecular scale insight and understanding of exactly how these crystals are forming and what effect that's gonna have on the lifetime of the battery, that's the dream, and that would be really useful. Right? And so there's actually some ways machine learning are coming in here, people, you know, you can do things like you put some additives in and then cycle the battery for a few cycles and then use machine learning to try and predict what it's gonna do after 1000000. So there are other ways you can use machine learning AI here, which is quite interesting. But in the long term, I think to like really crack this problem and solve it once and for all, you'd really wanna have this like full understanding of exactly what are all the different species forming at these surfaces of these electrodes and how does that affect the cycling performance of these batteries.

Nathan Labenz (13:22) Gotcha. Okay. Cool. So that gives me a good, you know, outsider intuition for why you would be excited to see crystallization happening in 1 of your AI projects. But let's take maybe 1 step at a time. So in terms of trying to figure this stuff out from a first principles perspective, my understanding is that what kind of comes between experiment and the AI models is super computationally intensive simulation. So that is to say, we're looking basically very small. These are like sub microscopic, right? You're talking about lithium atoms. It's a single atom. It's lost its third electron. We're really down to a very small shell of a So this would be very much in the quantum realm and you are then forced to compute numerically, I suppose. You're this is not the closed form stuff. Right? But just heavily computationally grinding out time steps in the in the wave equation?

Tim Duignan (14:24) Exactly. Yeah. So we do a thing which is called ab initio molecular dynamics, which is a bit confusing because the acronym is AIMD, but it doesn't have anything to do with AI. So, you know, if you've got first principles molecular dynamics as well as another term, I think. And the basic idea is you wanna solve Schrodinger's equation to calculate the forces on all of the atoms in your system. Right? So you're starting with quantum mechanics with Schrodinger's equation, and the key reason you use it is to compute both the energies and the forces, which are linked of every atom in your system. And if you can do that, then that's really useful because you can then do what's called propagate the equations of motion, basically solve Newton's equations to calculate how all of the atoms will move through time, make this kind of movie of how all the atoms are moving. So this what a molecular dynamic simulation is. And I did some work at a national lab in The States working on running these simulations on this huge DOE, Department of Energy, supercomputers they have over there, which is very fun. But, yeah, these things are really, really expensive. So, you know, you're talking huge supercomputers and, you know, you'll run these things for months, and you're looking at really tiny systems. So you might be looking at, yeah, something very simple like a lithium chloride electrolyte solution and maybe a 100 water molecules or so, and you can simulate it for, you know, picoseconds, which is, you know, fractions of tiny, tiny fractions of a second. So this is, yeah, far far actually, smaller than you can see really using any experimental technique. But it kind of makes it on is a plus as well. Right? Because it gives you access to kind of a spatial scale and a time scale that we can't access any other way really, because it's just moving way too fast and much too small to see using it was using light or x rays or anything like that normally. Mean, you can indirectly get at this thing these things experimentally, but you can't directly kind of observe them. So that's the advantage of it. Actually kind of good that it can access these tiny short timescales, but it does mean that it's inherently limited. Right? Because a lot of phenomena happen at larger timescales, larger spatial scales and longer timescales. Right? And so that's really what we'd like to be able do, take these techniques, which we know work really well for small systems. Right? There are certain experiments you can calculate and show that you agree with experiment to demonstrate that these techniques do work. The problem is they're just so computationally expensive that you need these huge supercomputers. And so it's not the kind of thing you can scale. Right? You're not gonna take this and simulate every possible different electrolyte solution or look at any kind of bigger, complex, more complex systems really with them. And that's where the AI comes in here.

Nathan Labenz (16:40) So these simulations are done on picoseconds. That's 10 to the minus 12 seconds?

Tim Duignan (16:45) Yes. So technically, your time step is a femtosecond, but you run it for thousands many thousands of femtoseconds to get you to picoseconds. Yeah. Which is 10 to the near 12.

Nathan Labenz (16:54) Yeah. So your time step is 10

Tim Duignan (16:55) to the minus 15 seconds. Yeah.

Nathan Labenz (16:58) And you're running thousands of those. So you're getting

Tim Duignan (17:00) a slice

Nathan Labenz (17:00) of time.

Tim Duignan (17:01) Yeah. That's the kind of time scale which intermolecular vibrations happen. So you and you really need to capture those. So that's really you're forced to use that kind of time scale.

Nathan Labenz (17:10) And is this when you're solving the wave equation, this is a if I recall correctly, the second order differential equation, and I believe you would solve that by a polynomial approximation, like, we're using, like, Taylor series type stuff?

Tim Duignan (17:25) Kind of. Yeah. So we're solving the time independent Schrodinger equation, which makes things a little easier. So we're not doing the full Schrodinger equation, which is time dependent, which can be very hard to solve, but there's some approximations we can use, particularly a thing called the Born Oppenheimer, some tricks we use. Basically, the electrons are way faster than the nuclei. So that's a very useful fact of, you know, the world that allows us to use some tricks to find easy ways to solve Schrodinger's equation approximately. So we don't have to calculate how the full wave function kind of evolves with time. In fact, we're just doing a thing called calculating the ground state energy and the ground state wave function. So it makes the problem a little bit easier. We're not trying to calculate, you know, this complicated quantum mechanical evolution of the wave function. It's just just 1 kind of wave function that we're trying to get a static kind of wave function. So that makes it a little easier. And yeah, there's all sorts of beautiful mathematics people have developed to kind of solve these things really fast and really efficiently. So the main way people use it is a Gaussian basis set set, actually. So you can run-in polynomials, you use Gaussians, and you put those on the center of your atoms, basically, and those kind of just you write your wave function as a linear combination of these functions to describe your wave function, or you can use what's sort of plane wave basis as well, and we can mix and match these things, which is actually the code we use called CP2K, 2 k, which is really efficient and fast at doing these things. So, yeah, there's a ton of really beautiful mathematics going into how do you go about solving these equations really fast enough to do some of these simulations. You just get to this kind of picosecond timescales and beyond.

Nathan Labenz (18:51) And this is an all to all forces on all other atoms in the system or all other molecules in the system are calculated to a 100 water molecules that gives you

Tim Duignan (19:01) Yeah. So that's the way we

Nathan Labenz (19:02) attention like relationship. It goes with the square of the number of?

Tim Duignan (19:07) Oh, yes. The scaling is very interesting. Yeah. So you have to calculate the force on every single water molecule. So you calculate a total energy, and you take the gradient of that energy as a as the function of the position of every atom to get the force on every atom. So and because the force is really what we care about because that determines how they move. So you get a force of every atom. The scaling is interesting, so it depends on which approximation you're using. So we don't you know, if you use the brute force method of actually scale exponentially with system size, which is a really nasty thing. So you want to avoid that. Right? But with the quantum chemists, that the this is the discipline that tries to do this. They've found all these beautiful ways of solving this Schrodinger equation really quite efficiently so that it doesn't scale quite so badly with system size. So you can't quite get it down to a quadratic, which would be nice, but you can get it down to kind of some kind of polynomial times depending on which approximation you're using, and that allows you to push to reasonably large systems such that you can kind of get a decent approximation of what a liquid is like, right, which is where we really wanna get so much important chemistry. All the biology happens in kind of this liquid phase. And the nice thing about these, you know, all these approximations we have now is I can get decent accuracy, but they can start to approximate the things that look a little bit like liquid will look enough like liquid that we can really start to make predictions about liquids phases, which is really exciting.

Nathan Labenz (20:21) Interesting. Just because I'm so curious, why is it exponential time instead of quite I mean, my naive intuition for this would be like each molecule or atom has some force relationship with each other 1, and that would seem to be.

Tim Duignan (20:36) Yes. So that so there are in fact, we can get linear scaling ways of computing force when we talk about neural network potentials. But when we're solving Schrodinger's equation, it's that's because you actually have to treat every single electron in the system. Right? So you're not even you're not treating the the atoms aren't the base unit. You know, they actually just are static. You just say, okay. The atoms or the cores of my atoms are just fixed in position. The electrons is what you're calculating. You're calculating the wave function effect. And the thing about the wave function is when you add electrons, the kind of space that you're in increases kind of combinatorially. So as you add more electrons, an electron can be coupled with every other electron in the system. So this kind of gives you this exponential growth in terms of the computational cost of the calculation. If you were to try to brute force this is using thing called full configuration interaction. If you're trying to brute force it out, every single electron can couple with every single other electron in the system, and so you get this very nasty scaling behavior when you try and do that. And you can't use and so what you're saying is, yeah, you could probably use an approximation where you only worry about things locally or you don't have to try and worry about everything being coupled with everything else, which is why you get this nasty scaling.

Nathan Labenz (21:45) Yeah. Interesting. So is this kind of a when you run these simulations, is it sort of a tick tock type of thing where you first at each step, you compute the forces, and then you propagate 1 step and things move, and then you recompute the forces, and you're just going back and forth through that cycle?

Tim Duignan (22:02) Yes. So it's kind of a combination of 2 things. We solve Schrodinger's equation first to get the forces, and then you go and use what's called a molecular dynamics engine to calculate how the atoms are gonna move. So you update the positions of the atoms using those forces, and then you will redo the force calculation in that new positions. K? So you can think it was a loop going back and forth between the energy calculation and updating the positions, and that allows you to make this trajectory. And the reason we do that is because it turns out that the atoms are big enough that they behave fairly classically, you know, so they follow Newton's equations, which is like the limit of Schrodinger's equation when things get big and high temperatures. Turns out you can just use Newton's equation. So we don't want to use the full quantum mechanics for how the atoms behave. Although there are rare cases where you do have to do that, in particular for hydrogen atoms sometimes, but mostly it's okay to treat the movement of these atoms using classical laws. Whereas the electrons are inherently quantum mechanical beasts. Right? They're just so small, so light that you have to use quantum mechanics to to treat the properties. That's why we have to solve Schrodinger's equation to get the forces because that's all about the what the electrons are doing. But then if you wanna know how the atoms are moving there, you use Newton's equations because they're more classical.

Nathan Labenz (23:08) Fascinating. Okay. Hey. We'll continue our interview in a moment after a word from our sponsors. So at what point does the wave collapse happen? If I'm thinking of quantum mechanics as being these sort of distributions over possible positions or position momentum combinations or whatever. And then you do this sort of next step like propagation. How should I think about the certain apparent contrast between a wave or a distribution and a discrete like this thing is moving this much in this direction?

Tim Duignan (23:41) Yes. So that's a great question. We never really have to worry about wave function collapse, which is quite nice, because that starts to get very complicated. Right? And this is why and the reason essentially is because of this thing called decoherence, which means that at room temperature, you get decoherence, which means everything essentially just the wave function is essentially always collapsed. And so we really only care about the wave function in its ground state, which means it's kind of a very well defined state, state, and it's kind of been collapsed into that state. So we're not too worried about things where you get the wave function and superposition, which gets very complicated. That's when you get back to needing these kind of exponential scaling calculations, which get very tricky. So that's a nice another nice approximation that we can make because we're kind of at room temperature and we're dealing with, you know, relatively large systems. And though there are cases where this happens, but certainly for a quantum computer or something, you know, they're trying to exploit these kind of phenomena like wave function collapse and put these put their systems in these unusual quantum superposition states they call it. So they have to be really careful to kind of isolate their systems and, you know, make sure cosmic rays aren't interfering or something. Because as soon as anything touches your system, suddenly that wave function could collapse because it starts interacting with its environment. So that's really all of these unusual quantum mechanical phenomena only really happen when things are isolated completely by themselves. As soon as they start interacting with, you know, something at room temperature, then that's all gonna immediately kinda get washed out, and things are just gonna collapse down to their to their normal state. So that's really what we're just trying to do, is just find the wave function in this kind of ground state, which is kinda almost the easiest problem you could have to do, which is nice in quantum mechanics, just finding, technically, it's the lowest eigenvalue of the Hamiltonian operator, and it makes it a little easier. Interesting.

