This episode is sponsored by Toyota Research Institutes. Check out their homepage here. Take a look at jobs here This episode is sponsored by Materials Today, an Elsevier community dedicated to the creation and sharing of materials science knowledge and experience through their peer-reviewed journals, academic conferences, educational webinars, and more. Thanks to Kolobyte and Alphabot for letting us use their music in the show! If you have questions or feedback please send us emails at [email protected] or connect with us on social media: Instagram, Twitter. Materialism Team: Taylor Sparks (co-creator,co-host), Andrew Falkowski (co-creator,co-host) Jared Duffy (production, marketing, and editing), Ramsey Issa(editing assistance).
Keywords: Toyota Car Catalysis Catalytic Converter Smog Exhaust Platinum Nitrogen Oxide Rhodium
I would like to describe a field in which little has been done, but in which an enormous
amount can be done.
This field is not quite the same as the others in that it will tell us little of fundamental
physics, but it will tell us much about the strange phenomena that occur just below our
perception.
In contrast to the natural philosophers of the past, the scientists of this field delve
into the recesses of nature and show how she works in her hiding places.
Their quest is to understand and create the imperceptible.
After all, there is plenty of room at the bottom.
Hello and welcome to the Materialism Podcast, an exploration of the past, present, and future
of material science and engineering.
My name is Taylor Sparks, and I am joined by Andrew, and today we're also joined by
Joey Montoya of Toyota Research Institute.
Joey, how are you, man?
We're doing great, thanks.
We're so excited you're here.
As you guys know, the time zone difference means that Joey had to wake up super early
for this, so we appreciate it, getting you out of bed.
Hope you're excited to talk about catalysis first thing in the morning.
Oh yeah, catalysis first thing in the morning, always a fun time.
So I was just telling Joey that catalysis is an enormously important field, but it's
one that I feel like I know very little about.
So I'm going to be playing a little bit, I don't know, not quite devil's advocate, but
I'm going to be asking some really simple questions, so I hope you put up with getting
the basics in place, but we're happy to learn everything you got to say about it.
Andrew, how have you been?
Doing well.
It's pretty cold here now.
We have snow in the mountains, but the colors haven't quite left our trees, so it's pretty,
it's very pretty.
Yeah, we were recording this the day after Halloween, and my kids were just delighted
to learn that the UK does do Halloween.
There were trick-or-treaters.
They did get candy.
It's weird candy, like what do we got here, we've got Curly Whirlies, which I stole from
their baskets, which are nothing like anything I've had before, but not too shabby either.
And I was able to head out last month to the, you know, we've been talking about the seventh
international conference on multifunctional hybrid nanomaterials.
That's the conference put together by Elsevier that was in Genoa, Italy, and I was able to
go and check that out last month.
It was super cool, so props to them for putting on a really cool conference in an amazing
location.
Okay, for today's episode, we're going to dive into catalysis.
Now that is a massive topic, which covers lots of technologies, which matter to all
of us.
So to get started with, Joey, do you want to tell us a little bit about yourself and
why you are in this field?
Sure.
So I started studying catalysts in undergrad at University of South Carolina.
That was a long time ago now, but I got a PhD in chemical engineering from Stanford
University and my specialty is in theoretical catalysis.
My thesis, I basically explained how catalysts are limited in their efficacy to convert water,
carbon dioxide, and air into useful commodities like hydrogen and natural gas and ammonia.
And I was also a postdoc with the Materials Project for three years.
I spent a lot of time developing open source software for things like DFT automation, data
mining, visualization, all topics that both you and Andrew are intimately familiar with
and I've heard previous episodes on.
I started at TRI, Toyota Research Institute, in 2018.
I've been working on similar projects there.
These days, my primary interest is in translating predictions, both from machine learning simulation
into the real world and specifically for catalysts.
I'm especially interested in the making part because I think there's still a pretty big
gap between DFT, machine learning prediction, and actually making a material or catalyst
in the laboratory.
Super cool.
Can you tell us about, so you're at the Toyota Research Institute.
For people that have never heard of that, that thought Toyota just made cars, what is
the Toyota Research Institute or TRI?
TRI is a research company.
It was founded by Toyota a few years ago to do deep, long-term research in the fields
of robotics, autonomous driving, material science, human-centered AI, which actually
includes cognitive science.
TRI's mission is to create new tools and capabilities, which are focused on improving
the human condition through research in those fields.
For our energy and materials team, which is the one that I'm on, our more specific objective
is to create new ways for material scientists to develop emissions-free tech.
There's a subtlety there in that our primary objective is to create tools rather than just
creating materials, but we anticipate discovering and already have discovered some materials
in order to prove that the methods we're developing work, so we do that kind of research too.
We typically focus on materials that are relevant to vehicle electrification, so things like
batteries, fuel cells, polymers, but because we're developing tools, we have a more general
focus on materials informatics as well.
That's rad.
We've had a couple companies on here that are big-name brands that also do research,
but it's not like it's a common thing, so it's awesome to see a big company like Toyota
making that investment, looking in the future and not just looking at short-term, what can
we do tomorrow to boost our sales for this quarter, but investing on five-year, 10-year
down the road technologies.
That's awesome.
Yeah.
Automotive companies have to do this.
They have these long timelines for research and product development, and so making these
types of investments is generally seen as a good idea.
Super rad.
Yeah.
It makes a lot of sense too.
Most materials research is going to be long duration and very expensive, but the potential
benefits are huge, and most of the time, that's locked to academia, but oftentimes, academia
just doesn't have the financial resources to take those long-term investments or do
that sort of development.
Yeah.
Joey, you're here to talk to us in this episode about catalysis.
Give us a very broad overview of why catalysis is important.
Is it something that's going to sort of change the world, or is this just something that
is more niche?
Tell us about it.
Well, I would argue it already has in many ways, and historically, I think some of the
historical examples are interesting.
First of all, catalysts, for those of you who don't know, are materials or molecules
that cause a chemical reaction to proceed more quickly without being consumed.
Frequently, the reaction would have happened anyway, but maybe over hundreds of years instead
of over seconds or over years instead of seconds.
