How does a company like GE use machine learning to further their research into new materials and technology? Find out in the first episode of our General Electric sponsored series where we sit down with two experts at GE Andrew Detor and Kareem Aggour.
This episode is sponsored by GE, a leading company in a multitude of fields. Check out their research 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: Materials Informatics Machine Learning General Electric
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. Hello and welcome to the Materialism Podcast, an exploration of the past,
present, and future of material science. I'm Taylor Sparks, your co-host, and I'm joined,
as always, by my trusty co-host. How are you, Andrew? Good. Finally graduated. Ready to start
chapter two of my life. We are not just joined by regular Andrew, but Master of Science Andrew.
Congratulations, Andrew. Proud of you. Thank you. And we are today super duper excited. We are
bringing you live to experts from General Electric as part of a fantastic new series that we have
together. It's a series of episodes we're going to have on General Electric, so stay tuned. There's
going to be about half a dozen of these things, and they run a wide variety of super cool topics
related to material science. Now, if you haven't listened to our prior episode, that's episode 20
on materials informatics, it might be a good idea to check that one out first because this is going
to build on that and get some of the jargon and vocabulary out of the way, but you're probably
fine to listen to this one regardless if you don't want to. This must be one of our all-time
most requested follow-up episodes. People loved the materials informatics episode, and maybe that's
because like Andrew and I work in that area and some of the enthusiasm bled through. I'd like to
think that's part of the reason. Whatever the reason, people have wanted more of that, and today
we are going to talk about materials informatics with General Electric as a case study and with
their perspective from industry. And to do that, we've got two folks with us. We're joined by Andy
Detour and Kareem Agour. Andy, Kareem, want to do an introduction? Yeah, sure. Taylor, Andrew, it's a
pleasure to be here. I really enjoy your podcast and just like to talk with you about what we're
doing at GE in materials informatics. So yeah, my name is Andy Detour. I'm a senior principal
scientist at GE Research, been with the company about 14 years, backgrounds in high temperature
metals, and recently been getting into machine learning and understanding how it can help us
develop new exciting materials faster. So really looking forward to talking about it. Fantastic. How
about you, Kareem? Okay, yeah, thanks for having me. I'm really excited to be here as well. Again,
my name is Kareem Agour. I'm a principal scientist at GE's Research Center in upstate New York,
and I informally lead a team that does a lot of research at the intersection of knowledge
management and data management. So what we're trying to do is capture a lot of the knowledge
that's typically captured, that's really locked away in experts' heads and trying to put that
into computable form so that computers can reason over that kind of knowledge to make decisions
in very similar ways that humans do. So Kareem, this is fascinating that here we are on a material
science podcast talking to a knowledge systems expert. What is your background? What led you
to get involved at General Electric? So my background is actually much more around scalable
data management. So I do a lot of work around digital thread and data management, particularly
around big data management. But really over the last four or five years, I've been really working
with folks like Andy Detour, and particularly I really kind of first got engaged around
additive data management. So I was really approached by folks like Andy from material
scientists who said, hey, within the additive space, we're generating a lot of data and really
a lot of variety of types of data, a lot of relational data, a lot of images, a lot of time
series sensor data. And we're really struggling to be able to put all this data together,
analyze it, really try to understand, do a lot of root cause analysis, understand
the sources of defects. And so they came to me and said, what are some ways and
approaches that we can start to bring this data together, link it together, and analyze that data?
So at the time, we were really using a lot of different knowledge-driven approaches, using
semantics and ontologies to model different types of data to link that data together. And so I
proposed that we use these kind of technologies, these kind of approaches to link that data and
analyze it. And so kind of one thing led to another. And so I kind of got hooked up
with Andy to look at materials informatics as kind of an approach to use ontologies and use
semantics to model data, link data, and analyze data. And that kind of led us to looking at and
exploring approaches to kind of looking at second wave and third wave approaches to kind of analyze
material science data to look at ways that we can start to address the challenge of how do we
actually kind of accelerate the development and discovery of things.
Rad. Yeah, I think it's a good application too, because data and material science is so varied,
right? A lot of it is quantitative, but then there's a lot of qualitative aspects to it too.
It's not usually as easy as just putting it into some sort of tabular form and running
a simple model on it, right? It has many different dimensions that are not always easy to interpret.
Well, before we get into what I'm sure is going to be exciting to hear about how that process
has worked for the few people that have been living under a rock for the last, you know,
a couple hundred years, Andy, can you tell us, you know, General Electric, what's your mission?