Nathan Labenz (25:21) Okay. I'm not sure I fully grasped that, and I suspect that probably a lot of people don't feel like they have full intuition for it either.

Tim Duignan (25:27) I don't think I I have either. So I

Nathan Labenz (25:29) think that's Yeah. Well, nevertheless, though, this I I think that's at at a minimum a great motivation for, you know, why it would be nice if we could have an AI that could figure all this stuff out for us. So tell me how you're now training AIs. And I understand, first of all, you're training the AIs based on the simulation data. So we've set up to the point where there's a quantum mechanical approximation solve based on where everything is and the forces that that exerts, and then there's a timestamp. So how does that simulation data turn into training data? And then what exactly is the model learning in terms of what are its inputs and its outputs?

Tim Duignan (26:06) Yes. Great. So, basically, by far the most expensive part of these simulations is the solving Schrodinger's equation. Right? That's very, very expensive, and that's where your computer spends almost all of its time. And so that's the piece that if you could speed that up, it would be really potentially transformative or useful. And this is exactly why people have been developing this tool called neural network potentials. Okay. And so the idea is that this is basically, it's a neural network, so it's just kind of a very, very flexible function with a huge number of parameters, and then you're just gonna adjust those parameters to predict what that quantum chemistry calculation is doing. That's solving Schrodinger's equation. Right? And essentially, so the input there is the coordinates, the positions of all of your atoms. So you've got an oxygen atom here and a hydrogen here and a sodium here, and each of those has a force on it, and there's a total energy as well. And that's what you're trying to predict. So you're trying to go from the positions of the atoms to the forces on all of the atoms. And that is the expensive quantum calculation that we're trying to replace with a neural network. And then so you do need to do at least some of these quantum mechanical calculations, right, to generate your training data, and then so we do supervised learning on that training data that you've generated from these quantum chemistry calculations. And these neural network potentials then can reproduce this mapping fairly well as long as you have in your training data something that's sufficiently similar to the case you're interested in, you then don't have to go back and do the full quantum chemistry calculation again. You can just use the neural network to calculate those energies and forces much more quickly. So this allows you this loop between, you know, the quantum chemistry and the classical, you can then massively speed that up because you're not spending so long waiting for the quantum solar to loop.

Nathan Labenz (27:41) So you're basically doing the same TikTok kind of loop, except instead of the numerical solution for the wave equation, Now you've trained a neural network to make a guess, and

Tim Duignan (27:54) it turns out that guess is quite good. Pretty good, normally. Yeah. Exactly. So we keep the keep the part where we're solving Newton's equations because that's actually pretty quick. Turns out that updating the positions of all of the atoms based on the forces on them is actually doesn't take too long at all. So the by far, the biggest cost is this quantum part. So if you replace that, then that gives you access to, you know, a whole new range of systems you can look at. So, you know, normally, that's the quantum you can speed that by a factor of 1000 or more. And so suddenly, you can look at systems that are 1000 times bigger and 1000 times longer, and that opens up such a range of systems that that alone is a really useful tool to look at a huge number of important things that we haven't been able to look at previously. And this idea has been around for a long time, actually. It goes back to the nineties, people came up with it. There's a beautiful paper in 2007, actually, Jorg Baylor came up with with Paranello, this idea of using these neural net potentials and they train them. The initial ones were just using a very kind of vanilla neural networks, you know, multi layer acceptor on just a standard kind of density forward neural network. Since then, we've started to develop these things a bit further and get a bit more sophisticated in terms of the ways we do this thing. So these tools are really getting a lot more useful now, is exciting as well.

Nathan Labenz (29:01) So I wanna you know, a lot of questions about the the nature of the networks. Maybe for starters, like, how big are they? Like, how much data is going into training? People, you know, I think are, at this point, most familiar with large language models. We're thinking in terms of, like, trillions of tokens and 10 to the 24, 25, maybe 26 flops, and many billions of parameters, if not even into the trillions with mixture of experts. How does the just size of, like, datasets, compute resources, and models, compare for for these purposes?

Tim Duignan (29:36) Yeah. I mean, is what I'm excited about is really they're very moderate, I think, compared to these things we've talked about, yeah, with these large language models are kinda hard to fit in the size of these GPU clusters and things to me. Right? So we're talking to generate my training data. So for me, I'm only interested in, like, 1 electrolyte solution, but to generate that training data, you need maybe a few GPUs for a day or so. Know? So nothing significant at all. You So I don't get access to you know, reels at all, really, just using kind of the standard allocation on our supercomputer here in Australia. So, you know, I don't have access to to that huge you don't really need these huge clusters or anything. Think to know? And the training data generation is still probably the most expensive part, and, yeah, that's fairly you can do that in a in a few days. Then running the simulations is not too bad, but you can do that on on, you know, at least for my system, simple electrolyte solutions. You know, just a few GPUs is really all you need, and you can run these simulations for really long enough to converge everything very well. So the the costs of these are really very unreasonable. Having said that, people are trying to do much more ambitious things obviously now with these neural net potentials. So there's a nice paper from the people who developed the neural net potential that I use. They showed they could simulate an entire HIV capsid. Right? So a huge system there, you know, millions of atoms. So that kind of thing, then you do want you should use 1 of these huge GPU clusters, which they use to simulate that kind of thing. So you can really push these things to huge system sizes when they start to get really expensive and start to get comparable to some of these large language models and things. But in my opinion, there's a huge number of systems that we don't understand fully yet. They're actually quite small, and if you just simulate those, you know, there's a ton of interesting insights you could get. So I don't think you need these huge resources to study interesting scientific problems, although there are certainly ways you could use these really huge resources to start if they were made available to us.

Nathan Labenz (31:14) Cool. So what you're doing the the recent results have come from a couple days of GPU time, and I guess intuitively that sounds like hundreds of millions of parameters in a model?

Tim Duignan (31:26) Oh, yes. So the parameter size, actually about, I think tens of thousands of parameters is Tens thousands. Yeah. So not too big at all. Wow. So so it's very small compared to it's funny because it's very small compared to, yeah, the large language models, obviously. But that's trying to, you know, model the entirety of the Internet. Right? So whereas ours, in fact, it's quite a lot simpler, think, the distributions essentially that we're trying to model, you know, than all knowledge. So we don't need that kind of number of parameters. People are trying to build these things called universal force fields, which maybe gets later, but those will, you know, have much higher parameter counts and getting up closer to that kind of thing that you need for a large language model. But it's actually interesting because there's a classical method of doing these simulations obviously, which we haven't talked about yet, but the basic idea is you kind of guess at some physical equations about what this force field should look like. And there you have this parameter fitting problem as well, but they actually end up using, know, a few dozen parameters. So in some cases, we've actually got way more parameters than the kind of classical approach to doing this, but still a lot less than these very large AI models. So it's an interesting we're kinda in this in between zone, but there's a lot of work. People are trying to see, can you get the parameter count down even lower, which would be very interesting. And so we'll see where the best models end up at is gonna be a very interesting question.

Nathan Labenz (32:38) Yeah. I'm really surprised that it's only in the tens of thousands of parameters. That's, like, strikingly small. That also suggests, I guess, if you're training something that small for days on GPUs, that would sound like overtraining on a standard thing, but is is there a sounds like there may be some sort of grokking type phenomenon going on here. Right?

Tim Duignan (33:01) Yeah. Interesting. So the training is actually not days. The training is normally done, yeah, within a day for my systems at least. It's mainly the days would be generating the quantum chemical data because that's the really expensive part.

Nathan Labenz (33:11) And then on GPUs too, so that's also, like, parallelizable?

Tim Duignan (33:16) Yeah. Technically, I did it on CPUs, but you can do it on GPUs now. There's now codes that do it on GPUs even faster. So that's really where things are going. Think we'll be running all these quantum chemistry calculations on on GPUs now. And, yeah, that's the kind of thing you need a few days probably to generate your training data, but the extra training yet is I don't know the exact timing, but it's probably, you know, few hours really. If is there a grokking phenomena? I I don't believe so. I haven't seen any evidence of that yet, but that would be very interesting yet if we saw something like that.

Nathan Labenz (33:43) So how about the architectures? Are we talking like transformers here or is other, you know, familiar architectures or are these exotic architectures you're using?

Tim Duignan (33:55) Yeah. No. So the 1 I use is called an equip developed by the Kaczynski lab out of Harvard, and that is a graph neural network, an equivariant graph neural network, so it's got missives passing, so you know, I'm sure people are familiar with those kind of things. There are also some there's 1 called torchMD, believe, which is a transformer, which has quite good performance as well, you can use a transformer there. And So the traditional ones, which some people still use are also just, you know, a plain, yeah, multilex eceptron as I said. So, yeah, these are all kind of standard architectures that will be familiar to people working in AI, I think. Yep.

Nathan Labenz (34:27) And then there's some interesting stuff around, like, how the data is represented, right? There's this when you the equivariance concept is critical to this whole thing, right?

Tim Duignan (34:40) Yeah. So equivariance is this idea, and this was kind of the big advance that we just had in in the last couple of years, I think. So as I said, this idea has been around for a long time, but the issue was always you needed so much training data that I always really struggled to get it to work efficiently and effectively. Basically, you need to keep going back and generating training data constantly to kind of look at the things I was interested in, in particular kind of these pairing of these ions, how strong were they, stick to each other on water. And so that was always a a real challenge or a frustration I had. But then we saw, you know, this paper that came out in EQuIPP is the architecture that we use. And the critical thing in that paper, they got this table and they say, we get lower error for a water box. They're trying to calculate the forces on a system of water. We get, I think it was 4 times lower error, and they got that using 1000 times less data. Right? So 3 orders of magnitude less data. So that kind of blew me away. So as soon as I saw that, I kind of jumped on this. Because it's not often in science, you know, you'll see a breakthrough with if something goes down by 3 orders of magnitude, that's the kind of thing that's that's inevitably gonna have an impact, I think. So that's really why I jumped on that quickly and to try to exploit these new tools, which have much more training data at once. Because then, as I said, the biggest cost is this quantum chemistry calculation you have to do to generate your training data. So if you can get the amount of that data that you need down, that's really exciting, because then you can start to really scale these things essentially and look at many different systems. So that's what I'm really excited about now, this tool of equivariance that allows you to kind of have this training data.

Nathan Labenz (36:05) And can you explain a little bit more like what that is? Is it a sort of like a module of certain symmetries? You're sort of finding a way to represent things that is in some way more general, less contingent on, like, the specific coordinates, but somehow relative to each other?

