Why There Is No “AlphaFold for Materials” — AI for Materials Discovery with Heather Kulik

is super good at Wikipedia level chemistry knowledge. I'm really interested in molecular design, like how do you find a new ligand that can go into a transition metal complex. And what that means is that some combination of atoms and it's gonna bind to the metal and it's gonna change its properties.

The thing I constantly do every time an LLM is updated is I just ask it, please design me a ligand that has 22 atoms. I can never get an answer that has 22 atoms.

Yeah. So, my my group, we work a lot in accelerated discovery of new materials. When I first started out, we were just really using AI to make predictions we'd normally make with computational models, just make them faster.

But the question I would often get when we're doing that was, okay, but what's what's surprising? What's what's sort of something from AI that, like, I wouldn't have already known if I were a really smart chemist or a really smart material scientist? And, you know, you make all these computational predictions.

Has anyone actually made in the lab something that you predicted? Recently, I was I was able to do a really nice demonstration where the answer to both of those questions, you know, was very clear from the work. So we were able to screen with artificial intelligence a set of thousands tens of thousands of materials where each individual experiment, if it were done in the lab, would have taken months to years.

And through AI, we uncovered this sort of unexpected chemical phenomenon that led to a emergent property in in what's known as a polymer network, so plastics, that would make the polymer about four times tougher. And when we showed, the design that AI had come up with to the experimentalists, they were really surprised. They would have never come on this on their own.

And then we were able to convince them to make it in the lab, and in fact, it was it was this tougher material. And where this has applications is if we can make plastics, tougher, then we, you know, can get more use out of them, and it'll ultimately address some of the problems we have with overall durability and use of plastics. So I think that's that's an example of some of the promise of AI and materials discovery.

sort of hard for me to think about how to how to explain it without getting too deep into the chemistry. But basically, these are molecules that have to break apart, and when they break apart, they make the overall structure that they're in tougher. So a little part of the material breaks, and that helps to dissipate the force.

Normally, the way you would think about making it easier to break apart these small molecular components might be to create a hinge so they can kind of peel open instead of sliding apart. But what we discovered was that there was a fully quantum mechanical phenomenon. There was really no way for us to predict this, you know, based on anything else where the electrons just move around in a different way so that at this moment where the molecule's gonna break apart, it's a lot more stabilized.

These types of concepts, they're sort of similar to what's kinda known about how catalysts and enzymes work, but it had never before been shown in in these polymer materials.

structural integrity during a earthquake by having a controlled break? Is that kinda Yeah. Yeah.

So we weren't the first ones to discover that phenomenon on its own. The general phenomenon that putting little places that could break to make the network stronger, that was published in Science Magazine a couple years ago. But the specific way we came up with to design the material to do this, that was that was our new contribution.

So, you know, I I was drawn to data driven discovery pretty early on, sort of before I even knew the phrase machine learning. And I guess I was just really excited by what you could learn from patterns and data. Back then, we were trying to call it cheminformatics, and just sort of trying to think about, you know, in what ways could you unearth trends in data, because I I started my career actually working kind of one molecule at a time or one material at a time, and I was just impatient.

I wanted to be able to sort of understand not just one molecule at a time and write one paper about it, which is something people would have been happy to do back when I was starting my career in the mid two thousands, but to actually kind of unearth broader trends in in how you understand how material is gonna behave. Somewhere around 2015, 2016, I realized it was a bad idea to call things cheminformatics, and it was a good idea to start calling things mushy and learning. And I had a brilliant student, John Paul Janae, who's now, I think, assistant director at AstraZeneca in Sweden, running their, inverse design program.

He and I originally talked about all sorts of ways of thinking about materials design, and he very quickly adapted that into training neural networks. And that's sort of when you know, I thought we were in the first sort of hype cycle, the first wave, but I think compared to what's going on right now, it was a tiny baby wave.

Yeah. Yeah. So even that polymer example I was giving, that would have been active learning in principle, but we sort of stopped after one generation because, because we had exhausted the space.

But, I think one of the areas where machine learning kind of just with what's out there right now has the most promising chemical sciences is in solving multidimensional challenges. So right now, we're working on a project in metal organic frameworks where we're trying to solve trade offs, relevant for, direct capture of, c o two from the air. And so in order to find a material that's good for that, we would worry about its cost, its stability in, say, aqueous humid environments, its ability to take in c o two over other molecules, its mechanical stability.

Is it gonna hold up under force? Is it thermally stable? Can you heat it up, and will it be okay?