What do you guys do? What's GE research about? Yeah, sure. Well, you know, when I say General
Electric, a lot of people think light bulbs and refrigerators, and we actually don't do that
anymore. So GE today is a high tech company focused in three main businesses. So aviation,
we make aircraft engines. Roughly two out of every three commercial flights is powered by a
GE engine, which I think is an incredible stat. Yeah. We have a whole power gen business,
so wind turbines, gas turbines, one third of the world's electricity is generated through GE
devices. And then we have a healthcare business that most folks don't really know as much about
that does a lot of imaging. So MRI machines, ultrasound, I think it's something like 16,000
scans per day done on GE machines. So these are the three core businesses and the research center,
we work with all of them. And about half of our business is directly engaged with
those folks trying to help get new products to market, trying to solve problems with existing
equipment. And then the other half is with the US government. We do a lot of
really far out there, government funded projects with DOD, DOE, NIH. So it's a really interesting
place to work because we get the real near term applied business research and development,
but we also get to work with academia and government funded programs. So it's a really
dynamic place. That is awesome. So what prompted you guys to make this big leap and start thinking
about informatics? Because it's not trivial, right? I think of just the old schoolers that
I've interacted with in material science who are like, this data science is a load of nonsense
sort of thing. Like, was it a hard sell to get GE on board with this concept?
It wasn't a hard sell. And I'd say at this point, we're really doing it out of necessity.
We've kept our eye on machine learning, materials informatics, and the great stuff that
academics are doing. And I'd say about five, six years ago, there was two real projects that
required really materials informatics to just wrap our heads around the complexity.
So Kareem mentioned it early on additive manufacturing. You know, GE makes and sells
additive machines. We make lots of parts in our engines with additive. And it's such a complicated
process. There's so many variables that go into it. And so we wanted to try to understand that.
And one of the best ways to do it is with these modern tools that can capture all that complexity
and allow you to see trends that you wouldn't be able to see otherwise.
So maybe we can hop into some of those case studies that you have,
specifically on additive manufacturing, and how GE approached that problem,
and what were some of the considerations and how machine learning helped you address those?
Yeah, sure. So yeah, Kareem, you know, jump in if you want on this, because you were heavily
involved from the beginning. But as a material scientist, we're interested, you know, in process
structure properties, right? That's the core paradigm. And so the program that we had specific
to additive, and we're still doing this work, is what are all of my processing variables? What are
all the knobs I can turn on my machine? Coupled with what are all the different material
compositions, the different alloy compositions I can put in my powder bed? And how can I,
for any arbitrary composition, define an ideal processing window, where I have a,
you know, robustness, where I'm not going to be generating any defects,
and I can build a material that has the properties that I need.
And so part of the work Kareem did was to really help build a platform to
ingest all this complexity of data and a framework where the average material scientist,
someone like me, could go in and use a platform to look for those trends and to try to identify
what direction to go in for the next set of experiments, and really try to get to an answer
faster than we could have otherwise. So Kareem, I think it's interesting,
because in materials informatics, we typically sort of see two approaches. There's some folks
who are like, we've got so much data, all we need is algorithms that can capture these subtle
threads of patterns in the data. But the data's there. The trends are there. All we need is a
tool to capture those trends and to then exploit them. But then the other school of thought is
really like, we actually don't have that much data. Instead, what we have is this massive
space with many processing or many compositional tunable parameters. And what we need is an
approach that can iteratively tell us what's our next step to most effectively explore this space,
so minimizing your uncertainty or something like that. So this is sort of like the Bayesian
versus frequentist approach. And I was curious if GE has approached this with one approach or
the other, or if you're trying to do both, or if you can speak to that a little bit.
Yeah, it's very much the latter kind of problem. And so we actually ended up going with the Bayesian
hybrid modeling approach, because the search space was so vast, and the ability to explore it
was so challenging, because the limited number of additive builds that we could actually run
compared to the size of the search space. So what we ended up, as I said, was going with
kind of a Bayesian hybrid modeling approach, where we developed a BHM model that would
analyze the data iteratively and give us recommendations on the next best experiments
to actually run within the search space, and then iteratively tell us, based on the amount
of experiments that we have and the amount of data that we have, what the next set of
experiments we should be running, so that it could iteratively narrow the search space down
to identify what the overall best search space or volume would actually be. And so as Andy was
saying, what it ended up doing was telling us what the overall envelope would be for the optimal
build space, so that instead of just giving us a single point solution saying, this is the best
process parameter combination, it would actually give us a window saying, within this overall
process parameter space, this is kind of the envelope, so that if there is slight variations
in the build, we're still giving us the minimal amount of defects for the parts that are actually
being produced. So I think most of the field of material science has kind of approached that the
same way, because we just don't have, first off, huge data sets. And even if we did, I think it's
kind of dangerous to assume that all of your learning is there, you just have to extract it,
and then you'll be able to make these perfect predictions of this amazing new material.
Because when it doesn't work, everyone's like, well, what's wrong with it, right? But the approach
that you described is totally different. It's like, hey, we're just going to get better and
better. We know that there's a lot of unknowns, but this is a way to systematically explore.
In other words, you were going to do a bunch of experiments anyways. Let us help you pick
the best next experiment using a statistics approach. Yeah, and I think the systematic
aspect of it is really the key there. Because a lot of times, if you're in a really high-dimensional
design space, it can feel like you're just playing a game of darts, where you're just kind of
randomly throwing it, and maybe you get a hit, maybe you don't. But having that systematic
progression through that high-dimensional space that's difficult for humans to visually
comprehend is key to actually leveraging the power of that data. Yeah, and I think it's so
important to make sure that the final recommendation, it can't be that you're
trying to balance on a nice edge and saying that the resulting process parameters can't be
this exact perfect value. Otherwise, you're going to end up with a part that's
terrible. You've got to make sure that there's some flexibility in the results, because you're
not going to get the exact laser power, the exact laser speed that you want. There's got to be some
tolerance for slight variation around that. So talk about high-dimensional search space.