Tim Duignan (36:22) Exactly. Yeah. So it's a really, really nice idea. It comes out of group theory and representation theories and beautiful mathematics that actually kind of revolutionized particle physics, a few decades ago. So as a physicist, it's kind of really appealing to me that these same tools now are are being transferred and being useful in chemistry and and useful in, you know, molecular sigration, which is really exciting in AI. And it's it's obviously highly mathematical. I'm not an equivalence expert, so, you know, I could give you kind of a high level understanding. And and the basic idea is that you you don't want to just feed, like, raw coordinates into these neural networks. Right? Because the raw coordinates are fairly arbitrary. In fact, if you, like, rotate the molecule, you know, the coordinates, the positions of all the atoms are gonna change, but its properties really shouldn't change. Right? Because if you rotate a water molecule, it's still a water molecule. It's still gonna have all the same properties of a water molecule, they're just gonna be rotated as well. And so what you really want to build is a neural network that kind of has this kind of rotational symmetries of space embedded into it, it knows about the fact that you're on three-dimensional space, if you rotate something, it's still the same thing. Okay. And the way you do this mathematically is you 1 way of doing it is you build tensors into the neural network. So it's not just operating on, you know, matrices or big list of numbers, but it's actually operating with these mathematical entities tensors. And the basic idea of a tensor is just has these nice rotational properties, right, it can kind of keep track of directions, so it has some idea of which direction it's pointing in is the basic idea. And it turns out when you do this, yes, it turns out the training data goes right down. And this was really appealing to me because it has actually got deep parallels to the way we model intermolecular interactions, you know, which is what's important for building these simulations, is the ways these things have been modeled for decades, really. A thing called the multipole expansion, which basically describe your interactions in terms of these tensors, you know, dipoles, quadrupoles, ox poles. You've heard these terms, but these are actually mathematical tensors, and that's the way we've been doing it for decades. And so it turns out that when you build that fact into the neural network, your trained data requirements go way down, obviously, because it can kind of exploit the fact that this is the natural mathematical way to describe interactions of molecules. So that was a really nice kind of example of how you could use fundamental physics ideas and combine it with AI to get you these really advances in performance. Now but there's a really interesting debate going on at the moment about this, though, actually, because AlphaFold 2 was the kind of 1 of the biggest breakthroughs, obviously, in AI for science, and it used equivariance in the way its architecture worked. But the latest version of AlphaFold, AlphaFold 3, actually got rid of the equivariance. And so and yet they still managed to get comparable performance to AlphaFold 2. So there's people debating, is this actually critical or is it not? And the and the trick they used was a thing called data augmentation. So they actually did feed raw coordinates into the model, and some other people have showed this works for neural network potentials too. But they took the dataset and then they rotated it in like, you know, 1000 different ways, and then fed that big much bigger dataset with all of these rotations into the neural network. And it turns out that when you do that, you can at least get error metrics, so, you know, your loss and things can get as low as the equivariant models with a similar sized training data. So it'll be very interesting to see, if these neural networks are actually as useful for practical tasks like running MD simulations, which I haven't really seen yet. So that's gonna be a very interesting question moving forward. Are there ways to kind of get around building this equivariance into the neural network potentials? Because the disadvantage of it is that it can slow your neural network down quite a lot because it's just a little bit more inherently complicated doing these kind of tensors having these tensors interact with each other rather than just, you know, normal matrices or lists of numbers. So that's a really interesting thing to keep an eye on how that field advances. And as an example, you know, there's a paper from Apple recently where they said swallowing the bitter pill, right, as an example perhaps of this, the bitter lesson, right, that really there's no way to be smart about how you build your neural network, the answer is just to make it big with lots of parameters and then get as much data as you can basically, and let the neural network learn everything, and don't try and constrain it in interesting ways. So that's an interesting example. I'm kind of rooting for equivariance to hold out because it's you know, you think if there's anything that that's worth building into neural network, it'd be, like, fundamental physical laws and, like, the the symmetries of nature. So we'll we'll see how that turns out, but it's very interesting.

Nathan Labenz (40:25) Hey. We'll continue our interview in a moment after a word from our sponsors. So let me just make sure I understand a couple things. It seems like the big idea is from a principles level, the coordinate system shouldn't matter. Right? But that's a very common thing in physics. Like, you can choose a convenient coordinate system, but in general, the coordinate system shouldn't matter. So you can either find a elegant way to represent your data in this sort of relative to 1 another form that is the equivariance form. And that essentially sort of abstracts away from any particular choice of coordinate or any particular choice of origin point in the representation of the data. But then your neural network architecture gets more complicated because it has to handle this relative representation. Or you can do what AlphaFold 3 did, and that is just systematically apply a bunch of rotations and translations to the data and then feed that all into a less complicated neural network, but more data points sort of synthetically generated. All of us is synthetically generated, but there's, like, the expensive waveform solution generations, but then and then also the sort of very cheap presumably, like, translation and rotation generations based on those, feed that into a comparatively simple neural network, more data points. And it seems like we're in a world where either of those can work and you might even think it could come out in the wash, or sort of, you know, in the limit, they might converge on kind of being the same thing.

Tim Duignan (42:06) Yeah. Exactly. I think that's probably where we hit it is that you can just it's kind of you can impose these you essentially, you need to tell the neural network that it's in three-dimensional space and it has these rotational properties. Right? But you can do that in different ways. 1 is you can do put it on the data side and augment the data, or you can kind of encode it, hard code it into the neural network itself, and tell the neural network that actually, you know, restrict the things the neural network can do so that it makes sure that it track of the orientations of all of these things in space. Yeah. So there's kind of the 2 alternative ways of doing that, and it'll be interesting to see which 1 wins out. I suspect perhaps for these simulations, running these molecular dynamic simulations, it may be worth having keeping the equivariance in the neural network itself just because it may give you better generalizability once you get away from the data set or once you're trying to do things like look at this crystallization phenomenon we saw. So I'd be really surprised, I guess, if these neural networks that impose it on the data side could still see this kind of generalizability properties. So that's gonna be something interesting to watch. But, yeah, that's essentially it.

Nathan Labenz (43:07) Yeah. Okay. Cool. 1 ML paper that I would recommend if anybody's interested in developing their intuition for this sort of equivariance concept, but also just a really interesting finding about neural networks is called relative representations enable 0 shot latent space communication. This is out of a a group in Rome and also some folks from Amazon. And this is a paper I kind of always go back to because they basically showed that the latent spaces learned from the same dataset across different training runs with different initializations seem to be converging to the same thing, modulo some sort of, like, rotation and maybe, like, a scaling factor. And with that, if they can, like, figure out, you know, what is the sort of rotation or scaling factor that's applied, then it creates potential for these model spaces to kind of talk to each other or to potentially, like, distribute training. There's could be implications for that. Model merging, like, who knows? Right? There's a lot of kind of interesting. If there's this, like, fundamental convergence, there's a lot of interesting downstream consequences of that. So that paper is relative representations enable 0 shot latent space communication to an aside. Nice.

Tim Duignan (44:24) Checking

Nathan Labenz (44:25) out. So, okay, going back to this crystallization phenomenon. So now we're all queued up to hopefully appreciate not on your level, but, you know, hopefully much more than we were at the beginning, why this is special. So what you saw was you trained a you collect a bunch of this data in the computationally expensive way of solving the wave equation. You train a neural network on it with positions in forces out. Now you're able to run this simulation propagating 1 femto or, you know, a couple of femto seconds at a time in this TikTok sort of way where it's here's the position, calculate the forces, apply the forces, update the position, loop through that. But you're doing this now with the network doing the force predictions. And lo and behold, you see a salt crystal forming in your simulated solution. You had said that it's there was no crystal in the training data. How do you know that for for 1 really stupid starting question?

Tim Duignan (45:26) Yeah. So, yeah, there's 1 other piece of the puzzle which which we should explain first perhaps, which is that that wasn't just a pure all atom simulation where we saw the crystallization. So we actually had 1 other step, which is we used coarse graining to get rid of the solvent molecules, which maybe we can get to that. So that actual simulation was just the ions itself, and we actually used a neural network potential to do that as well. So this is a really exciting, even more recent use of these neural network potentials, is not only can you go from predicting from the positions of every atom to predicting the forces on every atom, but you can also do this trick called coarse graining where you kind of ignore certain parts of the system. So we actually took a simulation, an all atom simulation, and then use this neural network potential to predict the energies and the forces just on the ions. So that simulation actually didn't have any explicit water in it. It was what we call an implicit water simulation. The AI had learned to kind of implicitly approximately account for what the water molecules are doing without having to exactly calculate them. So this is kind of exactly analogous to what you do with the normal neural network potential where you're getting rid of all of the electrons. So with a neural network potential, you can just calculate the energies and forces from the positions of the atoms themselves without ever kind of thinking about the electrons, and then you can do exactly that same procedure again, but getting rid of the solvent molecules. So now you don't have to predict you don't have to incorporate the position of the solvent molecules into the neural network, you just predict the energy and forces from the ions themselves. So this is a very another exciting way you can use these neural network potentials to kind of push to even larger scales. So you're not having to stuck with this simulations where you're treating every single water molecule in the simulation, because that kind of has inherent data limitations. Know, you gotta deal with, you know, when you wanna push to life systems, you gotta keep track of every single water molecule, and you end up just spending a lot of computational time wasted studying how some water molecule out in the bulk is buzzing around and rotating, which you don't actually care about at all. So this is the other exciting thing we did in this paper. So people had done neural network potential MD of electrolytes before, and if you do the all atom way, you won't probably won't see crystallization because this crystallization phenomenon normally happens on very, very long time scales. And so that's how we're feeling confident that it isn't in the simulation, right, because it's very rare to see these of kind things. It only happens typically if there's something really wrong with your forces if you're studying a normal salt like sodium chloride that doesn't normally precipitate, crystallize out a solution. But then you can also go back and, like, explicitly check these things. So we can compute a thing called the radial distribution function. It's basically just a histogram of the distances, and you could look at that and it looks totally liquid like. You know, it has all the features of the liquid at low concentration and so on. So there's no evidence in that simulation of any crystals forming. And then you can just, you know, inspect the trajectories more directly as well and just check and make sure that there's no signs of these crystals forming. So we did go back and check for that in the all atom and the full simulation that we trained this neural network potential on. And so but the exciting thing was, yeah, we saw this crystallization using this additional way that you can use neural network potentials where you also coarse grain out the solvent, and that was exciting because it means that you can do something that you can't do with these all atom simulations really, which is that you can look at very quickly look at things like crystallization with much less compute time, basically. Because to if you wanted to look at these kind of crystallization things with the all atom case, it would take a huge amount of computer time. You have to run it for very, very long times, and you have to use probably a technique called enhanced sampling. We kind of forced this crystallization to happen. We're here it has just kind of dropped for free out of the simulation. And the remarkable thing about that was in particular was when you do this coarse graining or getting rid of the solvent is that the solvent is actually really, really important in determining the interaction of ions. Right? It actually makes the interactions about a 100 times smaller. Okay? So they call dielectric screening. The water molecules are incredibly important in determining the interaction of the ions in water. So it's quite remarkable, but that doesn't occur in the crystal, right, because there's no water there. And so, yes, somehow this neural network potential now that's trained to get rid of the water molecules can in fact kind of predict that as you go from this water to this crystal, what the structures are gonna look like, even though the strength of these interactions could be changing by almost a 100 times as you go from these 2 different phases of material. So that that was what made that even more remarkable is that it's not it it would have been, you know, nice if you just looked at longer timescales with the neural net potentials of crystallization, but we could also use the trick where you got rid of the solvent molecules and yet it still works to see these phase transformations, whereas I would have thought that would never be possible. Know, I've been and this is actually what I was originally trying to build in my PhD were these continuum solvent models, right, which is these models which ignore the water and just treat the ions, which can be much, much faster, which is really their advantage. But the problem is the water is so critically important, but it's not clear at all how to do that. So I spent my whole PhD, literally years trying to build these things, right? And I didn't really get very far trying to build them by hand and come up with mathematical equations and fit these parameters and do all these things. Anyway, built some models that were okay, but they're really inherently limited in their generalizability. But here, the model, you know, this was trained, you know, in just a few hours, I think, automatically kind of realized and learned how to build 1 of these continuum solver models itself, and they could run simulations, and they could see these emergent phenomena. So that's really why it blew me away and was so surprising. You know, this is the kind of thing that, you know, I've been trying to do for years. I had kind of given up and thought was impossible, and yet it just dropped out for free to training 1 of these neural networks, which is kind of what's so exciting about it.