I'm just naming a few. But in total, right now in an active learning campaign, we're working on seven different objectives. And usually, just even for a not so accurate machine learning model, you get, you know, at least a 100 to a thousand fold speed up for every dimension you're optimizing over.

So the real promise is gonna be in searching for that needle in a haystack with, say, seven objectives and and doing something where you're not waiting for the models to be accurate before you start doing that optimization. That's really the promise of active learning. Yeah.

getting it the drug out to in people's hands is often the bottleneck for a drug. And, also, you know, what happens to the drug when it sits on the shelf for three months? Yes.

they're they're used for for? They're used most in gas storage, sensing, and separations. They're used in combination with polymer composites.

They have a really, strong promise for c o two capture especially, but people have looked at them for catalysis. The limitation on catalysis has been, you know, how stable are they. So one of the things we spend a lot of time on is trying to be able to predict their stability, but they're used for all sorts of things, even drug delivery.

You know, what they have the opportunity to do is really place precise chemical groups in specific orientations that can ultimately allow for what's known as host guest interaction. So basically, kinda create a glove to have a targeted interaction with a with a guest molecule in the metal organic framework. I see.

And just for the for the nonchemist, metal organic framework, LEGOs for for for chemistry. Is that Yeah. Yeah.

Metal organic frameworks, I think, are going to be a little bit more of a household name among some engineers because they, the discoverers of those materials just won the Nobel Prize in chemistry this year. So as as much as that can make something in chemistry a household name, but they're basically, like Tinkertoys or LEGOs, and they have different building blocks that can be combined in basically infinite ways to create very precise chemistry.

advance those? Like, what are the roles of the two? So how?

So I I started my career studying what's known as transition metal catalysis. If you look at the periodic table, the middle of it contains a bunch of metals, a good example would be iron. And all of those things sitting in the middle of the periodic table, they have, what's what's referred to as an open shell.

So the electrons in those those materials are not paired, and they're not they they're, as a result, more reactive. Normally, like, the the way that you understand how they're going to behave, so that, for instance, they give rise to, you know, different combinations of these metals give rise to the catalysts that are used in a large number of transformations, including the things that, say, feed and sustain most of the world's population, such as the Haber Bosch process for ammonia synthesis. And the way going back twenty, thirty, fifty years that people understood these materials and could enable their rational design is through quantum mechanical modeling.

Quantum mechanical modeling by using approximations to the Schrodinger equation, it can be very accurate, but it's very computationally costly. And so a single quantum mechanical prediction, depending on the level of fidelity used, could take hours to days to weeks. And that's what I would have normally been doing before I got started in AI.

Some of what we do these days is accelerating those quantum mechanical predictions as well as looking at you know, an area that I'm particularly excited about is that not all quantum mechanical approximations are equal, and you can actually use ML models to kinda predict what the best approximation to use is depending on the material studied.

In terms of which method is the right method to use? Yeah. So we it it actually turns out to be quite complex.

You can't just determine it from heuristics. So we actually, in one area, use the quantum mechanical wave function as inputs to neural networks to actually predict what is the right method to use and learn that mapping. I see.

That's probably gonna be an interesting question. Challenge. Yeah.

The benchmark.

My personal experience is that, and this will date itself immediately, is is that is that CHA GPT is super good at Wikipedia level chemistry knowledge. But one of my favorite things to actually throw at GPT as as an anecdote is I'm really interested in molecular design. Like, how do you find a new ligand, that can go into a transition metal complex?

And what that means is that some combination of atoms, and it's gonna bind to the metal, and it's gonna change its properties. And so the thing I constantly do every time an LLM is updated is I just ask it, please design me a ligand that has, 22 atoms. So the first time I've I've done that, there are many ligands out there that have 22 atoms, and then I say, wanted to bind to the metal with two nitrogen atoms.

I can never get an answer that has 22 atoms. So then you can try a range and see how many times you can get a range. And so that's that's maybe a trivial thing, but that's something that a expert chemist, you know, could do in a do in a second.

So there are really good introductions to chemistry that I think you can get through conversations with an LLM. You can get a lot of insight into an area you're unfamiliar with. And for sure, things have improved a lot.

Like, when I first tried typing in, you know, which exchange correlation functional should I use for this type of chemistry? The answers were completely wrong. They looked right, but they were completely wrong.

I think things have gotten better because that knowledge is out there on the Internet. It's in the training data. But I think there's a lot of things that probably backing up a moment.

You should learn chemistry well enough to know when when these models are right or wrong. And if you don't know any chemistry at all, it's hard to know if you're if you're assessing correctly. But I think that there are a lot of things that you don't have time to do a deep dive into that you can now get from, say, an LLM that can augment knowledge.