Let's talk high-entropy alloys. Oh, geez. I've seen the math on these.
So I'm not going to go through any of that math, right? That's like the first slide in
everyone's presentation at these high-entropy alloy conferences. But suffice to say, well,
for those who don't know, high-entropy alloys are very unique. They're unlike most conventional
alloys that are built on a base alloy, a base element, excuse me, like steels are based on iron.
Most of what's in steel is iron. High-entropy alloys sort of throw that out the window and say,
I'm just going to take elements from all over the periodic table and throw them in a pot,
mix it up and see what happens. And so you can have four or five, six elements,
all in significant quantities, 20, 30%. And it's just an unexplored space.
And it's surprising too, because they tend to make simple structures, right? Unexpected to me
anyways. That was very surprising. Yeah. They tend to make simple
structures and they tend to exhibit other properties that are very unique. So that's
why GE is interested in high-entropy alloys. We have actually a couple of projects now that are
really deep into these systems. I assume it's for your refractory,
like you want the high temperature stability. That's right. That's right. Mostly in refractory,
high-entropy alloys. And the promise there, conventional refractory alloys, the Achilles
heel has always been oxidation resistance, which is kind of funny because they're so good
mechanically at high temperatures, but they form such volatile oxides that you have to
apply coating. So the question is, can you get a better balance of properties using a
high-entropy alloying approach in the refractory space? And from an MI perspective, how do you do
that in such a massive compositional space? Right. Right. And this is where it was necessity
for us to adopt and learn, really learn on the fly, how to do materials informatics,
how to do machine learning, data strategy, high throughput experiments.
And so, you know, we had a project that was two years ago and we were looking at
refractory high-entropy alloys. And what we did was we set up a high throughput experiment
where we could run about a dozen alloys per week, which may not sound like high throughput to some
people, but high temperature structural metals, that's pretty, pretty quick. And we were trying
to do a multi-objective optimization process here because we needed oxidation resistance,
we needed ductility, we needed strength at high temperature. And so we set up these experiments
to quickly assess those key properties. And we fed all of this into a machine learning model,
which we ran in an active mode or sequential learning, some folks call it. And it was a cool
project because we, at the time we didn't yet, you know, quite fully trust machine learning.
So we ran our experimental campaigns one of two ways. One way was we said, we're just going to
let the computer tell us what to do. Let's do Bayesian optimization, give us a list of alloys
to run next and we'll go do it. The other half of it was let's use good old metallurgist intuition.
Let's just try to understand it in our brains and think about it. And we kind of went back and forth
over a 15 month period. We ended up screening 367 unique alloys and discovered a space that's
very unique, a Pareto front that's very unique using these machine learning methods that I don't
think we would have found if we did it, you know, quote unquote, the old fashioned way. So it really
opened our eyes at GE to the power of these methods. Dude, that last part is really what
gets me like, how do you prove that it was better than what was out there, right? So like sometimes
people do like random search, they'll take some of their experiments, like, you know, had we done
this with random search, this is where we would be at. But it's easy to like, it'd be like, okay,
well, you found some stuff. He's like, no, that was hard. Like how we would have never known to
look for these things. One question I've got for you, Andy, is how did you pull off the multi
objective optimization? Because I've seen several different ways where people try and do this. I'm
curious what you guys did. Yeah. So this one, this one, we were worked closely with our designers.
So we were trying to design a system here. And so there, there were requirements to sort of
minimum viable property values that we had to hit across all of our metrics. And so once we had
those sort of minimum viable, you need to have at least, you know, this melting temperature,
you need to have at least this strength. Once we understood those, then we just wrote out in
this case, we just wrote out a multi objective function. Okay, so each term had was one property
scaled by the min viable value. So your loss, your loss function was a single value,
which was a composite of several things you cared about. Okay, that's right. Yep. And then
the system was just seeking to maximize that objective function. I think that's the most
common way I've seen it done. But there's some cool emerging ways that get around that instead
of a single loss function, treat it in a true high dimensional space. Yeah, I wonder. Well,
a lot of the issues with building like a utility function like that is that you kind of set the
relative importance of these different metrics prior to actually running that. Just because a
lot of these things exist on very different scales. You know, one might be a qualitative
assessment where you're saying was this coding good or bad, and the other might actually just
be the hardness. And so trying to put those into a single utility can be an interesting challenge.