Nathan Labenz (50:47) Okay. I wanna dig in on this on a number of different fronts. That is definitely really interesting, and I wanna make sure I understand it all. I guess for starters, in terms of how you knew that there wasn't crystallization in the dataset, it's a combination of you're running it for a very short period of time, a shorter time scale than crystallization usually takes, and you also went back and screened the data and said, can we spot any apparent crystals by just looking at are there things that appear to be bunched together? And if you don't see that, then you can say, okay. There's no crystallization in this dataset.

Tim Duignan (51:29) Yeah. Exactly. So what you do see in the simulations is you see the sodium and chloride ions kind of periodically peering up, and then they'll come apart. And then occasionally, you might get, you know, triplet forming, but you don't really see any of these bigger clusters that look like the crystal really. So, yeah, that really indicates that it's not crystallizing the data. It's a bit hard to determine exactly what the biggest cluster that forms is or what the most similar thing in that liquid simulation is to the crystals. That's something to potentially dig into further. You know, how close is the training data? What is the closest example in that training data to a true crystal? But definitely, there's nothing like this crystal growth phenomenon where it starts growing larger and larger and larger and kind of phase transforming the whole system.

Nathan Labenz (52:09) Gotcha. Okay. So then in terms of the coarse graining concept, again, just for intuition sake, when we talk about solutions, I think of these measured in like molarity, right? So a 1 molar, you can tell me how many molar your solutions are, but a 1 molar solution of salt water would be 1 mole of salt in a liter of water. And so a liter of water is a kilogram and a mole of salt would be whatever, like 30 grams or something. So we're 1 part in 30, something along those lines?

Tim Duignan (52:51) Yeah. That's a rough order of magnitude. Yeah. So I think technically we were at 4 moles. So the molarity of pure water, I believe, is around 55 moles per liter. So if you're at 4 moles, you got 10 water molecules for every salt ion. It's probably reasonable.

Nathan Labenz (53:06) Gotcha. Okay. So now we got 1 ion per 10 water molecules. The idea with the removing of the water portion of the all atom simulation. So you're running in the simulation, you're you are computing all the water molecules, all the ions. But then is the intuition that the main forces come from the ions because they're actually imbalanced in charge, whereas the water is neutral and it has its polarity, but it's not the because of the fundamental neutrality versus the imbalance in charge. Is that where my intuition should come from that I might be able to abstract away from the water in the first place?

Tim Duignan (53:45) Yeah. So the forces from the water molecules are actually very large though. So, yeah, they actually play a really important role in determining the the force on the ions. So the technical thing we're doing is it's called integrating out or, you know, probability speak, it's marginalization. So the water molecules are still implicitly accounted for. It's not like you're ignoring them, although I may have said that technically. They're more kind of their average effect is still accounted for in the simulation, through a kind of average effect. And that's the real challenge of it is that these watermelons do play a big role in determining the interactions of the ions, but it's just you can account for that effect in an average way is the basic idea. That rather than treating, so that you've got these 2 ions interacting and rather than saying, okay. I need to know how every single water molecule around them is arranged to calculate the force on them. I can say, on average, what would the force between these 2 ions be if I average over all the potential configurations of the water molecules around And that's the really challenging thing. So that's why it's so incredibly hard to build these continuum solvent models because the water molecules are actually critically important. So you're right. The water molecules are neutral and the ions are charged, so you get these large charge charge interactions. The But water molecules are absolutely critical there because they actually reduce the size of that charge charge interaction by, you know, a factor of 80. So it's absolutely critical. The water is absolutely central there, and it's really, really hard to get rid of them. That's why it's so hard in my PhD trying to find ways of ignoring them. Right? It turns out it's just really hard to do. But that's the beauty of this these neural network potentials is that they can somehow automatically average over all of the effects of all of these water molecules, and yet still get you an accurate prediction of what the forces are on the ions are gonna be on average. Right? When you average over the the behavior of all of the water molecules. And so that's what yeah. It's really

Nathan Labenz (55:26) So did that suggest if I'm envisioning what's going on in solution, the ability to average in this way, which I guess procedurally is like you're taking this all atom simulated output and literally just removing the water from it. And now you're training the network on purely the positions of the ions. Yeah. That's what you know. Pretty simple. So each ion then I should guess I guess I should understand this like being surrounded by a lot of water model molecules most of the time. It's not like it's randomly interacting with 1 molecule at a time because then it would seem really hard to abstract away. It's like in a shell of water molecules.

Tim Duignan (56:05) Yeah. Exactly. It's kind a state. Yeah. They're surrounded by many water molecules. And so you have to give the neural network a lot of examples of the positions and forces on the ions alone. And you need to do that so that it can average over the water molecules. Right? So if you only gave it a few examples of ions without the water molecules, it wouldn't it wouldn't be able to do it because it needs to know all of the different ways that these water molecules could be arranged. Can we average over that? So you actually to train the coarse grain model, you actually need a lot more data or a lot longer simulations. The data is obviously smaller because you only need the ions, but you need a lot longer simulations to generate the training data for training these coarse grain models because they have to do this averaging over the all of the water molecules. So in a given position of ions, you don't have a unique set of forces because it's gonna depend on the waters, but you do have a unique average set of forces. And that's what you're getting the neural network to calculate, is average forces on all of the ions in these positions. Whereas the normal way we use the neural network potential, which we talked about before, there is actually a unique force on every single atom in the system which comes from solving Schrodinger's equation. So you're kind of using the neural network in 2 slightly different ways. 1 is to compute the exact energies and forces on every single atom, and then the second way you can use it is to calculate the average energies and forces on these things. And technically, we're talking about here is actually free energies. So we're using the neural network to calculate free energies rather than pure energies. There's some underlying statistical mechanical theory that kind of makes all of this make sense and makes this rigorous, which is quite nice.

Nathan Labenz (57:38) Okay. Very interesting. So now through the crystallization, how should we understand this? I mean, the it is a for 1 thing, it's a phase change. Right? You talked about, like, orders of magnitude, amount of force being exerted by the ions on 1 another when they're in crystal form versus when they're in solution form. I think you said it was, like, 80 times or something?

Tim Duignan (58:04) Yeah. Exactly. So you've got the steam water. The ions obviously come apart, and they dissolve, and they're floating around freely. And the water in between them actually reduces their interactions by this factor of 80. It's called dielectric screening. Whereas in the crystal, that's not gonna happen at all. There, you know, you have much stronger forces kinda sticking these ions together. So that's why I thought that it would never be possible to to look at crystallization and liquid phase with the same kind of continuum solvent model. So that's why it was kinda blew me away when we saw it happen. Because, obviously, you could see these crystallization phenomena in the all atom case if you kind of drive it to happen because there, you know, you don't have this problem with the water molecules kind of being very relevant in the solution and not relevant to the crystal. So that's what was so surprising about the fact that it was able to do that. We still don't have our head around exactly why it can do these things, is that something we're dig into kind of further? What I suspect is that because in the solution phase, because you get nice kind of peers forming and then these triplets forming and then possibly larger clusters forming transiently, that the neural network can look at those and kind of extrapolate and say, oh, I see what's happening. When these piers are forming, that water squeezed out, and therefore the forces are getting larger between the ions in most cases. And so it's kind of then learning that information implicitly, and then it can take that and extrapolate to what it's gonna look like in the crystal. So that's what's kind of interesting. Now but an important caveat is that it cannot do this kind of quantitatively precisely. You know? So I don't think, and in fact, we know that it's not gonna get the quantitatively exactly things like the solubility, you know, how much salt you need to make the crystal precisely right. That's not the kind of thing it'll get right because it doesn't have any training data on the crystal itself. You know, there's gonna be small quantitative areas there. So that's that's an issue, but that can still be fixed, I think, fairly easily. Once that you've seen the crystallization, then the obvious next step to do is to take those crystal structures, and then you could add that to your training dataset and do a thing called active learning and try and get those things quantitatively. So that's kind of where we wanna go from here, I think, to start trying add data on the crystal and see how precise we can get these things in terms of actually making precise predictions of experiment, which would even be even nicer. But, yeah, this stage is still a bit of a surprise that it could do this at all. So still just trying to work out exactly what's going on, where it's getting the information it needs to kind of predict what the crystal is gonna look like.

Nathan Labenz (1:00:17) I wonder if we should think about this as the network generalizing out of distribution or if this is sort of a property of like the outer loop? Because I mean, in in like language models, you're probably familiar with this. It's pretty well established at this point and certainly listeners heard me talk about it many times that there are these higher order concepts that are being learned, you know, purely as a means to next token prediction. But nevertheless, like, we can now go in and identify these concepts, first autoencoder techniques, and representation engineering, and even demonstrate that you can steer a language model by injecting 1 of those directions or activations or or features is a better term than activation. You can inject that at runtime and you can steer the output of the model based on that. Sounds like that's not quite what's happening here. I I don't think from your account that I should expect that there is a feature that's been learned in the network that is like the crystallization feature or concept, but rather it it would seem to be more of just like as things sort of approach a certain position, the forces are such that they lock into a grid and stay there for at least some amount of time. Does that seem right, or could we even resolve that with what we know?