But I think you have to start from somewhere and then use it as a tool rather than starting from zero and relying blindly on what an LLM will say.

challenges out there where the datasets aren't large enough or diverse enough, and so I think they've attracted less interest. So the ones closest to my heart are, reactivity predictions, so predicting which reactions will occur and and why, especially in complex, phenomena, like in, you know, multiple elements and and sort of predicting those transformations. Another thing that I think there's not enough data on is just more diverse chemical bonding and more diverse chemistry.

For me, that's transition metals, but there's also questions of warm dense materials, sort of exotic phenomenon. We have really good data sets out there for really boring chemistry. So we have, you know, probably even if you're not a chemist, you're familiar with, organic molecule data sets and organic molecules binding to proteins.

Those are the common data sets out there. There's lots of challenges out there where the physics is much more complex, and the things like how does matter behave when, you shine light on it and you excite it into excited states, all all sorts of things like that receive relatively little attention because, you know, there may not be a benchmark or a leaderboard yet for that.

there are all sorts of repositories of fairly low fidelity DFT data on crystalline materials. So materials project, open catalyst project, these do provide good leaderboards, but some of the limitations of that are the data comes from not very high fidelity density functional theory. So I'd say that's a second, challenge is that we're the smartest ML engineers right now are learning on data that is not going to be reflective of experiment.

There aren't big experimental data sets. For example, one of the advantages of things like CASP is that it comes from an experimental ground truth, whereas that aspect just isn't available in materials as much.

What would you want it to do? One kind of murky area so maybe I'm not gonna directly answer answer this question. One murky area for us is electronic structure calculations are expensive, and, they should, in principle, give you the right answer.

They should, from first principles, give you the right answer of how a material's gonna behave. And a lot of people are scaling these up right now with, machine learned interatomic potentials on training data. And every time someone comes out with kind of a new data set trained on a and they call it a foundation potential, foundation model, it looks really good.

And then you get it into your lab, and you say, okay, I wanna use it for this problem I'm really excited about. And it starts doing kind of wacky things, like molecules fall apart. I won't name names, but there was one that made a huge splash this summer, and people started declaring, oh, this method is dead.

This method is dead. We're all gonna just use these neural network models now. It's only in my hands, the one I'm still not naming, is only about five times faster than my fastest DFT calculation on a GPU, and it also doesn't work all the time.

So I would say we need a more transparent way of of trying to figure out, if these models can really replace conventional physics based modeling. If they could, if I could just give up ever doing a DFT calculation again and just rely on machine learning potentials, and if they were, you know, two orders of magnitude faster than the traditional approach, that would change that would change how we're doing science.

Yeah. So so there are a lot of really smart people working in high throughput synthesis and experimentation and autonomous labs. I think the thing that that's interesting to me in that space, at least, is that there are some types of experiments that, at least as of the last conference I went to on this, are really hard for autonomous, high throughput experimentation but are really easy for a human and vice versa.

And then there's, you know, the serendipity that a human might experience in the lab that a couple of people have tried to think about, like, well, how do you how do you introduce that noise into high throughput experimentation? So I think that's a challenge. Your question also brought to mind another point that I am by no means an expert on, but most people who actually work on getting materials to the device scale, say something that would be in your television or something like that, is they will tell you that it's not just the material, it's the process.

And I think we're at we're at ground zero. We're we're nowhere when it comes to, like, well, how do we machine learn not just the structure and the properties, but also the the role that processing plays?

and the example I was giving is something we know is stable, and we've seen a structure of it before, and it will fall apart with some of these models. The challenge here is that, what AlphaFold has done really well is is predict structures of globular proteins primarily with 20 natural amino acids. I could actually point to lots of cases where alpha fold fails too for more interesting chemistry.

The challenge is that you have a lot more than 20 building blocks when it comes to materials, And so there's lots of different ways to think about chemical bonding. And right now, no potentials are really robustly encoding all of that bonding, especially with respect to metal organic bonding.

Now there's simple regions of material space. You can pick aluminum. Aluminum is very boring, and you can write down people in the sixties could write down on on paper, you know, how you need to model aluminum.

That's something that is pretty easy to fit a neural network potential to. But then if you wanna get over to iron oxide, and then if you wanna get over to high entropy alloys, there are definitely cases where people are using these methods. But I'd say a big challenge is is that there's no real way to know if when you go to bigger length scales and time scales, there's no real way to know if you're right or wrong.