They're generally understood to do this by lowering the activation energy of the chemical
reaction.
Some of you in high school chemistry probably studied this and learned about the kind of
hill that you have to go over, and it basically splits the hill in two and makes both the
smaller hills lower.
Catalysts basically give you an alternative route to the main route in the thermodynamic
space.
That means you don't have to go to the top of the hill and you can get where you're going
faster.
In terms of importance, historically, one of the most important things that catalysts
have done is they've basically enabled us to produce food at a much higher rate than
previously because of the additives we put in fertilizer that make growing food much
easier.
The Haber-Bosch process takes nitrogen from the air and hydrogen, typically now from steam
for methane, but hydrogen and creates ammonia, which we then use to make reactive nitrates
that we put into a fertilizer.
There's lots of people who have speculated that this is basically the reason that the
world's population was able to grow throughout the 20th century.
Yeah, gosh, it feels like we could do a whole episode.
Maybe we'll have to do a future one just on Fritz Haber because what an interesting guy
who is like equal parts like villain and hero.
If it wasn't for him, like you're saying, the food that we have today would not be possible.
I just saw a stat, I was Googling this, but something like 40% of our food is now reliant
on the ammonia which comes from his process, which is absolutely amazing.
As of a couple of years ago, they were producing 230 million tons of ammonia, right?
Using this process annually, which is just, it's staggering.
The amount of CO2 emissions and energy tied up with this is absolutely incredible.
But this same guy, apart from feeding us, was also involved in the development of mustard
gas and he was on the wrong side of World War I.
He's this complicated, interesting individual as a lot of scientists have been over the
last hundred years.
Yeah.
Yeah, yeah, absolutely.
Yeah.
I like to put it that one out of every two nitrogen atoms in your body probably was on
the surface of an ammonia synthesis catalyst, which is usually iron.
Yeah, so you can think of it that way, that the molecules in our bodies are intimately
connected to catalysis and the catalysis industry.
Okay.
So, it's food.
And I've definitely seen those stats where people say like, hey, we're going to run out
of food, it can't support over XX population, right?
And then it just keeps going higher.
We keep finding ways to do it and it wouldn't be possible if it wasn't for catalysis.
Why does Toyota care about this?
Yeah.
I mean, Toyota is not a food manufacturer, but catalysts are important for Toyota for
a couple of different reasons.
So almost everybody, and when I talk to people about catalysts, one of the first things they
always ask me about is the catalytic converter in your car.
And the catalytic converter is responsible basically for taking the emissions from your
engine and converting them into something that's a little bit more benign.
So things like sulfur oxide and nitrogen oxides that result from incomplete combustion of
the fuel in your engine and hydrocarbons too, and CO, they basically oxidize all of
these and form nitrogen and CO2, which we know is not great from a climate change perspective,
but is much more benign than CO or the remaining hydrocarbons, which can be carcinogenic and
basically contributed to smog.
So we don't really see smog in California as much as we used to.
And part of the reason is because catalytic converters are both required and widespread.
So that's one reason.
I was just going to say, if you look at pictures of LA from like the 70s, I don't know, maybe
it was the 60s, it's like this orange haze, right?
Because these were not fully catalyzed reactants ending up in the atmosphere that produced
that really very dangerous stuff.
And I read an article once that talked about the catalytic converters having done more
to improve the environment, broadly speaking, in terms of broad impacts than any other technology,
which is absolutely incredible.
And we still put them today, but going forward, let's imagine a future where there are fewer
internal combustion engines.
Why do we still care about catalysis?
Yeah.
And we have to stop emitting carbon, quite frankly.
Our CEO wrote a great article about this, and it's basically talking about how carbon
is the enemy.
And so one of the big things that we have to figure out how to do is just how to power
cars without emitting carbon.
And there are several technologies that we're kind of thinking about for this.
Two of the big ones, battery electric vehicles, fuel cell electric vehicles.
The fuel cell electric vehicle in particular is one that depends on catalysts.
And so most fuel cells, especially the PEM, fuel-sized polymer electrolyte membrane fuel
cells, they basically, they use catalysts and the catalyst layer to catalyze the oxidation
of hydrogen into water and in the reduction of oxygen also into water.
And basically, these chemical reactions go very slowly, and they're what I would call
the rate limiting step of the fuel cell's performance.
They basically limit the power density, they limit the efficiency, because with a catalyst
like platinum, the fuel cell can operate at a certain efficiency and a certain power.
If you were to replace it with something like nickel or something like gold, for example,
it would just not operate nearly as efficiently.
So the efficiency of this device depends directly on the catalyst that's used.
Okay.
Now, it seems kind of unique that Toyota is doing this for both electric vehicles and
fuel cell based vehicles, right?
So these aren't just batteries, these are fuel cells.
I feel like most companies have done one or the other.
Why is Toyota taking the strategy of developing both of these technologies at the same time?
So I would argue that the simple answer is that we just don't know what's going to happen
in the future.
There are lots and lots of supply chain constraints around both of these technologies.
Batteries, battery electric vehicles and fuel cell electric vehicles, they both have the
advantage of being able to use the electric motor, right?
In some ways, it's a disadvantage, actually, because magnetic materials themselves that
go into the electric motors, those are probably going to start to come in some sort of supply
issues in the coming years, just because we're kind of depending on them for everything.
But there's questions about lithium, there's questions about cobalts in the lithium ion
batteries and the cathodes that we don't know how to answer yet.
And manufacturing of both of these things is going to get better over time, but we don't
really know which one is going to become more viable.
It might be the case that we don't actually achieve widespread adoption of lithium.
What I mean by widespread adoption is like one technology comes to dominate the whole
market, right?
It could be both.
We need both.
A great example of that, actually, is heavy-duty trucking.
So heavy-duty trucking, one of the biggest issues is weight, right?
And all the battery EV trucks that have been proposed are very, very heavy.
And so hydrogen actually makes a lot of sense in this context, both because you don't have
to worry about the weight, the same type of weight restrictions, which for trucks is
very important.
Anybody who runs a truck and drives a truck route, you see weigh stations on the highway
and trucks also cause a lot of wear and tear on roads.