I think multi objective optimization is really an interesting problem, because the way that you
tackle it can have a huge impact on the results that you get the approach and the and how far
into the Pareto front that you get. So I think that was, I think the utility functions a great
like approach to it simply because like, by combining all these things into one,
and I lost my train of thought. I'll say that it is clearly an open research area,
because when I go to conferences, I hear all sorts of people talking about like,
what is the right way to deal with multi objective? It's an it's an open question,
I think, in this evolving field. And it's enabled by materials informatics,
materials informatics is allowing us to do multi objective optimization in a quantitative way,
because now we have these surrogate models that are fast acting that we can run in any of these
optimization frameworks. The other direction I think we're we'll be going in the next decade
is is trying to couple these surrogate material models into multidisciplinary design optimization
framework. So MDO is very mature, right? They're integrating structural aero thermal models,
they don't yet really integrate materials in a quantitative way. So I think
that's going to be a big part of the future enabled, again, by materials informatics.
Well, pivoting back to this sort of like direction, then for GE, I don't remember when we
first met Andy, but I do know that one of the early times that we interacted was when I learned
about ml4m machine learning for materials. And this is a General Electric initiative,
which began what last year, I think in 2021. That's right. And I know that you've brought
in some speakers. I know that's how I got involved in it. I had a chance to sort of say,
here's what's happening in academia in my little corner of academia. But can you tell us more about
ml4m? And what you're hoping to achieve with that? What are the goals? What's the process
that you're following there? Sure. I remember when we met, I think it was at MRS Boston
conference. And I remember going to your talk. And I was like, that's the guy who has the podcast.
That's so great. So yeah, we started this initiative we call machine learning for
materials or ml4m at the beginning of 2021. And really, it was a conscious push that our
leadership kind of drove to build or foster materials informatics at GE research and across
the General Electric company in general. And so I and a team of some other folks who are really
passionate in this area have been working at over the last year and a half. And we really break it
down into largely three categories. So one's focused on our data strategy, kind of getting
our house in order around data. Another group of folks are looking at machine learning case studies,
which is, let's just try some quick and dirty projects and show people how effective and how
powerful these models can be. And then we also are looking into high throughput methods, because I
think it's it's like peanut butter and jelly, right? Machine learning and high throughput
experiments go hand in hand. So we're trying to think about ways we can generate data sets quicker
using high throughput. Rad. So, okay, let's step through some of the cases. And so you've got this
data strategy. We've talked about case studies, high throughput materials, third wave AI, where
are you in that? Are you trying to do all these things all at once? Are you sort of systematically
approaching it? What's the strategy here? Yeah, we're doing them all at once. And I mean, I could
go into detail in each one and bore you all with that. But we're trying to do everything at once,
we've got a team of I know, there's probably 20 people who are spending, you know, a fair amount
of time on this really trying to push it forward. We are also hiring in this area. So we actually
stood up a materials informatics lab this year, and where we're, you know, looking for the best
and the brightest, come check out GE research, because we're looking for people to have a
material science background. Everyone's invited except for Andrew. You have to stay here and help
me with the podcast. Yeah. That's awesome. Yeah. So was there something specifically you wanted to
Well, I mean, let me kick it over to Kareem. Like, tell me like, if this was at general or not,
if this was at say, Netflix or something, the data strategy would look one way.
What are some unique aspects about managing data at a big materials company?
I mean, really, I think that the big challenge is, you know, for a company like GE, I mean,
we've just got this, you know, massive legacy, you know, 150 plus years. I mean, the reality at
the end of the day is that GE is, you know, a materials company. I'll just say, I heard this
fascinating quote from Jim Warren. I'm curious to hear your opinion on it. He essentially argues
that the amount of data that's being generated, like there's more scientists alive now than have
ever existed in the past. Like that's, that's a true statement. And it might arguably be true
that the amount of data that we're generating for materials right now dwarfs anything that
existed in the past. So his sort of idea was like, who cares about capturing the old stuff?
Because it is hard. It's in like old lab notebooks. He's like, who cares about that?
Do we just focus on capturing, getting a plan in place that captures the existing data going
forward? Or should we really try and look backwards? I know Andrew's got thoughts on
this, but I'm curious what you think Kareem. No, I think, I think we absolutely need to
make sure that we have a rigorous process for capturing the data that we're generating today.
But I absolutely think it's critical to capture all the historical legacy data as well.
Because there's a lot of insights. And I think there's a lot of value to all the experiments and
all the, all the things that have been done in the past. I think there's a lot of, you know,
value to that legacy, to that history. But I think it's more than just the data, the bits and bytes
of the data itself. I think there's also a lot of value to what's kind of locked away into the
material scientists, experts minds as well. I view that as just as critical, if not more so than the
data itself. And I think that that's a really rich and really important part to, you know,
the kind of the bottom of the third wave AI aspect as well. So from my perspective, it's not just the
data, it's also the domain expert knowledge. That's really a key part of our strategy, part of our
materials informatics strategy as well. So how do you leverage that? Yeah, I think you really have to
understand how you can get at that, how you can capture that from the domain expert knowledge,
and how you put that into a computable form. That's really how you leverage it. It's fantastic
that, you know, Andy and, you know, guys like yourselves have that domain expert knowledge,
and you can kind of use that to guide and shape the work that you do. But I think you really have
to actually extract that from the domain expert knowledge, from the domain expert, and put that
knowledge into a computable form, and get that into a computer so that it can reason and use
that knowledge, you know, in much the same way that you guys actually do. So one form of domain
expert knowledge is equations, right? So in our heads, we like the Hall-Petch equation, we all
know, right, that yield strength will be proportional to one over square root of grain size. And so you
could put that into a knowledge graph. If you had the grain size of material, but you didn't have a
strength prediction, you could use Hall-Petch, for example. But then beyond that, there's, you know,
more qualitative reasoning that we have, like you, you know, you turn the temperature up, and,
you know, the flow stress will go down. But we don't really have an equation for that, per se,
but we know that if that goes up, this goes down. So how do you incorporate that sort of knowledge?
