Before we dive into Materials Informatics details, we should talk about what exactly Materials Informatics is and why materials researchers should care about it. In this video, we discuss how Materials Informatics can help identify structure-property-processing-performance linkages in ways that humans would not otherwise be able to do. I describe the history of the Materials Genome Initiative, a movement from domain experts to deep learning in natural language processing, and examples of machine learning successes such as AlphaFold, Al 7A77, and others.
Hey everybody, I'm Taylor Sparks.
Welcome again to my class on materials informatics.
Now today, for today's video,
we're gonna talk about
what exactly is materials informatics?
Why should we care about it?
Why are we doing this class in the first place?
So to answer that question,
we need to dive into what is the heart of material science
and the core, you've seen this before.
You're familiar with the Tetrahedron probably, right?
Material science is all about the study
and investigation of these structure,
property, processing relationships, right?
That if you know what property you want
to get a certain performance,
well then you know how you need to process the material
to get that structure, right?
How these things are all tied together.
Understanding these linkages is the core component
of material science and engineering.
So with that said,
some of these relationships we've learned over the years
and they're relatively simple.
For example, how about Hall-Petch, right?
You've probably seen Hall-Petch relationship
in your introductory material science course.
It showed something like hardness or strength
as a function of grain size.
In fact, they usually do grain size
to the negative one half power
because they found through empirical relationships
that as you reduce the grain size,
things get stronger and harder, right?
So to the right, these would be smaller grain sizes
since it's the negative one half power,
but that only reaches to a certain point.
And after that point,
basically you lose your benefit, right?
But this relationship, this sort of Hall-Petch relationship
has been a really important hallmark
of material science and engineering.
So that said, why do we need data science to do that
if we can do that on our own, right?
Well, data science allows us to do this
with much more complicated relationships.
Instead of these really simple ones
where it's just grain size,
for example, it's playing relationship,
maybe it's many different relationships
and maybe it's not a linear relationship.
Maybe it's nonlinear and complex, right?
Those are things that are gonna be really hard
for us to capture with our human intellect,
but those are things that we can capture quite easily
through data science, right?
So let's consider a few examples, right?
Data science is already doing some amazing things, right?
If you're like me and you follow machine learning
in the news, you'll see that there are crazy headlines
constantly popping up.
Take a look at this one, right?
It says, a new artificial intelligence can guess
whether or not you are gay or straight from your photograph.
How crazy is this?
So the idea was this, you took dating websites
which have people's faces, right?
And then it has their sexual preference.
And so now all of a sudden
you can build a classification algorithm
which takes the face as an input.
It does the machine learning magic
which we'll talk about this semester
and outcomes of prediction that they're straight or gay
or whatever they're predicting.
So in this initial study,
they found that it was 91% accurate.
Imagine this has crazy ethical concerns, right?
Now there've been some follow-up studies, right?
So Stanford, Wayne Kaczynski did a follow-up
and they found that it's actually 81 to 85% accurate
for males and a little bit less
for females, 70 to 71% accurate.
And then what's interesting, they started realizing,
well, it's not as if they're like recognizing it,
you know, cheeks or eyes or freckles or whatever
lead to this being one or the other,
but rather these algorithms were really good
at catching differences in the way
that we present ourselves.
So for example, the way that we have our facial hair,
mustaches or beard, or you know,
the way that we do our eyebrows
or the way that we put makeup on our earrings, right?
It was paying attention to those things
in much the same way that humans do
but doing it in a really quite accurate way,
far outperforming human people.
So here's one example of capturing really
what ought to be complex trends from data
and doing it quite accurately.
But what about materials, right?
This is a really cool one.
This happened just last November, right?
So when I was a kid in high school,
there were these softwares that you could download.
They were screensavers,
because back then screensavers were a thing
that I don't think anybody has those nowadays.
But back then, if you let your screen sit,
there were concerns that it would damage it
because we use different technology
than what we're currently using.
And it basically, if you had the same image on your screen
for too long, it would be damaging.
So instead, you change the image periodically.
And one way that they came up with
was to do protein folding with your computer.
What a clever idea, right?
Because protein folding,
figuring out these big complex protein molecules,
they can come together.