So hydrogen might make more sense for that particular sector.
It might, you know, it's something that there's actually a Mirai project, a hydrogen fuel
cell project that's been showing how these trucks can be constructed relatively simply.
And also distribution for trucking might be a little bit easier than creating a new
distribution network for hydrogen because you could probably just use it on highways.
Oh, yeah.
Yeah, just the main arteries.
Exactly, yeah.
So and long story short, it's simply in the Toyota.
Toyota is a very, I would call it, careful company.
We like to make sure that we do things right.
We have a culture, basically, of high attention to safety, high attention to reliability.
People who own Toyotas typically when they switch, they're like, oh, yeah, yeah, yeah.
They're like, they're like, oh, reliability is something that I got used to and that sort
of thing.
So and yeah, it's something that they want to be able to fulfill.
And so they're obviously invested in a lot of different future facing technologies here.
I'm going to put in the show notes an article I read in getting ready for this episode that
touches on some of the things that you just described.
It's related to the supply chain.
I just don't think people realize the embodied emissions that go into an electric vehicle,
how massive they are because of the mining.
It's staggering.
The amount of like the tons of earth is like hundreds of tons that have to be removed to
make a single battery, which is wild.
And so compare that with an internal combustion engine where most of the emissions come during
its operation.
In an electric vehicle, it's mostly during fabrication, and that's mostly the battery.
So fuel cells do offer some interesting advantages there because they don't like, for example,
if you want a truck that can go a further distance, which me living in Utah, where everything's
spread out, mileage anxiety is a real thing for me.
If I can't go through 400 miles, I may not make it to the next gas station sort of thing.
But that means that the battery has to scale with it, and that means your emissions scale
with it, and it becomes maybe not practical.
But a fuel cell is a different animal, right?
You can store hydrogen, and the same motor can store hydrogen at a very low weight basically,
right?
So it really eases the range anxiety, but it has its own challenges, right?
Andrew, do you have something you want to add to this?
Yeah.
I mean, the Scandinavian government, I think, recently commissioned a study where they're
actually looking at what happens when we need to actually start replacing the batteries
in the current fleet of vehicles or start making the next generation that comes after
this current generation of electric vehicles.
And there really just isn't that much lithium and cobalt out there to manufacture these
massive batteries.
So even if it is feasible for maybe the next 10, 15 years, there's going to be challenges
that have to be overcome in terms of the materials that comprise these batteries and
the overall structure and the sort of recyclability of these materials as well.
I know there are some fuel cell vehicles on the market right now from Toyota itself.
So what exactly are the challenges that the Toyota Research Institute is looking into
right now as far as catalysts go, right?
I know typically you're looking for high activity, low cost, and long-term stability.
But how does that actually manifest in terms of research?
Yeah.
This is a great question.
And so we recently wrote a review of our last three or four years of research on the
oxygen reduction reaction specifically, which is probably the most critical reaction in
fuel cells for the catalysis.
And I highly recommend it.
It's an energy environmental science that came out about a year ago or a little less
than that now.
So we actually have a really what I would call a healthy understanding of why some catalysts
are active and some aren't.
Part of what I studied in grad school was basically this.
We can use simulation actually to basically determine the relationship between the electronic
structure of catalysts, specifically metal catalysts, and their activity.
And there's some beautiful work that's been done.
Basically, one of the things that's really interesting is all the bond energies between
molecules that go through the catalytic cycle, the bond energies between those adsorbates
on the surface, they're all correlated.
And what this means is basically there can be fundamental limits on how good a catalyst
can be.
And platinum is kind of at the top.
We call it a volcano.
There's basically a volcano relationship, and there's a peak of the volcano, and that's
the best sometimes catalysts can be.
And this also helps us to think about designing new systems to maybe circumvent those what
we call the scaling relations.
But in terms of research, so we're thinking about that.
We're thinking about how to make the activity better and the power, which ultimately means
the power density can be better.
But the point we make in this article actually towards the end, because we kind of shape
it as a perspective, is we can actually make the Mirai at reasonable customer costs now.
We can sell them, and it's not too bad.
The amount of platinum we use is reasonable.
There is kind of a floor on how cheap we can make it because it's got platinum, because
the manufacturing of the fuel cells got some floor costs.
But we think it's kind of ready to give to consumers.
There's still some big problems to solve for hydrogen, right?
I'm not saying that fuel cells are going to take over or anything like that.
But yeah, we can make the whole thing pretty well.
But we kind of want to make it a compelling technology, something that people get excited
about.
And for that, I think we have to think a little bit more about durability.
Why catalysts degrade?
And also, if we want to make a cheaper catalyst, we have to solve some durability problems.
So alkaline fuel cells, for example, you can use different kinds of catalysts, things
like nickel molybdenum or nickel iron.
And for these types of materials, they degrade much more quickly, and we don't completely
understand why.
And I actually think there's a huge opportunity, generally, for corrosion and degradation to
basically be studied from a scientific perspective.
I don't think that we understand these processes, the degradation of catalysts.
So that's one of the big research topics we're thinking about in this context.
Can you just talk about a few of the main things that happen that go wrong when a catalyst
degrades?
Is it coarsening?
Is it dissolution?
Is it oxidation?
What's actually going on?
So it's all of the above, unfortunately.
It depends on the material, right?
So platinum doesn't really oxidize much.
It oxidizes a little bit on the surface, but it doesn't oxidize in the same way that some
other material might oxidize and flake off, the way iron does when it rusts, for example.
But coarsening, Ostwald ripening, basically what happens is in the fuel cell and other
applications, metal particles from one nanoparticle, one little basically sphere of platinum will
leave, and they'll go to another one because it's better to have a larger, it's more
thermodynamically stable to have a larger particle with a smaller surface area.
Basically, surface area is almost always an energetic penalty.
So basically, some nanoparticles will grow at the expense of others.
And what that means for a catalyst is you have less surface area, and less surface
area is almost always bad because the surface is really where the magic happens in a catalyst.
That's where the active sites are.
That's where the chemical reaction is happening.
And so that's one of the big ones.
It's probably maybe the biggest one now.