Yeah, is this like interviews, or taking like the write-ups and hitting it with NLP,
and then creating these, I guess you talked a little bit about this at the AIM conference this
February or April. Can you talk a little more about that? Because I think that that is such
an interesting area, but it's not quite clear to me how we capture that knowledge.
It's not quite clear to us either, but it's something that we're working on.
I love that you're working on it.
Yeah, no, we're working on it, because we see this data scarcity problem, and it's going to
always be a problem in our field. You brought up, you know, Netflix or Facebook. So I ran some cool
numbers here that I think are pretty provocative. If you look at just how much data is generated on
the internet every day, it's about two and a half million terabytes. If you look at all of the data
stores that are accessible in material science, plus, you know, the 17,000 design curves we have
at GE, it's about 57,000 terabytes. And so every 30 minutes, there's more data generated on the
internet than the collective whole of all that's accessible in material science. So like, we're
not going to have the levels of Facebook and Netflix, you know, data ever. But what do we have
that they don't? We have physics. We have, you know, these materials are driven by mechanisms
that we understand. We have domain expertise and experience. So that's the point is, how do you
include that on top of the data to, you know, push things forward?
Brad. Well, Andrew, I'm kind of curious from a case studies perspective, jumping back to the
case studies, among the two you mentioned, or others you want to bring up, had there been like
smoking gun, like amazing, just killer success stories, maybe not allowed to talk about them,
they might be preliminary, but I'm curious, has it been like, holy cow, like we've hit the holy
grail already? Or is it like, just tantalizing evidence that you're on the right track?
Well, I can't say that we had the holy grail for us would be to develop a new material or a whole
new process that's in production and, you know, being used out by the world, right? I can't say
that we're there yet. But there's been so many tantalizing aha moments have come about with
materials informatics that it's a snowball. I mean, it's just building now. I think our
businesses are starting to understand that where material development used to be like a 10 year,
10 plus million dollar prospect, and folks are very hesitant to invest in that, especially today,
where technology is moving so fast. What we're seeing at GE and what I'm sure other industries
are seeing as well as the materials informatics is going to sort of flips that on its head,
because now I think we can develop materials at a closer pace to product development.
Well, how about this? We were at this conference recently in Denver, and Andrew,
interrupt me if you want.
Yeah, I was just gonna say, I think it's really telling and important that GE is being so public
with its materials informatics initiative, because a lot of people in industry are still
kind of scared of it. And, you know, they're waiting for somebody else to take that leap.
And I think seeing some of these big engineering companies finally jump in and show success,
or at least progress towards using them is going to be huge and pushing other,
you know, players in the industry to also adopt those methods.
Totally agree. Awesome leadership role.
And we owe it to folks like yourselves in academia, right? Because you showed us the
way really. I've been reading papers about machine learning and material science for
over a decade. And, you know, starting from MGI, right in 2011. And to be honest,
when that first came out, I kind of wrote it off, right? Oh, this is just another
hype cycle. But I don't think it is. I think that this is the way material science is going.
That's awesome.
Couldn't agree more. Yeah, couldn't agree more.
So Andy, you and I were at a automated, Karim, were you there too? I don't remember.
But we were at this automated materials research and development conference, which was
fascinating. It's not something I've been to before. But imagine like, if you took the humans
out of the lab, and it was just robots, right? And that's kind of the idea, or that's like the
end goal, right? Could you actually remove humans from the lab or had them do something else where
their skills are more suited? This automated workflow clearly coincides with one of your
goals of ML4M, which is high throughput materials. What's the status of sort of automated R&D at GE?