They have all these different functional groups
that can interact, you know, bonding and light,
you know, strong versus weak bonding all over the place.
It's really hard to predict the actual structure
that these things will have.
And so they had this really clever idea.
They said, all right,
there's all these computers across the world
being used sometimes,
but mostly sitting there with screensavers running.
What if instead of running a screensaver,
you had a special screensaver
that was actually gonna be doing
the protein folding simulations?
And so this is a way to farm out the calculation of this,
what seemed this impossibly hard problem
to lots and lots of people.
So that was one approach, you know,
and they've tried lots of things over the years.
As this headline says here,
AlphaFold, this is a solution
to a 50-year-old grand challenge in biology.
So this is something that's been worked on for a long time.
And lo and behold, AlphaFold has now come up
with a solution to this,
the company's DeepMind that came up
with a solution to this using,
you guessed it, machine learning, right?
So here's a quote from one of the founders.
He said, we'd been stuck on this one problem,
how do proteins fold up for nearly 50 years?
And then to see DeepMind produce a solution for this,
having personally worked on this problem for so long
and after so many starts and stops,
wondering if we would ever get there,
this is a very special moment, right?
Pretty incredible how accurate it was able
to make these predictions using machine learning.
Rather than using explicit calculations
of all the bonding interactions,
it just used machine learning to make these predictions
and solve this incredible problem, right?
So that's getting closer to material science, right?
Proteins, let's get a little bit closer still.
If you're like me and you love 3D printing,
you can see one going in my background over here,
3D printing is so rad,
and yet you realize maybe that 3D printing
is actually a little bit tricky
and it doesn't always give you exactly
what you're looking for.
For example, when you're 3D printing metal alloys
in whatever technique you're using,
you'll realize that many times
you start with these really fine nanoparticles,
but when you do the actual densification
and printing of them,
they become prone to what's called hot cracking, right?
So this degrades the mechanical properties,
you get cracks forming during the solidification
of the 3D printed metals, it's just not great.
However, we have come up with a solution to this.
The first ever 3D printable aluminum alloy
was developed by a team,
and the team used, so it was based out of HRL,
that was where the lead researchers came from,
that's formerly Hughes Research Laboratories,
as in Howard Hughes, the airplane manufacturer.
Anyways, it's gone through some changes,
now it's HRL Laboratories.
And I love this, one of the researchers,
Brendan Yehada says,
we went from a haystack
to a handful of possible needles, right?
Because finding the exact composition
was like finding a needle in a haystack,
there's so many different alloy compositions to pick from,
how are you gonna know which one
will give you the properties of interest, right?
The idea here is that you can use machine learning
to help you down, so like from the enormous number
to a small number of reasonable possible needles,
like it says here, right?
So what they did is they actually used this,
they partnered with Citrine, a great company,
who's actually, shirt I'm wrapping today,
they're the ones who got me involved
in materials informatics many years ago.
They combined classical nucleation theory,
so theory about material science
with some rules on lattice spacing,
with thermodynamic stability, with density,
materials informatics,
and what they were ultimately able to do
is search through over 11 million combinations
of powders and nanoparticles
to end up at the one that is now
the very first 3D printable aluminum alloy.
It's actually been, it's got its own grade, right?
You have to register it with the aluminum,
I forget the organization,
but it is now aluminum 7A77,
which is a 3D printable aluminum alloy.
More and more case studies like this,
which are just remarkable achievements,
are being made possible by artificial intelligence,
machine learning, materials informatics, right?
So where did this all start from?
It's only a couple decades old,
that's what's remarkable, right?
There were other fields that were using big data
and machine learning long before material science.
Material science is late to the game.
It really got kicked off in the early 2000s.
For example, one of the big first articles
that came out was from Krishna Rajan in 2005,
the Materials Today article,
where he laid out some of the framework
of what this could potentially achieve.
What is this new paradigm?
What will it require?
What could it potentially do for us?
This was a, it's a good article, it's worth reading.
Krishna Rajan is a forward thinker.
In fact, he actually, at the time,
he was at Iowa State, I think, when he wrote this,
but he left and went to University of Buffalo
and founded a new material science department,
I think it's called Materials Design and Innovation,
and it's centered on materials informatics, right?