Basically, with these ripening processes, you'll end up with a lower power density
fuel cell over time.
So that doesn't happen at a really fast rate.
We use basically polymers to kind of immobilize the catalyst and the support, which is the
material on which the catalyst rests, can prevent this from happening.
But it is a problem that we have to solve, and we should do better on.
In many ways, this problem is one of materials, but also design, it seems, right?
Because you could have a certain material that works very well, but if the design doesn't
complement it, or if the design contributes more to its corrosion just by the flows and
the concentrations of gases, that can completely hinder it.
So how do you go about balancing the design implications as well as the material implications
when putting together a fuel cell and trying to find the right combination?
Because it seems like, to me, that you could optimize one to the detriment of the other
and vice versa.
Absolutely.
So this is a really interesting question and something we're thinking about a little bit
at TRI in the context of machine learning.
So yeah, the complexity of—and this is one of the things that makes catalysis beautiful,
but also infuriating to study, is that the system itself is incredibly complex relative
to the diagnostics you would do on the lab bench scale, right?
So when most people want to do an experiment to determine the catalytic activity of a fuel
cell catalyst, they'll do something called a rotating disk electrode experiment or something
along those lines, where they're really only getting a picture of the inherent catalytic
activity, the best it could do under some mass transport limitation or something along
those lines.
But what we really need is the device performance, right?
And so making an inference about the device performance, given all of the design parameters
in the device, is a big challenge.
One of the ways in which Toyota does it, quite frankly, is just by having a big team of people
who know about the different components of the fuel cell and how they interoperate and
that sort of thing.
But we're also thinking a little bit more about kind of trying to make inferences from
what we call the lab scale, the RDE experiment scale, to either the gas diffusion electrode
type diagnostic scale or the whole membrane electrode assembly, which is the piece where
we're really doing the magic in the fuel cell.
I think because of the level of complexity, this is a place where tools like machine learning
and AI may be able to play a big role, simply because it's a place where, basically, as
you say, there's big design constraints and that sort of thing.
And the behavior really isn't the same as what it will be on the lab bench scale.
And of course, the lab bench scale is interesting just from a scientific perspective.
But we've got to be able to do both.
And to connect those two things is still a challenge.
So I enjoyed the article that you mentioned, the one in energy environmental science.
You talked about some of the ways that you can reduce the cost or reducing the platinum
group metal, right, loading of these things.
And I saw three sort of three approaches, right?
You can do the core shell approach, which I think our listeners can imagine, right?
If the whole thing's made out of platinum, but it's the stuff on the surface that is
doing something, then just don't put the platinum in the middle, right?
Put it around something else.
That's one approach.
Another is alloying.
So instead of taking pure platinum, for example, or whatever other metals you're using, mix
it with something that's less expensive or more available.
I think you guys use cobalt, right?
And then rise that a platinum cobalt, but nickel or others could also be something you
look at.
And then the last one was had to do with the electrolyte.
Can you explain that last one?
I wasn't totally following it.
Yeah.
So we have an understanding, I mentioned, you know, of like the electronic structure
of the material and how it relates.
And there's this fundamental limitation.
But the electrolyte plays a big role in the kinetics, right?
Because you've got to basically take for oxygen reduction, for example, you basically immobilize
an oxygen molecule on the surface, and then it has to be attacked by the protons and solution
to form water.
And this has a couple of different proposed mechanisms and a couple of different possible
mechanisms based on the electrolyte conditions.
But the electrolyte will influence how quickly that happens, basically because the kinetic
barrier will depend on the solvation of that intermediate.
It's not just the bond with the surface.
And there's some really interesting work that we highlighted in that article from the
Yenchel Horne's group at MIT that basically describes how you can think about electrolytes
and their ability to facilitate this reaction as well.
And a certain extent, since the electrolyte itself is a catalyst, it's basically stabilizing
intermediates differently.
And it's, you know, the transport throughout the solution could be the rate limiting step.
It's usually not, but it shouldn't be at least.
It's also the case that we don't really know much about this from an atomistic perspective.
It's hard to do simulations of the electrolyte and its role in influencing proton coupled
electron transfer to the adsorbate.
Interesting.
So is it a matter of when you say reducing platinum group, is it finding the right combination
of electrolyte plus catalyst?
Is that what it comes down to is?
I mean, actually, I think this is a good frontier to it.
It's related to stability as well.
Like the stability of materials and different electrolytes is very different.
We, you know, we draw Courbet diagrams, which are these phase maps.
And the assumption of these Courbet diagrams is basically that you have the same electrolyte
all over the place.
And it's really not even that because all of the different equilibria are actually in
different electrolytes sometimes.
But we would like to know, for example, if you have a different electrolyte, will the
stability change?
Will the activity change?
And these are questions that we have, I would say, limited answers to in specific cases,
but we don't have a very good generic theory that can be derived from the first principles
of atomic simulation.
Because of that, you're going to be sort of designing your experiments and your measurements
based on what you do know.
But is there concern that you might be measuring the wrong thing or that the types of measurements
you're doing are going to bias the results in the direction that you go with the design?
Oh, exactly.
Yeah, yeah.
So great example, actually, we talked about degradation.
We don't know whether a degradation, so there's something called, there's a nice article,
I can't remember exactly what it was, but it basically lays out the different forms
of deactivation.
So typically the thing you care about is whether your catalyst remains active, but there are
several different reasons it could deactivate.
So we talked about those, basically oswald ripening, all these different things that
could happen.
But if you don't know that, if you don't know that ahead of time and you change a design
principle or you change the way you're designing your system in response to assumption you're
making about the activation mechanism, then you might actually make it worse, right?
Because you're thinking, oh, this is happening because of ripening when we actually, in
reality, my material is just degrading completely or it's being poisoned, right?
CO is in the feed stream and it's attacking my surface and there's no available active
sites left.
The material is fine, but it's just being poisoned.
I use that word, poison is a common term for catalysis when basically something that you
don't want on the surface of your catalyst stays there and blocks all the active sites
from the thing you want it to do.
And yeah, like that could be happening.
We have been working at TRI on building systems for advanced diagnostics too.