Are you not even there yet? Are you beginning to explore it? I know I heard one of you guys
talking about like, you're still arc melting things and how on earth do you automate like
arc melting, right? Or, you know, you could say the same thing with many other processes. So I'm
curious, what's your take on that? Yeah, I mean, we're just beginning,
to be honest with you. We don't have robots running around our labs doing all our experiments
for us yet. But again, I think this is something that is coming along now because of this data
scarcity problem. And it's, I think there's two ways to solve that, right? One is what we've been
talking about with third wave AI, and how do we get smarter about what knowledge and information
we can build into our models. But the other way is kind of like a brute force method, right? Can we
just automate our experiments to get them done quicker? And so I think that's a big part of the
future. We've got some examples of where, you know, I mentioned 12 alloys a week, but that was
a lot of it was manual labor. But it's just getting smarter about how you batch things,
smarter about how you run your experiments. And this notion of it's more about direction
than perfection. You know, something we hear about just to guide me in the right
direction quickly. But it's just starting. I'm kind of torn on this because on one hand,
I'm like, man, I have some grad students who just have like the touch, they can just make materials
and they're so good at it. And so on one hand, I'm like, man, if I could just get the robot to be
that grad student, they could just churn out great materials all the time. But on the other hand,
I know that a lot of the awesome discoveries in material science were accidental, right? It was
like that, oh, oops, I weighed the wrong amount in there. And then you go back and you figure out,
well, what went wrong to give you this amazing property? Like, are we going to sacrifice that
sort of serendipity if we move towards automated lab where they're just following the script from
the ML algorithm, like you've got a Bayesian tool that's saying, here's your next experiment.
And we just follow that descent like down to wherever it leads us. Are we going to miss out
on these sort of serendipitous discoveries, do you think? I think that's a risk. And we need to be
thoughtful of that. I think one way around it, I think, is that you look at high throughput
experiments, the cost now of doing an experiment is going to go way down. And so to go out and make
mistakes is affordable, right? And so one thing you could do to try and build that serendipity in
is open your search space up beyond what you would do if you had to run old fashioned experiments
that were really expensive. So I think that's one way to try to find those surprises is just
open up that search space enabled by high throughput methods. Yeah, I do a lot of work with
genetic algorithms. And one of the concepts within the genetic algorithm is this mutation
operator, where you randomly will introduce different mutations to your optimization dataset
that are stochastic in nature. And that allows you to break out of local minima or and prevent
premature convergence. So it's almost like you need like the Bayesian to tell you like, here's
your next experiment. And then you need to follow that with a genetic step. Chaos, yeah. I actually
kind of like that idea. Yeah. So as you starting to develop your own internal datasets, processes,
procedures, one of the big questions that's still not really settled in the materials informatics
community is, what's the format of your data? And is it compatible with a lot of the other formats
out there? Because Citrine has a data format, Materials Project does MPDF. Yeah, they all have
different ones. How have you approached formatting your data in a way that's not only compatible with
maybe people's expectations of how materials data should be formatted,
but also allows you to leverage as much interesting other aspects such as the
ontological facets of the data? This is a really hard question.
On one extreme, in our businesses with their design data, like our aviation business has
17,000 plus design curves, right? That data is pretty well structured. And there's a format
for it. And we have a product that categorizes that data and makes it queryable and APIs and all
that. But on the research side of things, our work's very messy. And we're all over the place,
right? And so I don't know if there is one structure that's going to capture all that.
I'm not sure if that's the direction we should be heading. One of the things we're experimenting
with, and actually, Taylor, this is building off of the data in brief journal, is we're calling it
for the moment, a data share article, which is we have internal reports that we can write a GE
proprietary, we could put them in our library. We're trying to spool something up now where folks
can publish their data sets and be able to make them accessible to all GE employees with a two,
maybe three page article that just describes that data, its provenance and what it all means.
I think that's a good next step forward for us. We've talked a lot about what should be the
perfect end goal. And this is like on a big Slack channel with materials informatics people. And
people are like, no, it should be this, it should be this. I think that we can't let perfect be the
enemy of good in this case, right? Maybe this data in brief or whatever this format you have is not
like the end all be all, but it's a step in the right direction. And it's initiating a culture
change where this idea of my data is not my own. It doesn't belong in my lab notebook for me to
look back on. This is company data. This is going to help many other people. And that means I have
to make it in a format that helps other people, even if it's not perfect, like taking a step in
the right direction. And where, where this really bites us is in our characterization lab. So we have
a really, really nice characterization of all kinds of microscopes and TMS and OJ. And we've
got all this. So we generate a ton of images and data in that lab. And it, you know, someone,
someone at some point came up with 90%, 90% of that information just got, it's kind of squirreled,
locked away into share drives and folks, hard drives, and we got to make that accessible.
It's too expensive.
I'm a big believer that communities really need to adopt, you know, standards, you know,
for sharing data and that, you know, internally you can represent data however you want, but you
need to map to some common data models and common data standards to be able to share data across,
you know, organizations and across boundaries.
So Kareem, I have a question for you now, as we're sort of approaching the end of the episode,
I'm curious about this third wave AI idea, right? So we're not just building models that
can find patterns in data. We're not just like building really good models that can deep learn
or new ontology for that. Like we are after new contextual understanding, new mechanistic
understanding, extracting insight more like to you, what is third wave AI and are we going to
achieve it? What, what, what's missing for us to achieve it?
So really this concept of third wave AI really comes from DARPA. It's, it's not, you know,
a GE idea, but it really, maybe we should take a step back and describe, you know,
the first and second wave as, as DARPA describes it. I really, I guess the first wave is as DARPA
defines it really, I guess, became popular in the seventies. And the idea of first wave is
these really kind of handcrafted rule-based experts systems where domain experts would
kind of create these very brittle, you know, if then else type rule-based systems.