Tim Duignan (1:01:44) Yeah. I mean, this is a fascinating question. I definitely don't have the answer. So the way to think about it though, I think, is to think in terms of, you know, matching surfaces essentially. So when we're trying to compute the energies from the positions, you can think of that as trying to match what's called we call the potential energy surface, which is the basic idea that as a function of the position, which is just a vector, right, the positions of all of your atoms, that has some energy value, and this is a high dimensional surface. You can imagine it's like a single value that you map to from the positions of all of your atoms. And it is exactly the same thing with these coarse grain simulations where it's now a free energy surface that you're trying to match. And so in the simplest possible case, so imagine you just had, you know, 2 positions. So you have a single atom and it is in 2 dimensions, that actually you can visualize that exactly as a surface, right? Every 2 points in that 2 dimensional plane, there's some energy, which you can imagine is kind of a height of that surface. And then you can generalize this to higher dimensions. And this is what the neural network potential is actually learning. It's learning that surface, right? And it turns out that the neural network potential, you can train it on 1 region of the surface, and somehow it can extrapolate to what it would look like in other areas of that surface. Okay? So then you're only training it where it's in the liquid phase and it's kind of matching the surface as best it can. It's got these kind of equivalence and these various constraints, which is trying to use to match that. And it turns out that somehow that that can also do a good job, at least approximately, of predicting what that surface is gonna look like when it's in the crystal structure. So you can imagine the crystal is some like a potential energy well, we call it. There's some like sharp dip in that surface, which means it has a lower energy, which means it's likely to be in that state. And so that's the way to think about what it's learning. It's actually learning that high dimensional energy surface, and then that's what it's using to then do these dynamics, because you use MD to kind of move the atoms around on the surface, and the atoms like to be in lower energy states, right, that's what statistical mechanics tells us. And so that's kind of what it's if you want to think about the concept as learning, although this is a very vague idea, what is even is a concept, you know, but that's kind of what intuitively I would think about what it's learning. And now there's interesting parallel with this to large language models, right, which is that people often talk about large language models as fitting a surface as well. Right? They're just fitting a different kind of surface. In fact, they're fitting, you know, the probability surface of next words. Right? So that's what these GPT, these big transformers are doing. They're predicting the next word, and that's just a probability. Right? That's every vector that represents your sentence, every word has another probability. So that's a probability surface vector fitting. And in fact, this actually has deep connections to statistical mechanics because the log of your probability density in statistical mechanics is actually an energy, right, to do with thing called the Boltzmann distribution. So there are actually deep connections between what's going on with these large language models and what's going on with the training these neural network potentials. Now when it comes to, yeah, this idea of these concepts or this, you know, the the fact that these large language models seem to have this ability to kind of have some notion of concepts and generalize out of exactly what they've been trained on, which is kind of this surprising phenomena. I mean, I would speculate that it may be something similar going on. I mean, that it in a sense that when you learn this kind of probability surface very accurately, that then allows you to extrapolate to other regions of that probability surface and to make predictions there. And so probably, you know, I mean, we don't really fully know what's going on in these things, but there may be some parallels there. You actually see some evidence of this in the way these large language models work. I mean, they use a softmax layer in them, which means they're which is their softmax is this normalization layer. That is actually a Boltzmann distribution they're using. That's actually an idea that comes from statistical mechanics that comes from these MD simulations. And so there are, you know, they always talk about log probabilities. You know, you hear that very often. Well, a log probability in statistical mechanics is a free energy. So there's actually some deep mathematical connections going on here. I don't have my head fully around it, but I I really think this is a direction is worth pursuing is thinking about these connections between statistical mechanics and large language models, because I think there may be some kind of deeper connections going on there.

Nathan Labenz (1:05:47) Yeah. Okay. Definitely very interesting. So more about the generalization potential. You had another phenomenon that you reported as well, which was self ionization of water. I guess we're now back to the all atom case there. But tell us about that.

Tim Duignan (1:06:08) Yeah. So that was another kind of thing that really kind of blew me away and surprised me. And now that people have actually seen this as well, and I think no 1 really knew quite what to make of it. But it's very interesting. So when you run these water simulations, so now you're you're including all of the water molecules, and in this case, yeah, we're just trained on, you know, pure water or maybe some ions of water. When you run these simulations for long enough, you actually see this phenomenon called water self ionization. And basically what that means is the hydrogen pops off a water molecule onto its neighboring water molecule, and it forms an Oh and an h 3 0. Right? So instead of 2 h 2 o's, you have these 2 separate what they're now ions. And this is a phenomenon that we know happens in real water. In fact, about 1 at a degree, I think it's 10,000,000 atoms or 10,000,000 water molecules will be in this kind of ionized state, and it's very, very important. That's the pH. That's what the pH of water 10

Nathan Labenz (1:06:58) to the minus 7. Yeah.

Tim Duignan (1:06:59) 7 to So we know that this happens in real water. Right? And so the remarkable thing is the neural network potential is actually kinda predicting that will happen. Right? It's saying, oh, you know, possibly what you know, is running these simulations, and then every now and then, occasionally, 1 of these hydrogens is hopping off onto its neighboring atom. And that's not something that it's seen in the training data, and yet somehow it knows that that's kind of a fairly plausible thing that might happen. And then the really impressive thing is it keeps the correct shape of that hydromium ion, that h 3 0 ion. It keeps it in this kind of pyramid type shape, which is really surprising. I would have thought it would kinda once you get outside of where it's seen, I would have thought it would just kinda break and go crazy or do something very unphysical, which you do some you do see quite often, honestly. So I would have thought it would break and not be able handle that, but it actually keeps running. The simulation stays stable, and then you see the thing called growth is hopping, which is then another hydrogen will hop off from that HgO to its neighbor, and you get these chains of hopping hydrogens, which is actually this really important mechanism which explains why acid diffuses through water so quickly. So this is actually a really important physical phenomenon. The neural network potential is just kind of guessing that it's gonna happen, which I thought was quite impressive. Now this is very well studied, so it's not really telling us anything new, but it does imply that maybe there will be cases where it could find things or suggest things that are new that we haven't thought of, which is quite an interesting idea. Now but there's a big caveat which I need to mention, which is coming back to the same point about crystallization, that it doesn't get these quantitatively precisely. Right? So it's not gonna predict the rate at which this occurs exactly. In fact, it does it much too fast. So it shouldn't really happen in a normal simulation. You shouldn't see this happen on the kind of, you know, this picosecond time scale we normally access or we can get to nanoseconds or much longer with these NMPs within your network potentials, but you still shouldn't really see it. You should have to force it to happen. So the fact that it happens is is not necessarily physically correct. And so that's an important caveat that, you know, it can't get things quantitatively precisely outside its training data, but it's almost you know, it's not necessarily a bad thing if when it goes outside its training data, it can still make reasonable guesses and suggest interesting phenomena that might occur. And that's something that we can easily fix if we want to. And in fact, people have now fixed this and trained these neural network potentials explicitly to get this process right. So you add training data exactly on this hydrogen ion hydrogen atom as it comes off the water molecule, get lots and lots of quantum chemical training data on that, and then simulate the whole thing very well, and you can actually precisely predict now that that pH 7 belly of water, which is very nice.

Nathan Labenz (1:09:17) So how does that process work? I guess I'm imagining when you're generating the training data via the simulation, you could, like, take 1 initial condition and just run it indefinitely, but I guess you could also set up a bunch of initial conditions intentionally and then run from a bunch of different starting points. I'm guessing something like that is what's happening where people are saying, let's bring 2 water molecules into the right, you know, form the right configuration for that to happen so we can start to have more of that in the training data. Is that right?

Tim Duignan (1:09:51) Yes. Exactly. So what I normally do is just start with what's called an equilibrium molecular simulation. So you just give the atoms, you tell it, you know, the temperature, room temperature, and then you just run the simulation and generate your training data from that. But that won't sample what we call kind of rare events, which are things that happen, you know, very infrequently. Things like crystallization or this or this hydrogen hopping off the water molecule type phenomenon. And so there are various strategies you can use to kind of get some of those more rare events. 1 is you can run it at higher temperature, or you could, as you say, run a ton of these things kind of in parallel and hope that that explores your phase space better with different initial starting coordinates. Another thing you can do is called enhanced sampling, where you kind of drive these systems out of their local minimum. You know, the positions they'd like to sit in, you kind of force them to different arrangements and things. But then another really nice idea, which probably is the most powerful 1, is this idea of active learning, which people have shown now works really well for these. And the idea here is you actually you train a neural network, and you go and run much longer molecular simulations with that neural network. And then you have some method of estimating the uncertainty in the predictions of the neural network. And then whenever that uncertainty spikes, whenever it says, oh, I'm not sure about this 1, then you take that data, and then you go and add that to your training dataset and do new quantum chemistry calculations calculations on on it, it, and then you can expand your dataset to cover kind of these rare events that don't normally happen in the simulation when you're just running these short ones to keep your training data. And people have shown now that this is a really powerful technique and you start to kind of and you can imagine this is kind of a feedback loop. You can start to build your training dataset, cover more and more space, get cover more interesting phenomena, and then you can start to think about simulating things, yeah, like chemical reactions, which is really important. So people now are doing some really nice things on this simulating, you know, catalysis, so the reactions on on surfaces using these techniques, and they're getting really powerful. So that's really exciting as well.

Nathan Labenz (1:11:42) Yeah. That's interesting. That honestly reminds me in some ways of just down the fairway LLM app development that I've done at times where there's this sort of bootstrapping loop in a lot of cases where especially if you're working with a pre trained model as we normally are in a language model case, you don't need that many examples to fine tune. And, you know, I often recommend, like, as few as 10 can be enough to do that initial fine tune. And then it's all about identifying edge cases and supplementing that dataset. And

Tim Duignan (1:12:13) Exactly. Yeah.

Nathan Labenz (1:12:14) Yeah. Doesn't have nice more quantitative way to identify your edge cases than I typically do.

Tim Duignan (1:12:19) Yeah. Because the problem with the you know, the hallucinations and things is there's no just function you can go call and go say, you know, what is the ground truth here? And so, you know, it's very hard to tell or or to add to your training data the correct things. Often, you'll have to go and get a human to kinda label those things. I guess this is the way you do the same thing for a large language model. But the nice thing here is you can kinda do it all within a computational loop. Right? So you can kinda automate it and get it all internally happening on on on, you know, 1 supercomputer. And so you can really start to get these things scaling up in a kind of automated way, which is quite nice because you've always got that kind of ground truth you can go and just automatically refer to them and calculate, which makes it a kind of really nice powerful technique, I think.

Nathan Labenz (1:12:56) Yeah. So if we're seeing this level of generalization where we're seeing these self ionization and crystallization phenomena happening with really rather modest compute resources to generate these datasets, then I guess I wonder how far can this go? And you're probably wondering that too. But has it gone any farther than this yet? And what would be prospects for I guess there's a lot of different dimensions. Right? I guess I don't really know how you're dealing with temperature. You said a second ago that you can you know, you put your temperature into the simulation. Do you also put the temperature in as like a free parameter into the network? Do these networks have or have the prospect to accept a bunch of other free parameters like temperature, pressure, whatever else?