So something that's come out in the past five or six years, at least partially because
of this notion of stability is something called ICP-MS.
ICP is a mobile technology and so is mass spec, but basically online ICP-MS can monitor
the solution, the electrolyte to see if there's any of the catalyst there, if it's degrading
in the solution.
And so this is something we're excited about using to test out some of these fundamentals,
like why is the catalyst degrading and then to be able to use that information in the
future basic design of both the catalyst layer and the fuel cells as a whole.
So looking forward at future technologies, right?
You can always try and, I don't know, change things, make them smaller, get better core
shell, these things.
But at some point, what if you try and move away from these PGMs, platinum group metals
altogether?
And maybe before we answer that question, what is it about those metals that makes them
so spectacular as catalysts?
And you sort of alluded to the scaling relations and volcano plots.
Can you describe that in a way that maybe a basic listener might understand them?
Yeah, yeah.
So I'm throwing out a lot of terms.
I'm going to throw out one more.
So this is something called the Sabatier principle.
Sabatier was a French scientist who came up with this idea.
So it might be better described as the Goldilocks principle, though my friend, AJ Medford at
Georgia Tech likes to call it the Goldilocks principle sometimes.
And basically, what a catalyst needs to do is it needs to bind a molecule that it's going
to catalyze into something else strong enough to get it to stick and get it to stay long
enough to do the chemical reaction, but not so strongly that it can't let it go after
it's done.
And so it has to have this intermediate binding energy, not too strong, not too weak, not
too hot, not too cold like Goldilocks.
And so platinum basically sits right in the middle.
For hydrogen, it's actually pretty simple.
It's hydrogen bond energy.
You can basically write the...
That's for hydrogen evolution.
For oxygen, it's the oxygen reduction.
You basically have to have this oxygen bond energy that's sitting right in the middle
of where it needs to be in platinum and palladium and lots of other of these ruthenium to a
certain extent.
They sit right in the middle.
They basically sit right in the sweet spot of catalysis, the peak of the not too hot,
not too cold, not too strong, not too weak.
There's another reason for that, which is about electronic structure.
You can basically describe the bonding the same way you would describe molecular orbitals
using platinum.
Maybe that's too advanced, but I think the Goldilocks principle, I think, is something
that maybe can be illustrative for lots of people.
So if those materials are amazing because of how they bond with the stuff that you care
about, is it simply a matter of finding sort of wild new materials out there that have
similar bonding characteristics?
And in the article you shared with us, you were looking at some pretty wild stuff.
You looked at single atom catalysis.
You looked at putting things on zeolites or moths.
You looked at different chemistries.
Can you talk about what the future holds there?
Yeah, yeah.
So it is about kind of finding things that are like platinum, but I'd argue that what
we're actually trying to do with those materials is we're trying to depart from the behavior
that governs platinum and other metals completely.
So platinum and other metals basically have this limitation that seems to apply both to
them and to oxides.
One of the things we talk about this paper was this discovery a few years ago that oxides
follow a similar trend, and they're actually penalized by the electrolyte.
So you can actually show that oxides for fuel cells can't ever be as active, in acid at
least, as metals.
And typically, if they're in alkaline, they're only going to be good as metals, and there's
some differences in stability and things like that.
But basically, for moths, for nitrides, for sulfides, for lots of these new material classes,
I think the thing we're thinking about is we basically know enough to know that these
metals aren't going to be much better than platinum.
That said, there's kind of cool things you can do.
Basically, you can interpolate.
So if you pick two things on either side and alloy them, they often have the properties
of the thing in the middle.
There's this beautiful example of this in ammonia.
So this is actually a cobalt molybdenum.
It's kind of on either side of the volcano, and you just kind of bring them together.
And there's a physical argument for that, too, which is kind of nice.
But yeah, with moths and zeolites and all these other things, I think the thing we're
thinking about is we basically realize the design constraints that platinum imposes,
and we don't really foresee beating them in the way, basically, by any metallurgical
engineering.
So the goal is to just find completely new systems.
So that kind of poses an interesting challenge, because you were kind of mentioning that a
lot of the theory behind catalysis is still being discovered.
But if you're going to create something entirely new, you sort of have to throw all
that out and go into a place where there's no theory.
Do a lot of preconceptions, then, cause limitations?
And how do you avoid that?
Yeah, this is a great question.
And of course, I wouldn't say you throw it all out, because there's still chemical
mechanisms that are happening.
So the methods that we've applied to learn about and develop the theories of metal catalysis,
I think, still can apply.
It's just a question of whether or not there are lots of things we don't know.
And I think that there's still a lot of serendipity in this kind of new field, too.
That's one of the beautiful things about it, is we don't know.
We don't know what's going to happen with moths, and we don't know necessarily what's
going to happen with all these other materials.
And there is some evidence to suggest that they have their own limitations that kind
of resemble platinums and all these other metal materials.
There are what we call scaling relations for other materials, too.
It's just not clear that they're going to behave in the same way.
Basically, the slopes of the lines and the intercepts might be different.
But yeah, at the end of the day, we don't know.
And that's honestly what's most exciting about it, is we know that we don't know enough
about this to say with confidence that it can't work.
And so that's why we're exploring it.
And so some of the most interesting results in catalysis have kind of been accidents,
in the sense that somebody predicted something would work for a specific reason.
It didn't work for that reason, but it worked for another reason.
And it was just because you were exploring.
That's the beautiful thing about science, right, is you don't know.
You don't necessarily know what you're going to find when you try something new.
So that's a good segue to the next thing I wanted to chat about, which was one of the
papers you sent us uses machine learning and an active learning framework, right?
So if you've heard, if you talk to me for more than five minutes, you've heard me rant
and rave about materials discovery, how it is largely driven by sort of luck, right?
Serendipity.
And machine learning has the potential to change that dynamic at least a little bit
in using probability and statistics to guide us towards things that have a higher probability
of being what we're looking for.
And in the paper that you sent us, you've done that.
So you were after these what you called poly-elemental heterostructures, which I didn't
know what that meant, but I read the paper and found it interesting.
Essentially, what you're looking at is combinations of lots of different elements up to eight
dimensional space, right, eight different elements.