I can kind of think of like, you know, TurboTax and these systems, you know, work really well
with the input data conformed exactly to, you know, what the system designers kind of expected,
but they really, you know, fell apart if there's any variability, you know, in the inputs.
And these are, you know, kind of second, second wave systems, which is really,
I think where we're kind of at today. And it became really popular in the nineties and early
two thousands. And they're as DARPA describes them, but really kind of the statistical machine
learning approaches and the basic idea there, you know, and, and really the second wave techniques
are anywhere from, you know, simple, you know, linear regression all the way up to, you know,
what's ubiquitous today, which is kind of like neural nets and deep learning systems.
And really the idea is to really extract, you know, hidden patterns in data.
And, you know, they're really very powerful and there's been a lot of, you know, really
big successes in these second wave approaches. But there's also been some, you know, major issues
and limitations. And one of the big challenges with them is they're typically, you know,
black box systems. And so we can't really understand kind of the patterns that are
being extracted in them. And so they suffer a lot of limitations because they're really,
you know, constrained by the limitations of the data that are being used to train those systems.
And so if there's any sort of, you know, skewed or gaps in the training data, they can also often
lead to errors. And so what DARPA really means by, you know, third wave AI is they're really
looking for the ability to have AI systems that can do what they call contextual reasoning and
adaptation. And so instead of having these, you know, black box neural net systems that,
you know, take a million images of cats and try and learn some sort of statistical representation
of what a cat is, they actually want an AI that can develop kind of a contextual model of what it
means, you know, to be a cat, you know, what's the, you know, the body of a cat look like,
what's the head, you know, shape of a cat, you know, pointy ears and whiskers and so on and so
forth. And so when you see like a new video or a new image of a cat, I can say, oh, you know,
I understand structurally what a cat looks like. And therefore, you know, this looks like a cat.
So I can conclude that it's a cat. And so if you know, you know, if you tell an AI system,
you know, it's never seen a tiger, never seen a lion. So, you know, say a tiger is a big cat with
stripes. You know, you can, you know, show it, show the video or an image of a tiger. It says,
oh, you know, it's kind of got the structural shape of a cat and it's got stripes. Therefore,
I can conclude, I can adapt and conclude that it's a tiger. And so I think that's kind of the
basic premise of what, you know, DARPA means by third wave, you know, so you can have contextual
models to actually understand what these, you know, real world things are like a cat and actually
adapt to make conclusions and, you know, kind of come up with new concepts.
It certainly sounds like the more broad definition of like generalized AI,
right? Can you build a model that does more than just what it was strictly trained on,
but generalize it out, right? This is a big problem for all of machine learning research
right now. And it'd be cool to see it happen with materials. So Andy, this has been fascinating to
hear about, like in theory, what this looks like. Do you have an example for me? Like,
I would love to hear where you've actually tried this out.
Yeah. So third wave AI, as Karim explained it, it can, it can sound really daunting and like how
it can't, it's tough to wrap your head around it. So one of the things we tried to do, Karim,
myself and a, and a small team at GE research, we tried to apply this to a real material science
problem. So let's like, so we, we picked a problem that was like a classic materials development
challenge, coatings on high pressure steam turbine blades. Okay. All right. This is like
as real as it gets, these coatings are subject to harsh conditions, erosion, steam oxidation.
All right. So we want to develop a better coating for a steam turbine blade. So what we did is we,
and it took about a year and a half, we sat down with Karim's team, knowledge representation
experts. We had about three or four material scientists, including the specialist who runs
the coder that was doing the coatings. And we built an ontology that describes this whole
project all the way from the variables that are used in the coating system to the microstructure
and how we describe it. That's developed all the way to the properties. In this case,
we're doing steam oxidation and erosion. We put that ontology into a knowledge graph where we
connected all the, all the pieces, right? So we know the connections between all of those,
all of those attributes that we defined in the ontology. And then we said, all right, great.
We got this thing in a knowledge graph. Now what can we do? So the first thing we did is we built
in analytical knowledge. This is kind of the easiest one to do. We got about, it was 10 or 12
equations that worked in this ontology that, that we wrote. So you know, the simplest one is like
deposition rate. If you don't know your deposition rate, but you know, your thickness and your
deposition time, well, you know, you can figure out your rate as an example of a very simple
equation that we connected into this thing. And then Kareem and his team built a natural
language processing interface so that a material scientist could go in and say,
you know, could ask specific questions on data that was contained in the knowledge graph. So like
for part ID number ABC one, two, three, what was the erosion rate, right? And it would come back
and give you an answer. But then you could also ask questions where the answer was not in the
knowledge graph, but equations that could piece together parts of the knowledge graph to get you
to an answer. And it could be two, three, four, five or more equations deep. The system automatically
can figure out that equation chaining and come up with an answer for you. So like there was an,
you could ask a question about what was the ion flux for this particular coding? And it would have
to chain together, you know, multiple equations to get you that answer. Cause it's not in,
you know, the data that we ingest into the knowledge graph. So it's, it's really been a
cool process and we've learned a lot along the way. One of the things we learned is expert knowledge.