Tim Duignan (1:13:40) Yeah. So the really nice thing about this method is that the potential energy surface, you know, these energies and these forces that we're learning, those don't depend on temperature. So when you solve Schrodinger's equation, it doesn't know anything about temperature. Right? That's just given the positions, give me the energies and The way that temperature comes in is actually in this this classical motion solving, you know, solving Newton's equations where you calculate how the atoms are moving. That's where you add a thing called a thermostat, which just basically controls the temperature. Right? Just like a normal thermostat. If it's getting too hot, you'll remove some energy. If it's too cold, you'll add some energy. So it's actually really trivial to run these things at high temperatures. So we've run all these electrolyte solution simulations at different temperatures, and they and they look perfectly stable. So that's quite exciting as it should be, yeah, definitely generalizable to different temperatures and as well. Microsoft just had a big paper called Madison where they were looking at this, and they they showed, you know, you could simulate over a huge range of temperatures, which is really nice. So definitely, yeah, that's the kind of directions we wanna be going in. Right? Because this is where it gets very hard experimentally. You can you gotta measure everything at different temperatures, and then when things get hot, you know, your equipment starts to run into issues. So this is a really nice aspect of this computation. I think it's almost trivial to change the temperature. Now you may in some cases, you probably need to add training data at these higher temperatures, say, or so there may be some cases where things break at different temperatures if you don't have enough training data. But in principle, yes, changing the temperatures is a very trivial thing to do. This is the nice thing about kind of coupling it with this more physical solver, know, this kind of physical engine, the physics engine, because that means that you don't really have you know, if you were relying on the neural network to learn everything, it would have to learn, like, oh, how does everything change with temperature. Right? And so that would be a much harder kind of AI problem, I think. Whereas in this case, you can kind of use the fact that you're learning something that's independent of temperature, which is a really nice feature.

Nathan Labenz (1:15:23) So if you have this thermostat and you can turn it up and down so freely, can you turn it down to freezing point and see everything crystallize or turn it up to boiling point and see everything turn to gas?

Tim Duignan (1:15:35) Yes. Yes. You can definitely do that. The issue with just cooling everything down to try and crystallize it is that you get this called amorphous crystallization, basically. So like a glass. You know, if you take silicon and cool it down too quickly, it amorphously crystallizes and forms like a basically, just like a solid liquid. Just a liquid that stops the atoms stop moving very much. So if you just turn the temperature down too quickly, that's often what you see there. But so it's not it's not as easy to induce crystallization as you might think. But, yeah, that's certainly something, yeah, that we wanna explore more. Yeah. Can you do those kind of things? Excellent.

Nathan Labenz (1:16:06) When I freeze water in my freezer at home, I'm getting an actual crystal. Right? Or am I getting a sort of water glass?

Tim Duignan (1:16:14) Yeah. You can get water glasses, but those have to be done in the lab because you have to freeze it very, very quickly, I think. In in in the normal forms of ice are a type of crystal. Yeah. They have a nice crystalline structure kind of where you get this hydrogen bonding structures forming. So, yeah, you're getting a crystal when you freeze it at home, but there are certain types of ice. You know, there is actually a whole phase diagram of ice. It can form all of these crazy different phases of different temperatures and pressures and speeds. So if in a simulation, you know, because we're using these picosecond time scales, if you just turn the temperature down too quickly, it will just you'll probably just get an amorphous freezing. But people have people are now using these neural net potentials a lot to look at ice phases and ice water equilibrium doing all sorts of fantastic work there. So you can definitely probe the crystallization of ice water. And there's actually a paper so there's some precedent for the salt crystallization we saw. There's a nice paper where they looked at simulating water, and they showed basically the same thing. You can see little water structures that look a little bit like the ice structures. Structures. And so you can actually train a neural network on water and you use that to say interesting things about ice as well. So there's probably similar parallels going on with ice crystallization water crystallization, which is ice. Yeah.

Nathan Labenz (1:17:22) So what other dimensions of generalization do you see or expect? Like, you could imagine training on 2 kinds of salt solution and trying to predict what happens with a third. I mean, obviously, you could really run with this thing a lot if you're scaling up the datasets. And it sounds like there are quite a few orders of magnitude of scale up that would be, like, start to get expensive, but certainly not out of bounds of the sorts of models that we're seeing trained today. You're talking about a couple days of compute time here or there. Certainly, you probably have 1000 x room to go that you could, like, plausibly write a grant for. Right? What would you expect to see as you go up these orders of magnitude?

Tim Duignan (1:18:05) Yeah. Definitely. So that's the interesting thing, yeah, moving forward. And people are already doing this. In fact, bite dance actually, you the the company behind TikTok actually had a paper on electrolyte simulation on exactly this. Right? They're trying to do the scaling up. Right? So they're not I'm I'm trying to get these things really, really accurate and precise. They're I'm trying to scale up and look at many different electrolyte solutions, and they build a big database of electrolyte solutions, and they saw exactly this. They could train their model on certain electrolytes and then generalize it to new ones that hadn't been seen in the training data, which is very interesting. So they were simulating organic electrolytes, I think, where it turns out, you know, these things are made up of carbons and hydrogens, you can kind of extrapolate from 1 to another, which is quite exciting. So, definitely, we're starting to see evidence of the ability of these things to general generalize to new cases where they don't have the training data on them because there's enough similar examples where you do have the training data. And so that's really the exciting direction to go in. And, yeah, so this is starting to become a big field. You know? This and you can people talk about these things as the universal kind of machine learning potential. And the basic idea is you wanna train 1 model that then you can put any elements into it and any kind of arrangement of atoms and calculate the forces on every single 1 of them. And so it's quite different to what I'm doing where I'm trying to build a model just for 1 particular case. They're trying to build these universal ones that could handle any arrangement of atoms. And this the big tech companies are starting to get into this as well. You know? So the ByteDance has the electrolyte 1. Microsoft has 1 called Matasim just came out. Google had this nature paper recently where they trained 1 of these called Gnome, and they showed that they could find, you know, millions of new crystal structures, which was very, very cool. Meta are working on 1 of these as well. So, yeah, all of these, this is starting to really take off now. It's the kind of thing that I think maybe industry is better placed to do because they have these huge compute resources, and they're starting to scale up these models and build these really universal ones. And what we're hoping, and I think we're starting to see evidence of this, certainly these papers are starting to show evidence of this, is that you'll see the same thing that you saw with the large language models when you scale them up. Right? That, yeah, exactly as you were saying earlier, you don't need to add nearly as much additional training data once you've got this good, really good foundation to to start with. You know? So it's learning all these kind of general features about the ways atoms interact. And then if you wanna get if you wanna study a new system, just adding a little bit of extra data on that new system will then get you a really accurate description of it much more quickly than you would need if you're starting from scratch. And so that's really exciting because you can start to think about, well, this could be a universal tool that we start to use as kind of first estimate to get an estimate of what any system is doing. Need to just go straight to 1 of these universal force fields and simulate it if they could run them quickly enough and start to get a picture of exactly what's happening at the molecular scale for any system you're interested in, which is a really, really exciting idea. I think it could be a profoundly kind of transformative scientific tool once you get this thing really up and running efficiently and and get it in people in the hands of scientists who can really use it and to do exciting things with it. And there will always be cases where it breaks and where it fails. Right? Just like check-in-depth, hallucinates and goes wrong in different places. You know, there'll be systems where where it has problems and has issues, and so you'll need to keep in to keep it in this kind of interactive loop where you're always adding extra training data, going in, extra quantum chemistry calculations, and then essentially fine tuning it. Right? Fine tuning for particular applications is probably what we'll have to keep doing, but it's a really exciting future, I think, as people start to scale these models up. And there's a nice 1 called Mace 0, an open source 1 recently, which is very cool. And the remarkable thing about that was they trained it on crystal data, so solid data. It's kind of the inverse of what we saw. We trained on the solution and saw crystallization. They trained it only on crystals, and they showed that you could simulate liquids with it. So that was very exciting too. So I think these things are are really gonna start taking off in the near term future, and that's really exciting.

Nathan Labenz (1:21:35) Yeah. There's a couple of big themes there. 1 is the possibility of what I've term I borrowed from a friend, the big tech singularity, where it seems like we we've seen plenty of, fiction over the years where, like, 5 companies take over the world or whatever. But this is another dimension on which, man, if you really just have to scale it up and there's only a handful of companies that have the ability to do that, these big tech companies potentially start to come in and dominate material science. Right? I mean, is that kind of what you would expect the future of the field to look like?

Tim Duignan (1:22:11) Yeah. I think it's certainly a possibility, and I and I think it's a risk. And I think, you know, the governments and and the public sector, yeah, needs to get onto this and start saying, okay. This is a serious area. We need to fund, you know, independent groups who aren't, you know, in private companies to build these things as well, build these tools as well so that we can have experts who are who are outside these companies who really understand these things. I think that's a really important thing that needs to be done. Yeah. I I think, you know, there's no reason we can build you know, there's, like, soon, you know, that the large hadron colliding. Know, there's examples of huge science projects, you know, where the public sector put large amounts of investment and built these big things. So I think there's no reason that we have to leave it to these big companies. Right? It's not the kind of thing that's suited to, you know, maybe academia. We have these small groups with 1 professor and a and a number of PhD students. Like, I think we need to move away from that model, but I think there are examples of science. You know, these big interdisciplinary collaborations, you know, we could start to think about getting all these people together. And and the paper was a really nice example of that, where they got, you know, huge team of people together all collaborating to build 1 of these foundational models. So I think there's hope yet that we can build these things and open source them. But, yeah, I think there is really a possibility that these things stay internally in these companies. And so, you know, we saw this big debate around this with AlphaFold 3 where they're, you know, they're hanging on to keeping those models private at least for now. And so that'll be a very interest interesting to see how this plays out. Will they release these models to allow the scientists to use them? Because because a big concern I have is I've trained these big models. Right? There's so there's so many incredible simulations you could run with these things. And yet, internally in these companies, you know, there's not that many people to kinda run these simulations. So I think it would be way more valuable if they release these tools to the public or to to others other scientists in academia to use them. But having said that, you know, these things take a huge amount of resources to build, you know, if you're spending, you know, huge amounts of GPUs on building these datasets, you know, that's a a big investment of money, and it's understandable that these companies, you know, wanna recoup their investment. So, it's a complicated issue, and it'll be very interesting to see how it plays out. But but I think definitely there's room for kind the governments and academia and the public sector essentially to get into this area as well.

Nathan Labenz (1:24:09) Yeah. Another big theme is the fact that China is not really all that far behind. I am 1 who wants to avoid the AI arms race with China almost at all costs, maybe not at all all costs, but, you know, in terms of ways the future could go badly, I feel like a throw caution to the wind AI arms race with China would be a bad 1. And 1 way I think we might trick ourselves into going down that path is if we dilute ourselves into thinking that we have a big lead and we're gonna win the race. And so I always note these examples where, you know, Chinese big tech companies or whatever academic groups in some cases as well are producing top notch stuff. And as far as I can tell, don't actually seem to be all that far behind. That could change, but I don't even need to comment on that unless you have 1,

Tim Duignan (1:24:55) but it is notable that

Nathan Labenz (1:24:56) they're not far behind on this dimension either.

Tim Duignan (1:24:58) Yeah. There's definitely a lot of fantastic research in this area coming out of China. Yeah. And it, you know, I think it has big big economic implications. You know, you could start to really do materials design, understand biology. You know, there's so much you know, whenever you get big scientific breakthroughs, it has massive technological and economic implications as well. So, yeah, I think it's something that we should really be thinking about a lot more. And, yeah, how to avoid the arms race? I'm not sure. That's a very tricky question that's above my pay grade, unfortunately.

Nathan Labenz (1:25:25) Unfortunately, I don't think there's any designated adults either that we can see that question too. Regardless of pay grade, we might all have to take that 1 up.