And you form these little nanoparticles using a technique that I'd never heard of.
It was called scanning probe block copolymer lithography, which is a mouthful.
What on earth is that?
I read a little bit about it, but it sounds pretty wild.
What are they doing there?
So basically, Chad Merkin's group at Northwestern is able to, they pioneered this technology
of synthesis, right, of synthesizing nanoparticles.
And they've been involved in lots of different nanotechnology applications, things in medicine
and all this other stuff.
But the way this works, and I'm not the foremost expert on this, but I've worked with them
for a while now.
And they basically are able to create this array of nano-sized pyramids that they coat
with ink and basically depress onto a substrate.
And the inks contain some sort of polymer and then metal atoms more or less suspended
in the polymer.
And then they're able to evaporate it.
And when they evaporate it, they basically get the nanoparticles to coalesce.
And they can do this millions of nanoparticles at a time.
So millions of well-distributed, well-dispersed nanoparticles on a chip that, quite frankly,
they're at the synthesis scale.
They're synthesizing things on a scale that's hard to characterize at this point.
That's kind of the limitation.
It's amazing.
Try to go through it.
Oh, it's beautiful technology.
And what we've been able to do, the way we've been able to contribute is by helping them
with the machine learning part, right?
And it's relatively straightforward machine learning.
It's composition-based.
We're using pretty simple features from magpie, which is at this point.
I mean, the materials machine learning is going so fast.
This is only, I think, five years old now.
But that's an eternity in machine learning years.
And we basically do a Bayesian process where we are able to estimate the uncertainty and
the phase count.
And the goal of this was to show that we had control over the structure, right?
To be able to say, we want to be able to find just nanoparticles with a single interface.
If you think about a sandwich, just with two pieces of bread, right?
Instead of like five different things that we're trying to put in the sandwich.
So it's a relatively simple one.
And part of the reason this is exciting is because we want to be able to suss out the
fundamental characteristics of just these things.
We want to be able to structurally control things, design experiments where we have a
lot of control and say, this is what's causing the catalytic activity to be high.
This is what's causing the thing we want.
And if we can do that, basically it enables design of the system at a totally unprecedented
scale.
Yeah, when I read about these poly-elemental heterostructures, I assume they're going to
be alloys, right?
That maybe at the small scale, things don't phase separate.
But that's not the case.
When I looked at them, they absolutely segregate.
It sounds like that's by design that you're looking for those interfaces.
And it was interesting.
Yeah, when I saw this, I was like, OK, you benchmarked against random search.
Sure.
And you did Bayesian.
And then I saw you called this this UCB, which I'm forgetting what the UCB stands for.
But it was essentially a composition base, right?
You added, which was, I thought that was interesting.
You're essentially using the same composition based feature vector, which has made so
many other materials informatics approaches outperform because you're injecting some
domain knowledge into it, which is pretty slick.
Yep, yep.
And UCB is upper confidence bound.
So part of what we include, too, is the uncertainty.
And that helps us make a decision about whether or not we want a tight uncertainty, typically,
for the first sample.
Or if you're trying to optimize something, you're using a relatively low uncertainty.
So was this a success then?
It said that you made 18 samples, right?
How would you characterize this?
What was the figure of merit that you were after?
Was it just matching what you expected there to be in terms of numbers of interfaces?
Did you actually look at?
Yeah, matching the prediction, I think, was the primary goal.
We were actually surprised that it was successful as it was.
I mean, so you could take a phase diagram and you could kind of guess about this.
And that's kind of interesting.
One of the things we would like to do in the future is embed information from the bulk
phase diagram.
But these are nanoparticles.
So they have different behavior because of the surfaces, right?
But yeah, the predictions basically were almost always very accurate in the sense that when
we predicted there would be one interface, there almost always was.
And where we kind of wanted to push an experimental side was to see how high, how complex can
we go before our predictions start to break down?
And we're able to go all the way up to six elements and still see single interface.
And that meant that there actually were alloys, right?
There was one side of the interface that had three or four and one side that would have
one or two elements in it.
And so this complexity kind of tackling this very high complexity, which we don't have
the tools to simulate, and we don't have maybe the phase diagrams, we don't have six
dimensional phase diagrams necessarily in the literature, was really exciting.
Well, and the training data you had, looks like it was 45 entries.
So all this is happening.
So even though the exploration space is in hundreds of millions, by your arguments, you
had a tiny fraction, which is realistically where most materials informatics is at, tiny
amounts of data, a lot of uncertainty, and algorithms that can deal with that uncertainty
to guide us, they tend to do really great.
Here's a cool example of it.
Yeah, yeah.
And you know, yeah, still a long way to go, I think, in all of materials informatics,
but it was nice to see this particular success for us.
So you mentioned there's a lot of discovery, but as we kind of mentioned, depending on
the combination of different materials, some things might perform better or worse.
Do you think there's an element of rediscovery that might happen or is happening?
Uh, probably.
And, you know, there's a couple of things to say here.
One is that, you know, there's a lot of embedded knowledge in industry that quite
frankly, isn't public, that's either because of trade secrets or because of, you know,
IP and doesn't necessarily get discussed in the literature.
The other thing is that, yeah, I mean, I think that the entire kind of field of the
materials informatics is to a certain extent going to probably rediscover things, but
it's in a new context, right?
So, um, I don't know if you talk to like a conventional metallurgist, they might tell
you, yeah, like, like this is, this is, this is the, this is, I figured I could have told
you this or something along those lines, but the fact that we were able to do it with
machine learning itself is kind of like interesting and I think worth knowing.
The other thing is that, you know, throughout history, we've kind of invented heuristics
to guide us through all these processes.
Some of them have been enormously powerful.
I mean, the scaling relations that I'm describing are themselves kind of a heuristic,
right?
They basically say that everything correlates along the line, but machine learning can,
can like basically give us something that is both general and has, can kind of take
into account all the complexity, right?
And the risk there, of course, is that you overfit, which I'm sure you're both intimately
familiar with overfitting problems in material science and in the literature, but, but
nevertheless, you can kind of not necessarily just use heuristics to drive the science,
but you can have something that's even more comprehensive.