It's there. We tried some things to encode it, but there's just a lot more work that needs to
be done in that area specifically. That is so rad. It's cool to see that from sort of beginning
to end sort of adaptation taking place. I think that is so rad. Yeah. Normally you'd just be like,
Oh, if I don't have this information, I guess I'll have to go run some experiments or I guess
I just don't have it and I'll have to do analysis without it. But even just having direction there,
because I mean, you know, nothing's going to ever supplant raw data, but if you can get a
very good guess or a good direction of what it might be, that's incredibly valuable.
Yeah. And one of the reasons that this happened is because we have folks like Kareem and his team who
are not material scientists, but they're experts in their field. They're not knowledge representation
and reasoning that field. And then you have the material science experts and, you know,
I can walk down the hall and we can get in the conference room together. So that, that's just
been really exciting for me. Kareem might have other thoughts, but for me, it's been really
exciting over the last year and a half to try to put this together and just, just learn.
I mean, all of this has happened because it's been a fantastic partnership with Andy and the
team. I mean, it's been a wonderful experience for myself and my team as well. It's been really
great. Well, as you've heard, GE is serious about materials informatics. I'm thrilled that they are
because you're a leader in this space and other companies are going to hear this podcast, I hope,
or hear about what you're doing and say, it's time for us to get serious about this too. So I
appreciate the leadership role. And obviously what you're doing is a big effort and it's going
to need people. So I assume you're hiring in the space. You mentioned that we can put a link in our
show notes, perhaps if people have, you know, if there's job openings that you want to fill now,
or if there's an informational source where that people could learn more about this, we would be
happy to point good people your way because you're doing awesome stuff. And it sounds like an awesome
career if you're asking me. Yeah, we are absolutely hiring and not just in materials informatics, but
across multiple materials labs. So we'll put those links in.
Okay. Anything else you want to tell us before we wrap?
No, I think this has just been a lot of fun. Taylor and Andrew, thanks for hosting us. It's
really been a pleasure to be here and chat with you guys. Yeah. Thanks so much for having us.
Pleasure. Okay. Well, special thanks to our guests, Andy and Karim. This was a blast to learn
more. I think you guys are doing amazing stuff. I can't wait to hear about what you do next,
especially this ontology stuff. Figure out how to get the knowledge from my brain, from Andrew's
brain, from Andy's brain into our ML models in an effective way. And holy cow, that's moving us
towards the general AI that we need. So can't wait to hear more about what you're doing there.
Okay. Until next time. So this is obviously just the first in a series of half a dozen or so
really cool episodes where we're going to be talking with GE. And this first was on materials
informatics, but stay tuned. We have future episodes on super alloys, on barrier coatings,
what else? On CMCs, on all sorts of really cool topics. So you don't want to miss them.
They'll be trickling out over this course of this year. As always, special thanks to the
sponsors of this show. Obviously this episode is sponsored by General Electric and we are
very grateful to them for not only being forward looking in terms of things like materials
informatics, but also for reaching out to us, a podcast about materials as a unique way to get
the word out about what I think is a really, really innovative company that I actually applied
to when I was a postdoc, right? Very cool place.
Super alloys, titanium aluminide, the introduction of ceramic matrix composites,
and the first 3D printed metal jet engine parts have helped propel commercial air travel beyond
the Wright brothers first 12 second flight at Kitty Hawk to some 100,000 flights happening
around the world every day. Materials innovation has always been at the core of what GE does
and central to the progress our products have driven. My name is Joe Vincicuera and I'm proud
to lead the Materials and Mechanical Systems Technology Organization at GE Research. Together,
we are an interdisciplinary team of aerospace, mechanical, materials, chemical, and manufacturing
engineers and scientists working to advance the state of the art for complex mechanical systems,
innovative system level designs, advanced materials, and revolutionary manufacturing methods.
Every day, our researchers explore the boundaries of cutting edge technologies
that are poised to change the world. Whether shaping the future of flight or aiding in the
transition to a zero carbon energy future, our team helps GE stay at the forefront of innovation,
enhancing our products, and delivering for our customers. If you're ready to see, move, and create
the future, consider joining our team at GE Research. To view our latest job openings, you can visit us
at www.gecareers.com to learn more. That link is also available in today's show notes and thank you
for listening to today's episode. We are also brought to you by Elsevier in the journal
Materials Today. You've heard us talk about this company before and this journal before. It is the
flagship journal for Elsevier. It is a phenomenal publication outlet. I regularly find myself
looking for subject matter experts, whether it's a killer review article, or a really interesting
perspective piece, or an in-depth actual research article. And regularly, Materials Today is right
there as a great resource for that stuff. So if you haven't considered publishing there or
checking their content out, I recommend it. But they're much more than the journal. Elsevier has
conferences, they have a whole, you know, community they've created around material science. So props
to them for that. Check out Materials Today. Thank you 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. Make sure you subscribe to the show on iTunes, Spotify, Google Play, 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. We'd also 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. Catch you next time. See ya!
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.