Everything they do is infused with data science
put into the material science program,
so this was one of the early pioneers in this area.
But I would say that in the early days
of materials informatics, it was pretty rough.
There's a psychology professor at,
I think he's at Duke, who I just love
a lot of things he has to say.
He had this tweet back in 2013.
He says, big data's like teenage sex.
Everyone's talking about it.
Nobody really knows how to do it.
Everyone thinks everyone else is doing it,
and so everyone claims that they are doing it.
And that was definitely the case
with materials informatics in the early days.
And I'm no exception, right?
When I think of some of the work
that I was doing early on, we got onto this trend
of wanting to look at big data really early on.
We were in 2012 when we started doing this.
And what we were doing, what we thought was big data,
was really just writing very analytical review papers, right?
Something that was really great is we were using
data from literature to assemble these large databases,
and then we were able to generate these sort of websites
that would allow you to visualize data
in really unique ways.
Let me show you an example of one.
So, for example, right here, you can see on this website,
which is still open and running, you can check it out.
If you go to, you can select your different parameters.
Whatever you want to plot against one another
on the X and the Y, you can select
what you want the marker size to scale with.
And then you can generate these plots,
and you can even sort your data in different ways.
And in the end, you're able to generate
really cool bird's eye view plots of different fields,
and you could hover over the points.
If you clicked on it, it would take you to the article
where we took the data from,
which is pretty exciting, right?
So this was a cool way to, I'd say,
visualize and organize data,
but there's no artificial intelligence here.
There's no, there's very limited data science to this.
So I'd say that it was a good first showing,
but there's a lot more that could be done.
And so, in the years that followed,
we really got much more into understanding
what data science can do for us.
And that's important because, as a field, we need to.
Just take a look at this, right?
Why do we need materials informatics?
Just take a look at this, right?
This is showing the number of publications
by keyword as a function of years.
Now, this is a log scale on the y-axis,
so this is log linear plotting.
And take a look at battery, magnetism, crystal growth,
thermoelectrics, polymorphism, metal organic frameworks.
All of these are basically linear in this plot,
which means that it's actually exponential growth.
And just ask yourself, how many,
and I guarantee that your field looks very similar.
There's this explosion of research happening right now.
And what are you doing with all of that data?
Are you actually reading 10,000 papers a year?
Because that's about how many are being published
in most of these fields of material science right now.
And if you're not reading 10,000 papers a year,
which you're not, if you're like me,
you're reading maybe a couple hundred,
maybe a thousand if you're going nuts, you're skimming.
Then what are we doing with all that information?
A lot of work and effort went into generating that data.
Are we just gonna toss it out?
That seems crazy, right?
No, we need to take advantage of it.
And to take advantage of it,
we need to rely on computers to help us
because it's just intractable right now.
It's way too much data for us to handle on our own, right?
Now here's another problem.
Just throwing more and more researchers at a problem,
hiring more scientists, making more universities,
hiring more research institutes,
isn't necessarily going to fix our problem.
Take a look at this example
from the field of thermoelectrics,
which is near and dear to my heart, right?
So the figure of merit ZT is the one that really scares,
scales with efficiency.
And yet take a look at this.
Over time, ZT had this sort of like roof
of like one for many decades.
And then in the early 2000s,
we started to eventually break above it.
But on average, as more and more researchers got into it,
the average ZT from the data that we collected
was actually going slightly down.
It was a slightly negative trend.
So just because you got more and more researchers involved
didn't necessarily mean that you solved the problem
because most of the time,
the breakthroughs were the exception to the rules.
It was something unexpected.
It was unusual.
And we need a way to go about rationally
designing new materials to allow this to happen
without relying on the happenstance breakthrough, right?
Here's another one.
We can use, there is a ton of useful information
embedded in large data sets.
And you don't necessarily need
domain knowledge to extract it.
Here's a great example, right?
So if you're like me and you love science fiction,
I grew up and Star Trek was on TV,
although I was a Star Trek,
the next generation guy myself growing up.
But they had these devices.
They'd go to these alien planets
and all of a sudden they could just be talking
with one another.
And you know, how's that work?
