Hey, I'm Diana, and you're watching "Physics Girl."

There's this thing that you see happen.

You see it happen every single day.

Phase transitions.

Liquid to solid.

Solid to liquid.

Liquid to gas.

And scientists understand these phase changes.

We understand them.

But there are some phase transitions in nature that scientists don't fully understand because they involve quantum mechanics.

The most mysterious phase changes are called quantum phase transitions, and they happen at the extremes.

For example, if you get a magnet hot enough, you can get it to suddenly lose its magnetism.

Or in the coldest temperatures of, say, outer space, you could get insulating materials to suddenly turn into superconductors.

And so, yeah, we understand what's going on with the individual particles inside of, for example, water when water boils.

But we don't understand what the individual particles are doing inside a superconductor.

It's an exciting frontier of physics because it's a problem not yet solved by humans.

But what if something else could solve this problem?

There is something that's better at solving certain problems than humans are.

Something that could give us better product recommendations than we could give each other.

Something that could diagnose medical conditions better than doctors could.

That thing is a machine Yeah, it's a weird thought.

But what if I told you there's a technique that scientists and engineers are researching that can help us solve this quantum phase transition problem?

That technique is quantum machine learning.

This is my editor Jabril.

Hi.

But he also has an awesome YouTube channel about computer science.

And since I know nothing about machine learning, he's gonna help us-- me-- figure out what machine learning is.

Yeah?

Oh yeah.

Yeah.

Oh yeah.

So you want to know what machine learning is, right?

Yes.

So, simply put, it's teaching an algorithm to learn almost anything that you want it to learn.

An algorithm.

An algorithm.

So, for example-- OK, for example, let's create this fantasy scenario in which, like, you're a gardener who loves their plants, OK?

I have a garden!

Yes, that's right.

All those garden photos that you post.

OK, so let's use you as an example.

OK.

So you're a gardener.

You love your plants.

And let's say that you run into this problem where you notice every once in a while your plants die, OK?

It happens.

After a while, you deduce that the problem is either from one of two things, all right?

A, it's either from insects.

You know, ruining your plants.

Or B, dehydration.

Maybe you're not watering them good enough or something like that, right?

You love your garden, and your plants are mysteriously dying.

OK.

If you could find out their cause of death while they're dying, then you can treat them.

However, there's a problem.

The only way that you can find out the cause of death is if you pull them from the roots and snap them in half.

You're not a very good gardener, but-- I'm sorry, I'm sorry.

We'll roll with this scenario.

JABRIL: And so what you do then is you start taking thousands of photos of your dead plants.

So now you have this data set of a bunch of images and their cause of death, OK?

So you will then train an algorithm by inputting all of the photos you took matched with their labels.

You know, the insects.

Insects.

Dehydration.

So on and so forth.

And the algorithm learns the differences in features of images labeled as dehydration and the images label as insects.

Yeah, so the process in which the algorithm learns all this is actually really fascinating.

And I actually made a video over on my channel if you want to come and check it out.

Jabril, focus!

Sorry, I'm sorry.

But actually it is really good.

I'll put a link in the description to it.

Thank you.

You can now-- this is the exciting part.

You can now take a new photo that the algorithm's never even seen before.

Feed it to the algorithm, and it'll give you a likely probability on the cause of death.

OK.

So now you don't have to kill your plants anymore.

Problem solved.

Yes!

Thanks to the machine learning.

So machine learning is this process of training an algorithm on images that you input with labels to learn enough about the features of the images so that if you input a new image, it can look at that image's features and then figure out which of the categories it fits into.

Correct?

Kind of.

That's only one type of machine learning that we went over today.

There are more?

There are lots more, yes.

OK, cool.

That's awesome.

But now to bring it back full circle to quantum phase transitions, was it?

Yes!

Right.

We're gonna unpack more of that.

Thank you for your help, Jabril!

Yeah, any time.

Get outta here.

[LAUGHING] OK, so, how can we apply machine learning to quantum phase transitions?

I talked to a researcher at San Jose State University who works on this exact problem.

Yeah, so I study properties of quantum systems of few particles by doing simulations, numerical simulations of these systems.

When you mention quantum phase transition, I just want to make a distinction that those are strictly at zero temperature.

DIANA: So technically to be quantum phase transitions, they have to happen at absolute zero, the coldest theoretical temperature in the universe.

But we can't actually get anything down to that temperature besides in simulations.

And Ehsan's group does use simulations in their work, but even their simulations incorporate some thermal or heat energy.

You generate a large number of system configurations.

And by analyzing them, you can infer where a phase transition might be taking place.

DIANA: The phase transitions that Ehsan's group studies are magnetic phase transitions.

They simulate a number of electrons and then see what happens with different concentrations of electrons and what the magnetic fields of the electrons do.

These systems act differently just below a specific critical temperature, the phase transition temperature.

Above that temperature, they see no order in the system.

And below it, they get kind of a messy checkerboard pattern.

But it's hard to see that transition.

Something goes on in these configurations that are not possible to be seen by naked eye.

But what we realized was that we could design an artificial neural network that can distinguish these configurations at very high temperatures.

There is no order in my system.

From of course the other type of configurations at temperatures below the critical temperature where I have an order.

So this artificial neural network can look at these, and you can train it by showing it lots and lots of these images.

So, in conclusion, if we can figure out what's going on with the individual particles and transitions to superconductors, we could potentially make a room temperature superconductor.

That's a huge field of research right now.

Because if we made a room temperature superconductor, well, any electronic components like your laptop use or your phone, neither of them would get hot because there would be no resistance.

It's a superconductor.

So as the electricity goes through, none of the electronic components would heat up.

That would be amazing!

Probably far off in the distant future.

But it is incredibly exciting physics.

All right.

I hope you learned something fun in this video.

I know that I did.

All right, that's it.

Thank you so much for watching, and happy physicsing.