One of the things I'm excited about is, you know, we talk a lot about structure property
relationships and, and structure property relationships and catalysis are particularly
difficult because we don't have a lot of structural characterization and the structural
characterization we have to do has to be at the surface, right?
That's where the catalytic activity is happening.
And sometimes that's really different than the bulk.
And so one thing that, that is exciting, there are big data sets being created.
It's something called the open catalyst project that has this massive, yeah, massive
data set.
And what they're trying to do is trying to like answer the question, what if we had
more data?
Like, will we be able to actually create a framework for designing these?
Imagine you had like, you know, the Microsoft paint of like, of like catalysis, where you're
like just able to like create a surface and then say like, this is, this is what I expect,
you know, to happen.
I have, I have sophisticated enough thermodynamic models to sample that.
That's not, that's not really possible.
Like the force fields we have, especially for catalysis, they're not, they're not quite
good enough to do that, but, but they're, the first step is we have to collect a lot
of data and these efforts are starting to do that.
How do you design experiments to maximize the information return while also making them
as quick as possible?
Right.
Because in any sort of data science regime, right, you're going to be limited by how
quickly you can collect data.
But if the data you collect is garbage, it's not actually that useful.
So how do you balance that?
Yeah.
And, you know, one of the things I've talked about Chad's technology, Chad Merkin's
technology, the, the, the high throughput nano libraries, mega, mega, nano mega libraries,
mega libraries and nanoparticles.
But, but hardware like is, is, is one of the things we have to think about.
So we also have a big collaboration with John Gregoire and John Gregoire, he created this
beautiful system, basically retrofitted a t-shirt printer to print about a thousand
catalysts at a time on a glass substrate.
And it's, and also what he did was he was able to use something called a scanning droplet
cell where he basically drags a single drop of water across the whole thing to individually
sample the catalyst.
And he's able, he's able to create one of the largest databases of experimental catalysis
that you can actually see on our, on our website, one of our websites, data.material.
If you, if you browse the, the, the high throughput J-CAP data, it's one of the things he's involved
in.
And it's funny because like, like high throughput catalysis has actually been around for, for
a long time.
We talked about Fritz Haber.
He had a collaborator named Alwin Mutasch who tested over 2,500 compositions and over
7,000 experiments to try to find the best catalyst.
And, and it was basically a big parallel reactor array.
And when I was an undergrad, I also worked kind of on a parallel reactor array.
It only had three or four channels, but still we're always kind of doing things in high
throughput.
So that's, we had to actually develop strategies to like get a lot of data at once and kind
of build this comprehensive understanding.
We also have to be able to contextualize.
We talked a little bit about this before, but contextualize the results in the context
of kind of device performance.
So, excuse me, being able to infer basically the catalyst, it's really like a feature,
right?
Like it's, it's like, here's some low fidelity experiment, quote unquote, low fidelity experiment
that we did that's related to the device performance, but doesn't necessarily, it's not the
only thing that's related.
So being able to relate those two things is part of it.
We spend a lot of time too, in terms of kind of getting returns when we create AI methods
or machine learning methods, we like to embed them in processes and then, and then try to
determine the performance of the process itself.
So let me elaborate on that.
Cause it's, I'm being a little bit vague there.
We'd like to set up simulations of acquisition.
So, so basically you'll, you'll say I have access to some set of materials and I want
to see how quickly my active learning or adaptive learning methods are acquiring new
materials.
And we, we did this in a paper we wrote a couple of years ago.
We simulated this process for bulk materials, but active learning of, of, of catalytic of
catalysts is also something that we wrote about a few years ago as well.
And you can basically say, here's the machine learning model that we're using to predict
the thing we want.
Here's the acquisition strategy or the acquisition function.
And you kind of want to see how these basically like virtual robots are performing in terms
of how well they, how well they get new materials.
It's kind of like the experiment, the simulation experiment that you want to do.
So that's one of the things we do.
And that helps us also to tune the hyper parameters on the whole stack of, of, of active
or machine learning.
What does the, what does the future of catalysis research augmented by data science look like?
So I would like to think, and I'm not sure.
So there are a few things that I think we don't quite know if they're going to work
out, but we're pretty excited about.
So I talked about kind of big data.
I think, you know, with open catalyst projects and similar efforts, one of the things we
hope to see is a sufficiently accurate, very fast method of simulation that can enable
researchers just to, just to try new things very quickly to say like, I want to see what
happens if I have this geometry, I want to see if it happens, if I have this geometry
and this electrolyte.
So an improvement of simulation where people are able to really both accurately and quickly
test their hypotheses about the structure of catalysts and how they influence catalytic
activity.
I think we'll also see, we'll start to see an expansion of, of the methods that I've
discussed to determine mechanisms of degradation.
So that could be imaging.
There's, there's new methods of imaging in situ is going to be something that we start
to see more and more.
There's a really nice technique called ambient pressure XPS that's come, kind of come
online in the past couple of years that helps you both because it gets information about
the surface and it's getting information about in situ, whereas XPS typically is done
in vacuum.
Ambient pressure XPS is still only done at like 10 TOR.
So it's like, it's not really ambient necessarily, but it, but it's, it's getting
information.
Yeah, it's just information we've never gotten before.
And you can actually see electronic structure about electronic information about the
catalysts in real time.
There's a beautiful set of experiments too, by imaging Bert Bakehusen.
I'm not sure where he is, but basically he's able to take like videos and see reaction
events happening by, by specific spectroscopic signatures of transition states.
And that's one of the coolest things I've ever seen.
You basically can watch the catalysis happen in real time, like the individual atoms or
individual molecules that are changing from one form to another.
And it's, it's really beautiful stuff.
So, I mean, I think that's part of it and I don't want to discount the role that new
experiments are going to play because I mean, they're really going to be, I think one of
the biggest things, cause we've, we've come a long way very quickly in theory and simulation.
And I think that I wouldn't say the experiments are catching up as much as, as, as they're
they're finding new ways to like, just to see things that they've never seen before.
They've got a much, much more difficult task.
So, and they also don't have the benefit of, of kind of like being accompanied by a
revolution and computing and data science.