Tim Duignan (1:25:33) Yeah. That's

Nathan Labenz (1:25:33) true. Our little part on it. Is there, like, a safety concern with these sorts of models? I mean, obviously, the people talk a lot about highly generalist models and are they gonna do what we want them to do and not? And then with biology models, you have sort of inherent dual use of like, if you can cure a disease, can probably also cause a disease. Is there something similar here or is this more of a 1 way bet type of technology? Like, I have a little bit harder time imagining this sort of harmful material science use case. But maybe it's that maybe I just have limited imagination.

Tim Duignan (1:26:07) No. No. So, I mean, 1 point is that these neural network potentials are not at all just limited to material science. Right? People are also using them in biology to to really interesting effects. So there's a paper from the David Baker group in science recently where they used any 1 of these neural network potentials to design new drug molecules. That seemed to work pretty well. And the state of the art, in fact, for this problem of drug binding to proteins, you know, calculating how strongly a drug molecule will bind to a protein, recent preprint actually showed that the state of the art was using 1 of these neural network potentials, at least for the core component of it, for that actual drug molecule. So these things definitely have implications for biology as well. So just the same pathway that, you know, biology and AI, you know, is a very powerful tool could be potentially a concern there, how you could use these neural net potentials to design, you know, drugs and things that could potentially be harmful. So, yeah, that that's definitely a concern. Yeah. In terms of material science, yeah, I don't think so much it's not probably not so much of an issue there, you know, other than to say that, you know, material science is really important for, you know, a lot of defense industries, obviously, if you're trying to make, you know, various new materials. You know, a key limitation of a of a lot of things is is the ability to design new materials. But I'm not too concerned about the safety things at this stage. I wouldn't be too much too worried about it. But, yeah, you can imagine at some point when these tools get powerful enough that they start to have, you know, enough technological implications that it's something

Nathan Labenz (1:27:29) you wanna start thinking about quite seriously. Have you seen the recent paper on cans instead of the tag mark group?

Tim Duignan (1:27:37) K I n's. Yeah. Yeah. That's a very

Nathan Labenz (1:27:40) interesting paper. But it's basically new architecture. Right? Where instead of learning the weights that are edges between nodes in a network, they are learning functions that sit on those edges. So instead of just a a single number at each edge, it's a function at each edge. And then instead of having an activation function on a layer by layer basis at the nodes, they're just adding the inputs from the different learned functions on the edges. So this makes more intuitive sense if you look at a diagram, which they have in the paper and definitely recommend that and hopefully, again, have an episode coming soon about it. They seem to think that this is going to be most applicable in science because it seems to address the sort of composability problem where you can now learn, like, functions of functions and, you know, do all sorts of interesting more and it's also, like, much more interpretable because you can actually just look at the shape of the individual functions that are being learned. Do you think that has promise for the sort of work you're doing, or how would you connect the dots between your work and that new potential tool?

Tim Duignan (1:28:50) Yeah. That's definitely really exciting. I think it certainly has potential. I really like this idea. The problem essentially is a simple word, it's like fitting a surface. Right? It's a very high dimensional function. You have some the positions of every single atom and you're trying to predict the energy, just a single number. And so that's exactly what these, you know, k a n's are kinda do, trying to find good approximations of these multidimensional, high dimensional functions. And so that yeah. It it makes sense to me as an approach that could really perhaps help with the generalizability and build these things that that are generalizable. And I like that I think they showed in the paper that you could improve the generalizability without spoiling the the what they'd already been trained on. Whereas with some of the normal multi valoreceptron type approaches, perhaps you would lose some of what it had been trained on previously when you added additional data and fine tune these things. So, yeah, there's potential advantages there. I definitely think these things should be explored. I think a big issue, though, is obviously the the computational limitations. What's gonna be computationally most efficient? Yeah. Because in the end, we are looking to run these things as fast as possible. And so you really need to demonstrate, you know, that you get some really solid advances in order to justify kinda switching to a new architecture. But I definitely hope someone is trying out building a a neural network feature with these cans. I think that would be really cool and really interesting to see how they perform. I think that makes a lot of sense. I think, yeah, my kind of philosophy and the thing I love about the machine learning community, which is that it's it's very empirical. You know? So there's a lot of trial and error and just empirically testing things out, seeing what works well, and seeing what doesn't, but then also kind of using kind of fundamental theory where you need to is inspiration for coming up with new ideas. And so that can was a nice example of that. But it'll be interesting to see, yeah, how it plays out if they are actually generalizable. Saw some people arguing that they were kind of you could map them to a simple multilayer perceptron in in a kind of in a way. So whether they are genuinely give you additional new properties that that or new behaviors will be very interesting to see or whether it's just a matter of, whether they reduce training data requirements. It's always very hard to predict in advance, you know, what what's gonna work and what's not. So it'll be great, interesting to see, but I think this just needs to be empirically tested. Yeah.

Nathan Labenz (1:30:51) There's definitely a strong trend of some sort of fundamental equivalence between a lot of very apparently different things.

Tim Duignan (1:30:57) Yeah. Sometimes I get frustrated. I feel like there's some deeper theory missing underlying these things. You know? And I I think it would be really great if if we could get a more rigorous description. And I I know a lot of people want this. Yeah. That's exactly why these neural network potential or, you know, neural networks in general work so well. When do they work? When are they not gonna work? You know, that's really lacking. I don't know. That's a big area of research, so it would be very interesting to see what comes out of that.

Nathan Labenz (1:31:20) So just zooming out, like, how do you expect all this to shape the pace of discovery in science over the next few years? It seems like you've touched on a number of obviously super high value use cases. And I guess the general expectation very similar to the, you know, episode we did on the biological applications of this and this sort of overlaps with that, but it's definitely more fine grained. I guess the loop is like, we're going to run a lot of experiments on the computer. We're going to identify things that look good. Then we'll do actual experiments with those to try to verify that they're right. If they're right, we'll have new discoveries. If they're wrong, we'll kind of feed that back end and patch the models and try to get better on those failure modes. And it seems like in biology, we're looking at potentially 1 to 2 orders of magnitude speed up in terms of how much, like, overall resources, you know, human and and financial maybe would need to go into discovering new things. Maybe even more than that. It seems like 1 is kind of a bear case and, you know, 2 is maybe like the mainline case and above above 2 orders of magnitude would be like obviously a Super Bowl case. Is that your expectation? How do you expect the the sort of trajectory of technology progress to change based on these tools?

Tim Duignan (1:32:51) Yeah. I mean, I'm really excited and really optimistic, frankly. And and it's you wanna be careful because you don't wanna overhype things and, you know, other people will disagree. Obviously, a lot of senior professors, in fact, you know, are quite skeptical of a lot of these things. But in my opinion, I I think they're genuinely a a transformably useful new tool, which will have a profound impact on, you know, every discipline of of fundamental science, I think. And so I'm really, really excited about them. The central challenge you know, AI is clearly an incredibly powerful tool, but the central challenge of it is the training data requirements. Right? You always you know, these are very data hungry beasts. So you wanna be thinking about where are the problems where you can generate large amounts of data efficiently and and increasingly diverse but very high quality amounts of data. And so I was stuck with, you know, machine learning and AI for science for a long time because for electrolyte solutions, you you actually have not that much data. You know, it's fairly limited amounts of data that you can use to train these things. So I didn't really get into this until around AlphaFold came out quite late. But since then, it's become, you know, really clear that you wanna that there are cases where you can, yeah, exactly as you're saying, have these stages where you can iteratively generate larger and increasingly diverse datasets. And 1 way to do this is with kind of automated labs and or having experiments in the loop, you know, where you're constantly generating extra data. And there's an interesting article in the New York Times about, you know, it's a pharmaceutical company where they have these, you know, huge automated labs where they're generating all of this new data, which is really 1 way to go, and I think that's really promising. And then the other way to go, I think, is to use kind of quantum chemistry fund fundamental physics to generate this data on the fly as well. I think that's a really nice, exciting avenue. And I honestly don't see, like, what the limitation is. You know? I'm I'm there's probably things that are that are unforeseen, know, that'll cause these things to to not be useful at some point maybe. And but at this stage, don't really see what the limits are in terms of, like, as you scale these things up, get larger and bigger models, why you won't be able to simulate a really large range of interesting systems and really get a lot of insight into many, many different systems that we've been waiting for a long time to really understand. Because if you look at the way, yeah, traditional experiment works, drug discovery in particular, there's huge amounts of trial and error. So working with experimentalists, there's a lot of just intuition and guesswork and trying different things and then getting some interesting results and then saying, well, we don't know what's actually happening at the molecular scale, and then trying to work out some way to model it. And there's this really just, you know, you can just feel the progress being held back by a lack of ability to just really understand what's happening, particularly this molecular scale was so critical with so many important processes happen. So if you have this ability to simulate these processes really fast and really accurately, I I think that could be really transformative and useful, and I'm critically excited about it both for biology and for, you know, material science and chemical engineering and even things like climate change as well where there's a lot of important molecular scale processes go on. So I think it's a, you know, really exciting new time to be involved in science.

Nathan Labenz (1:35:43) You had a couple interesting thoughts on the importance of or the opportunity of taking tools from 1 domain and applying them to another. And I guess you can comment on that in in any number of ways, but I'm curious if you see, like, untapped opportunity right now of tools you'd like to see move from 1 place to another. We talked a little bit offline about the diffusion models roots in physical processes and how that's obviously now, you know, gone kind of horizontal and is in use in a lot of different areas. But tell us more about the sort of borrowing tools from 1 domain to another Past. Present, and future.