There's no way that English is the language of the galaxy.
That's silly.
So instead they had what seemed kind of like
this equally silly object,
a little object that was their universal decoder
they could talk into.
And they would just hear the language
being translated in real time.
So that was the science fiction dream.
And then came along reality,
which were these little handheld devices,
which were atrocious.
You'd sort of like slowly type in the text
and then it would translate it,
it would get it way wrong,
it would miss context,
it would have, you know,
it didn't know what to do with homonyms
and different meanings.
It was just problematic.
And then nowadays,
we have incredible things like Google Translate,
where now with the Google earbuds,
I don't know if you've tried these out,
but while you're wearing them,
you can have it switch into translate mode
and it will translate in real time.
So if somebody's speaking French to you,
you will hear English, right?
Which is incredible.
We're getting closer and closer
to the science fiction dream
of machine translation being just universal.
Now, why do I bring this example up?
The reason I like it,
because in the early days of this,
they relied on domain knowledge.
They hired linguists.
They hired grammar experts and said,
all right, if you wanna translate this phrase
to this phrase in this other language,
here's all the grammar rules that you need to follow.
And they were basically doing it by hand.
They would programmatically come up
with the steps to do it.
And then there's this pioneer
in the field of natural language processing.
His name is Fredrik Jelinek, right?
He died in 2010.
This Czech American researcher, fascinating guy.
He has this often quoted statement,
I'm gonna tell you today,
which was that when he was working
on natural language processing,
trying to make this process work better,
he once said the following.
He said, every time that I fire a linguist,
the performance of the speech recognizer goes up.
So in other words,
when you fire somebody with domain knowledge
and just take them out
and instead replace them with a data scientist,
a statistician, you can do better.
You don't necessarily need domain knowledge,
especially when you have lots and lots of examples.
At some point, you can just learn from those examples
without knowing the reasons behind them, right?
So can we do the same thing with material science?
That was the premise
behind the so-called Materials Genome Initiative.
And it's fitting that I talk about this today
because it is June, 2021,
and exactly 10 years ago, June, 2011,
we had the Materials Genome Initiative
for Global Competitive come out.
This came from the Obama White House
and the idea behind it,
it came from actually from a study
at the National Science and Technology Council, I think.
They had this thought that there's all this data
and other fields are really benefiting from using it.
Could we do the same thing in material science?
Or in other words, is there a materials gene, right,
that we could identify?
Just like we have human genes,
it gives me freckles, right?
And, you know, blue eyes.
Is there the same thing
that allows you to do the same thing,
but for materials that leads to high strength
or to ductility, right?
That was the idea behind this.
So they kicked it around,
they came up with some ideas.
And one of the main goals it had,
which we will often hear repeated,
is that it was going to help us develop new materials
twice as fast at a fraction of the cost, right?
Which would be amazing
because right now we develop materials,
you're probably aware of this,
it's a decades long process.
It is just painfully long.
So we wanna do much better
and that's gonna require a couple of things.
First off, one point of materials genome initiatives
was we're going to equip the next generation workforce.
We're gonna train new students
through things like YouTube videos
on materials informatics
with the tools that they need
to do materials informatics, right?
Second, is we're going to enable a new paradigm shift
in materials development.
I'll show that in the next slide.
Third, we're going to integrate experiments,
computation and theory
in a way like we've never done before.
And finally, we're going to facilitate access
to materials data.
These were the premises behind
the materials genome initiative,
which would feasibly help us discover materials much faster.
So what exactly is this paradigm shift?
Well, take a look here.
On the left, you're seeing the traditional approach
which shows how materials are discovered,
which basically you start with publications,
you look in the literature
and you basically find a system of interest.
You just like, oh, they discovered such and such
in this material.
And then you think to yourself,
well, what if we swapped out this atom for that one?
Or we changed the microstructure or the grain size
or whatever it is that material scientists do.
We come up with our design of experiments,
we synthesize, we characterize our materials,
we get some sort of performance or property,
we publish it and we just repeat.
And this goes over and over.
But what this leads to is,
I would call it local optimization,
meaning systems that get studied get really studied.