So how are you going to handle the differences in fidelity between new and old data then?
Because with all these new characterization techniques, right?
It means that some of your data is higher fidelity than, than others.
And how do you incorporate that and weigh it appropriately in your model?
So I think that, so there's two different approaches.
One, you can kind of throw the old data away.
And if you can, if you can collect new data quickly enough, you don't have to worry too
much about the old data.
But you can also, there's multi-fidelity approaches in which you can encode kind of
the quality of the data, right?
You can say, I want to weight this a certain way, or rather than using the old data as
a kind of, as a strict prediction handle or something to regress on, you can use it as
a regular, regularizer.
There's lots of different ways to kind of treat multi-fidelity data.
We actually, we wrote a paper on, this wasn't for catalysis per se, it was on photo absorbers
to see like how quickly you could acquire photo absorption data.
And we asked an interesting question in this, in this paper, which was basically like, if
you want all of the DFT data and like a little bit of experimental, experimental data, does
your rate of acquisition of experimental data get better?
Right.
And the answer is yes.
The answer is yes, but maybe not as much as you would hope.
And that could be, you know, of failures and modeling or failures in machine learning.
But, but it's, it's interesting to think like, has like, can, can, you know, simulation data
as a big blob, will it actually help your active learning process?
And the way we achieve that is just by one-hot encoding the source, basically saying like,
this is DFT, this is, this is experiment.
And the success of our, of our active learning method is purely in the experimental space.
We basically say how quickly can we, can we acquire experimental space?
Candidates, experimental candidates.
One thing that I think is an unsolved problem that I'm really interested in is how do you
navigate both spaces at the same time, right?
Because you typically can acquire low-fidelity data.
I mean, obviously in the past it's a little different.
But you can acquire low-fidelity at a much higher rate, but it's not necessarily easy
to reason about how you should acquire it in order to support the higher, higher-fidelity
data.
And so we kind of wrote a few algorithms on this that I think, you know, are probably
crude, crude, crude first steps in this field.
But I think I'm excited to see more of that from the field is how do we, how do we treat
the data from the past exactly as you say, and how do we like treat what we would call
lower-fidelity data streams?
Very cool.
I'd like to say one more thing actually about, about this.
Yeah, go ahead.
About the past.
So, so one of the things that we probably want to do too is to figure out a way to extract
it and to like put it into a format that's useful.
And one of the things that's going to be necessary for that is NLP.
So at TRI, we fund a big project actually on natural language processing of, of the
literature, right?
And one of the things that's become very cool is being able to extract named entity relations.
Basically, like if you have, if you want to like parse the entire literature and say,
I want all of, all the catalysts or all the supports, and I want to like have some associated
information about them.
You can, you can, you can do that relatively efficiently now with tools like Math Scholar,
which was developed at LBL under TRI support.
They basically were able to take the information extracted from the literature and actually
like write representations of the elements that helped them discover some new thermoelectric
materials.
But I think this is going to be really critical to using the past.
Because basically the first thing we have to do if we want to get data of any fidelity
from the past is to get it right.
And we have to, we have to comb the literature either by hand or ideally by, by automated
methods.
And I think NLP is going to be critical to that.
That's rad.
Jim Warren talks about this sometimes that like, if you look at the rate at which data
is being generated, it basically dwarfs everything that existed in the past.
The same idea that like there's more scientists alive today than all the scientists that have
ever lived ever, right?
And so I, I heard the argument and I'm not sure if it's right or wrong, but all we really
need to do is figure out the data capture problem.
And then we don't have to worry about scraping data from the past because it's going to be
dwarfed anyways.
And I think that's kind of an interesting paradigm to think about.
At the Toyota Research Institute, we're laying the foundation for the next generation of
mobility technology with groundbreaking research in energy and materials, robotics, human-centered
AI, and machine learning.
Visit tri.global to learn more about our mission and our exciting work, or to browse our career
opportunities.
If you're interested in materials informatics, check out our GitHub page, github.com slash
tri-amdd, where you can find open source software for the early prediction of battery
lifetime, autonomous density functional theory calculations, predictive solid state synthesis,
polymer molecular dynamics, and more.
You can also see various apps and data repositories from our research portfolio at data.matr.io.
If you're a graduate student interested in TRI, we're currently accepting applications
for grad student internships for the summer of 2023.
This year, we'll have projects on battery informatics, machine learning accelerated
spectroscopy, polymer discovery, and for the first time this year, carbon neutral techno
economics.
Past internship projects have resulted in patents, peer-reviewed publications, and conference
proceedings.
Our past interns have gone on to found companies, become professors, lead government agencies,
and many have joined us full-time at TRI.
You can find more information about available internship positions in the careers section
of our homepage at tri.global slash careers.
The Materials in podcast is sponsored by Materials Today.
Visit materialstoday.com to stay up to date on the latest happenings in the material science
field, and you can read a couple of their articles.
In fact, I'm going to have them include two articles on catalysis that I thought were
pretty rad in the show's notes.
Those will be available for free for a period of time, I think six months, so you can check
those out.
But in general, we think that they're a cool journal.
They are leading the field of material science, so check out them.
You can also check out lsevere.com to learn more about the journals, conferences, books,
and things that they put together over there.
And thank you as always for listening to this episode of the Materialism Podcast.
If you have questions or feedback, please send us emails at materialism.podcast at
gmail.com.
And make sure you subscribe to the show on iTunes, Spotify, Google Podcasts, or wherever
you find your podcasts.
And if you like the show and want to help us reach more people, consider leaving a review.
It helps us improve and it exposes new people to the show.
Finally, you can check out our Instagram page at materialism.podcast and connect with us
and let us know what new material you'd like to hear about next.
Before we go, we'd like to give a shout out to Alphabot and Colobite for making the music
for the podcast.
They both make a ton of really cool synthwave music, so go check them out on Spotify and
YouTube.
The inventors of fire, electricity, magnetism, iron, lead, glass, silk, cotton, the makers
of tools, the captors of lightning, the architect, the engineer, the musician are all beneficiaries
of the materials of this world and are bound only by their imaginations in manipulating
those materials.