The experimental data is not there. Experiment even interpreting, say, looking at an image of an experiment surface, which you would wanna do, it requires some degree of an interpretation of that image. So it's just it's just hard to know from experiment or from other computations if these types of models are are correct, and they're certainly not correct across all of chemical space.

And I'd say they could fail more catastrophically than AlphaFold obviously fails, though there are definitely failures of AlphaFold too. Switching gears a little bit.

Yeah. So we started, I guess, about five years ago. So when we first started doing it, we were just doing sort of standard natural language processing and graph digitization.

These days, we use LLMs. But just to try to extract from the literature datasets of properties, wherever wherever people are widely reporting properties. And what we noticed is that there's a lot you can learn from these models.

So you can you can even on the scale of a few thousand data points, you can then do things like predict the temperature at which a moth will break apart based on experimental reports. But one of the funniest things I think we noticed is that you can get the temperature at which a material will break down two ways. One, you can get it from the graph, and two, you can get it from what the authors say about how they interpret the graph.

And those two things do not line up. So people you know, one of the challenges I think with literature extraction from papers is one would be the obvious mistakes people make, you know, no one's perfect. But the other would be just, you know, people interpret their results in different ways.

And so if we're building models based on those interpretations, that's a challenge. In terms of LLMs, they've come a long way in terms of literature extraction, but they're still definitely sensitive to false positives. And I think the amount of time we spend checking on LLMs to make sure that the data we're ingesting is is accurate definitely is an overhead on those types of workflows.

I see.

then maybe it's biasing you towards the previously reported results instead of something know? You know, one of the ways we try to address that is we try to train a model on that literature, but then apply it to new structures that have never been seen before and try to really look at how far we can extend the model. But we are trying to answer this in general.

There are repositories out there of experimental data where you can have a sense of when it was published, what the structure is, what it was used for, and we're really trying to build generative models on top of that now to try to be able to say, well well, if I know about the first thirty years of a field, can a model trained on that predict the next twenty? I think that's an that's an open question. And what model is best?

Know, and maybe it won't get all of them, but maybe some of those discoveries that we think are new in the most recent twenty years, maybe maybe some of them are trivial for a model to generalize to, whereas others are are not. I think in an ideal case where where we have the available literature data and we don't know, we could use uncertainty quantification to then identify, okay. These would be the most interesting materials to get into our dataset.

as well as metal organic framework activation stability, water stability. Other groups have curated other measures of stability. They're all out there.

They're on our website, that kind of thing. Awesome.

like, which really will drive the field forward in your mind? I think the National Science Foundation has one initiative. I've also heard about things with with, foundations before, sort of being interested in in putting together Cloudlabs, so things that users can on demand make use of high throughput automation.

I definitely think I think having user facilities where a computational researcher like me could design an experiment and have it executed would be awesome. Having all that data collected in sort of a public way would be great. You know, the way that research right now gets published into papers, it's very hard to then extract back out.

We spend a lot of energy trying to get it back out. And so some of this is a need also for, you know, maybe systematization of how results get reported so that, they can be machine learning ready from from day one when they're published. Some research subfields are trying to do that, but it's it's not really developed across material science.

But for sure, you know, I think there will be more sort of shared facilities where people can, make use of data from high throughput experimentation, and that would be really, really awesome. I don't know if it'll come from companies donating equipment from National Science Foundation or from, you know, private foundations.

I ask myself that all the time or more recently in the past year. So in particular, there's a lot of compute that companies have access to that academics don't. So I ask myself, you know, what can we do that's more creative that doesn't require just brute force compute?

And I think there is there is, like, an a lot of stuff that we can still do, but we have to ask those questions. For sure, Microsoft, Meta, those those ones are are kinda like the companies that have basically infinite resources. And as an academic, I don't have infinite resources, you know, but we have an interest in problems that, you know, haven't crossed the radar of those companies yet.

And I think as long as we you know, whenever someone poses a problem to me now versus a few years ago, I try to make sure that we're not just in the process of trying to do something that you're throwing a lot of compute at it would would would solve it. Yeah.

So I I think there is still a place for chemistry. I will say that. But my group develops a code for transition metal complex structure generation metal organic framework screening.

It's called MOLSimplifier. When we're working on MOFs, we call it MOF Simplify. There's website versions of it that you can look up and not install anything, but it's also on Conda and GitHub.

And if you do have an interest in transition malcomplexes, you know, just try it out. It includes machine learning predictions, but it also make novel structures. And I'm just really interested to hear ever if people are using it.

I know a lot of companies are using it, but we sort of find out sort of after the fact. So, if you're interested more in this material space, I'm I'm definitely interested and open to feedback.