A bunch of people dogpile on those
because it's very unlikely for you
to pick something totally out of a hat
because the odds of success are pretty low.
Now let's shift gears and compare that
with a data science or materials informatics approach.
We're still gonna start with literature,
but now from this literature,
we're going to apply data mining.
We're gonna extract relevant information
from those studies in the literature.
We're going to create a database
or possibly a user website.
We're going to create machine learning algorithms
to help explore new chemistries.
From those, we can actually generate systems
that create, for example, recommended candidate rankings,
which now gives us a ranked list of interesting candidates.
We could then screen those using sustainability
or chemical intuition or feasibility of synthesis
or whatever you want.
We still have to make our experiment
where we're going to make these things,
fabricate them, measure their properties,
measure their performance, and then publish them.
But because this becomes a loop
and we start with these much better lists of candidates,
the odds of you getting successful outcomes is much higher,
which is exactly what we're after
in materials informatics.
This is the way that we're going to dramatically
accelerate materials discovery.
So obviously in the last few decades,
there's been this surge of interest.
It went from being less than 10 articles per year
to now being hundreds of articles a year
using my Scopus keyword for materials informatics.
And I'm sure if you used other keywords
like machine learning, there'd be even more.
And yet it grows and it's growing and it's growing.
So the question is, is materials informatics
a passing fad or is it here to stay?
Are you taking this class with,
is it going to be worth your while in five years
or will it not be?
Because take a look at the Google Trends
for machine learning in the bottom panel.
The trend rose and in 20, what, 18 or 19,
it peaked and now it's actually declining
as, for whatever reason, less interest,
maybe people don't, they're not buying
into the hype as much, but who knows.
Is that going to be the case with materials informatics
in the same way?
Are we going to see it peak as well?
Let's see.
Why might it peak?
Well, there's a company called Gartner and I love it.
They put out this Gartner Hype Cycle
for emerging technologies every year.
And new technology, as it comes on board,
they say that most technologies fit along this curve
where things have this innovation trigger
and then you have this peak of inflated expectations
and then you have this drop to the trough
of disillusionment where you realize,
you know, it got, they said it was going to do
all the things for us and it's actually hard
and it doesn't work very well.
And then you get this slope of enlightenment
and the plateau of productivity.
I think of like voice to text, right?
I remember talking to Siri in the early days
was like, whoa, it's going to do what?
It's going to schedule your hair appointments for you,
all this stuff.
And then in reality, like, oh, this isn't great.
And then over time, we talked to Google and Siri and Alexa
and all of these things all the time.
Now it's definitely reached the plateau of productivity.
Well, in any case, five years ago,
machine learning was the top of this chart.
They said it was going to do all the things for us.
Machine learning is going to, you know,
we wouldn't even need scientists in the laboratory.
We just have robots and machine learning.
And I think in the years that followed,
we definitely saw it reach its trough of disillusionment
while we realized, ah, this isn't working
quite as well as we thought.
But have we reached the plateau of productivity?
It predicted for the general machine learning
in two to five years, we'd get there.
Is that the case for material science?
I'd say yes.
Certainly seen some great examples.
I'm going to show you many more
over the course of this semester,
but that doesn't mean that we're done.
For example, this was five years ago,
the Gartner Hype Cycle.
Remember it said in two to five years,
it would reach the plateau of productivity.
Well, where's it at now then?
Are we just there?
If you look at the emerging technologies for last year,
I don't have it for this year yet.
Take a look at where machine learning and AI is now, right?
You've got explainable AI.
You've got embedded AI.
You have responsible AI, AI augmented development,
generative AI, composite AI, adaptive machine learning,
and on and on and on, right?
Two-way brain machine interference.
So many cool things are coming on board.
So I think what's happening is that we're realizing
that the basic vanilla application
of machine learning to material science
can do great things, but it's also limited.
But there's so much more that we can do with it.
And so that's why I think it's worthwhile
for you to take a course like this
to continue to expand what we can do
and what we will be able to do
with data science applied to materials research.
So in our next video,
I'm going to jump into this big question
of how are materials actually discovered, right?
Because I think that's an important thing
to compare how we're going to be doing it
versus how it's been done in the past.
So look forward to seeing you in the next video.