Tim Duignan (1:36:21) Yeah. So there's this brilliant book I was reading recently called The Incomplete Guide to the Art of Discovery by Jake Oliver, who is a scientist involved in discovering plate tectonics, which is this big paradigm shift in geology. We realized that the Earth's not static. It's made of these tectonic plates, which were all moving around. And he made this fantastic point that there's a really easy kind of hack to do really exciting science, and that's to find tools from 1 field or discipline and take them and apply them in your own. So this is what they did to discover plate tectonics, particularly taking a ton of tools from physics and things and applying them to discover that these plates were moving around. And I think when you start to look at science, you see this cropping up everywhere, you know, historically. So the classic example is, you know, physics itself, right, when Newton realized that if you took, you know, tools from mathematics, calculus, and then applied them to physics, you get this transformatively useful tool. Right? You could suddenly explain gravity and huge amounts of physics, which was incredibly exciting. And that's so kind of ingrained in physicists now that we don't really even think of that as being kind of interdisciplinary, but but it really still is a type of interdisciplinary. And then there are many other examples of this. Right? X-ray crystallography, for instance, all that you use to get crystal structures of proteins, know, as came out of physics and you're applying it to biology. So it's a very, very common feature of of important science. It's taking these tools from 1 field and applying them to another. And yet diffusion models are actually a fantastic example of this, where if you go back and look at the original papers that motivated these diffusion models, they're actually explicitly citing nonequilibrium statistical mechanics as inspiration for these models. And in some sense, using actually the mathematical tools of it. So using thermal annealing and using a thing called long term dynamics, you know, which is actually the equations or the algorithms that we use to do molecular simulation. Invented by physicists to simulate molecules moving around, and then, you know, people thought, oh, we could take this and adapt this to generating new images. And then you got these fantastic results. You know, we can make these incredible images now using these algorithms. So that's another example of this more recently. And then the coolest thing now though is that we're starting to go back in the other direction and kind of complete the loop, and you had this fantastic guest on the other week talking about, you know, all the ways these diffusion models and things were being used to impact biology and problems that we would traditionally have been tackling with normal molecular simulation. So you're starting to complete this loop where you're using these same tools again now back in biology, which is very exciting, and doing molecular simulation, which is very, very cool. Although you wanna be a bit careful because then it turns out that you may end up just doing molecular simulation again without realizing it, which I think there's a little bit of that going on. You're running longebond dynamics and you're using it, this thing you call the score. But there's a great paper out of Microsoft recently where they showed that if you train a diffusion model on molecular simulations, the score, which is the thing the diffusion model is learning, is actually mathematically identical to the force field. And so you've actually come back full circle, and you're just doing molecular simulation again. But now you've got a bunch of extra exciting nice tricks you can use to make it really more efficient and and make it more generalizable, which is exciting. So, yeah, I think this is a really profound example of this importance of interdisciplinary, right, and thinking across different fields and trying to take tools from 1 field and use them for another. And I think looking forward, another example of this is this idea of taking tools and I think understandings from statistical mechanics and actually applying them to understand the theory of large or, you know, of AI itself, machine learning of what's actually going on in stochastic gradient descent. What are things learning? What does it mean for things to learn? And I think there's a ton of tools from statistical mechanics which could really be brought to bear on this, and and we're starting to see some examples of this. So the paper just the other day showing that the the distribution your parameters go to when you do when you're training a model go to a Boltzmann distribution, right, which is a is a tool from statistical mechanics. It's also the distribution of molecules go to this Boltzmann distribution. So we're starting to see really interesting parallels emerge. And then another really lovely example of this was a Japanese group actually looking at the mathematics of phase transformations. So this is a we have beautiful theories which can explain phase transformations, which is exactly like crystallization. And what they actually showed us when you're training these neural networks, you start to see some of these same mathematical features emerge in these training of these neural networks, which is a very nice idea. So then you could start to think, oh, a dropping. You know, this phenomenon that seems so mysterious, actually, it has a lot of parallels to something like a phase transformation physics. Right? And when you look at something like a learning curve, it actually looks a lot like an energy curve when you're equilibrating a in these simulation. You know, it starts high, it drops down to a lower value, and it'll be jumping around some plateau with some noise, and then you make it a phase transition where it suddenly drops lower. Right? That's exactly what you see when you're trying to training a model and you see these rocking phenomena emerge. So this is a really exciting direction, I think, as well using tools again from different fields to try and understand things within your own field.

Nathan Labenz (1:40:53) You see something like Sora, and the OpenAI folks behind that are less interested in the video generation and more interested in the, like, world simulation. And, you know, it's obviously, like, macroscopic world simulation. What do you think is the outlook for like 1 model to rule them all across all these orders of magnitude versus when you see this core screening? Is that something that sort of inherently suggests that we're going to have a hierarchy of models that handle these different orders of magnitude of system, or can that all be integrated into 1 thing if you just dump everything into 1 1 dataset, 1 training run, 1 network?

Tim Duignan (1:41:37) Yeah. That's a fascinating question. Right? Yeah. That would be the dream if you just have 1 neural network to rule them all. Yeah. I don't know that we'll get there. I mean, the way I think in the future, you know, thinking long term how we'll do these molecular simulations is you use, yeah, kind of a hierarchical coarse graining approach. Perhaps you could incorporate this into 1 model, but we're so you would have a neural network that ignores the parts of the systems that aren't relevant, but then it will have higher resolution pictures of the system that are relevant. So you think about something like a drug binding to a protein. Right? You at the local region where that that's a really important process. Right? And you probably need an atomistic description of that. You need to know where every atom is as a function of time to simulate that. But further away from that protein, you could probably use a coarse grain description where you're not treating every water molecule, and you're only kind of treating them in kind of average way using this kind of coarse graining technique. And so I think that's really exciting long term thinking about how you can build models that kind of combine different scales, kind of multiscale modeling all within 1 neural network to both study the kind of atomistic detail where it's important, but then where it's not important and you can kind of average it out, it can do that kind of on the fly. And so that's really exciting. You can start to think about potentially, yeah, building models that can do both scales at once or look at multiple scales. Because there's a huge range of phenomena where it's not just happening at 1 scale. Right? Because molecular scale processes that then determine larger macroscopic things. And so you need to have this kind of multiscale approach, and that's a really challenging problem. It's gonna take us some time to work out. But I think at the end there, yeah, machine learning and AI will be a critical tool to kind of try and help us do that efficiently.

Nathan Labenz (1:43:07) Would it be like 1 would the first step in your world toward that be training a single network on the all atom versus and also the coarse grain as opposed to, like, separately? Have you tried that? No.

Tim Duignan (1:43:22) I haven't tried that, but it's a great idea. And then having it or allowing it to kind of automatically jump between descriptions where, you know, it thinks it's best. You know, maybe I need a all item description here, so I'll go back to a higher resolution picture of it here would be very cool. So that's, yeah, something I think, yeah, people should be thinking about more long term to try and do something more ambitious here. Because this is the real key problem is that you want to use this coarse grained picture so you can look at bigger systems, then also these microscopic phenomena are really important in many cases. So having some way that you can couple these things in 1 neural network would be very cool. But, yeah, that's a tricky technical problem. I'm not too sure about how to go about.

Nathan Labenz (1:43:57) Cool. There's always more to do, least for now. Yeah. Until the AIs are doing it. For a few years, at least, there's

Tim Duignan (1:44:02) Yeah. There's always

Nathan Labenz (1:44:03) more to do.

Tim Duignan (1:44:04) I mean, this is 1 other point I would I would make is that, you know, I really encourage you know, there seems to be a real excitement and attention around making an AGI or artificial superintelligence. But, you know, I just encourage people that there are so many important, incredibly exciting problems you could tackle with machine learning. You know? And I I think this this goal of kinda creating something that's smarter than any human is maybe something that, you know, it comes along with all these kind of risks and uncertainties. And I know it's an appealing idea, but it's maybe not the kind of thing that society actually wants us to do. You know? So I encourage, you know, machine learning researchers to think, is this actually a goal that that that we should be pursuing? So you know? Because they're throwing billions and billions of dollars at at the moment. Right? But I think if you go and ask the average person, like, do you want something that's smarter than a human in every single way? They'll probably say no. Right? They'd say no. I want something that, you know, can do my dishes and do my chores, but not something that can make art and music better than I can. Right? So and similarly for scientists, you know, I don't want an AI bot that can come along and do hypothesis generation and and come up with new ideas and read the literature. But, like, those are the things I love doing, you know, and solving problems. You know? That's what I like doing. So what I would kind of prefer is people to build tools that I can use to do things like massively accelerate simulations, for instance, to take the kind of really hard problems that we're stuck on in science and give us new powerful tools that we could then use to solve problems. So that's if I there's 1 thing message I could give to that community, it would be kind of just try and find tools to solve problems for humans to do these things. Because building something that's smarter than any human is is not really something that that I think any of us really wanna do. And so, you know, like geneticists could be, you know, using CRISPR and genetic editing to make the humans that are smarter than anyone, but not something that they're trying to do because it's not something that society has decided that we want to do. You know? So I think it's we want us to take a step back maybe and sometimes think about at a more fundamental level, what are we trying to do and why are we trying to do it. You know? So Sam Altman talks about this idea of, you know, wants to build an AGI that can discover the grand unified theory of physics, you know, unifying gravity and quantum mechanics and then explain it to them. But, like, I think that's something we should save for a human. Right? And we should build tools that help us do it, but I don't think we just wanna have an AI explain it to us. Right? That feels to me like just looking in the back of the book and just reading the answers. Right? I mean, there's a a source of meaning for for so many people, and I think it's maybe something that we should think about trying to preserve. And I understand that, you know, an AGI would be incredibly practically useful. You know, it allows to help solve cancer and things. But I think there's ways you could use AI to help solve, you know, these problems by developing really powerful tools rather than just by developing something that tells us all the answers.

Nathan Labenz (1:46:34) Yeah. Yeah. Amen to that. I'm very largely with you. I think there's an interesting distinction there too between, is it the power or is it the generality that should concern us more? I tend to fall more on the fear of generality side, and it's an odd feature of the current reality that like the path to a lot of the best performance on even narrow tasks is like through pretty generalist systems. But I do share your concern that first of all, that I'm, like, not sure society wants an AGI or certainly a superintelligence. Obviously, I definitely do think the risks of that are not to be dismissed, and it it does feel to me like a kind of fundamentally ideological project. It's not necessarily framed that way, but I do think that it, you know, it it probably should be understood that way. But when I I think the only point of divergence, I think, from what I understood you saying just now is if I could have a narrow system that could do some of these, like, frontier things but couldn't do anything else, you know, sort of the alpha go of, like, understanding physics, I think I would be on board for that. Even if that meant that the human who counterfactually would have made that discovery, like, gets deprived of that you know, eureka moment, the joy of that eureka moment. I still think I would probably be inclined to make those trades, but the the part for me where it becomes, like, very concerning is when that physics solver is also, like, a general purpose intelligence that can do anything better than a human, and then we have all these sort of control problems along the lines of the research coming out of anthropic. You have these sort of they generalize in pretty nasty ways, it seems like too, where they they just showed 1 where simple flattery or sycophancy toward the user starts to generalize toward, like, much more problematic forms of deception. And it's like, yikes. I think the path through this, like, superhuman intelligence to a good outcome is definitely fraught to say the least. If we could find a a path to a narrow physics solver or material science discoverer, that was better than humans at only that narrow task, I think I would be on board with that. But, unfortunately, right now, like, that path is not not super obvious.

Tim Duignan (1:49:03) Yeah. No. Fair enough. I I understand if we could build 1 of these things, yeah, it would probably inevitably get built. I just think I don't know if we wanna throw that many resources at it. And it's a question of where we direct our resources to. You're building a narrow AI physics solver. Yeah. It's easier, then sure. Yeah. Then doing it you know, building tools for humans to use, then, yeah, maybe that is the way to go. And, yeah, the other, yeah, massive obvious counter argument is obviously the only way to get something that's really useful is to build this generality into it. So maybe there's no way to build things that are narrowly really good at individual tasks, and the only way to really go is to build this kind of general intelligence. Yeah. Which may be just an unfortunate fact of the way reality works. But, yeah, it will be interesting to see how it plays out. It's gonna be a fascinating time.

Nathan Labenz (1:49:44) Yeah. No doubt about that. Alright. I'll say it again. Thank you for getting up early in the morning, and thank you, Tim Duignan, for being part of the Cognitive Revolution.

Tim Duignan (1:49:53) No problem. Thanks for having me. It's a pleasure.

Nathan Labenz (1:49:55) It is both energizing enlightening to hear why people listen and learn what they value about the show. So please don't hesitate to reach out via email at tcr@turpentine.co, or you can DM me on the

Nathan Labenz (1:50:08) social media platform of your choice.