I'm Laurence Maroney.

I'm reporting here from the tensorflow summit.

We're in the Tensorflow Cafe on.

It's a great pleasure to be chatting with Ian Lang More.

Who's a software engineer on?

He works with tensorflow in nuclear physics in welcome.

Hey, Well, thanks for having me on.

So tell us all about it.

What What do you do?

There's a goal of powering of getting power from nuclear fusion.

Right.

So it's incredibly energy dense.

What you do is you take, uh maybe I could borrow a couple of these coffee grinds.

You put together two small nuclei And what you get out ways a little bit less than what you put in.

And the remainder is a huge amount of energy conversion, amassed energy, and I get E equals M C square.

Exactly.

Yeah, And so it's so incredibly energy dense that to provide the power that humanity would need over the next 100 years would require, say, boron, that the world currently produces for just seven months.

So say, set aside the next seven months of boron and we have enough energy for the next 100 years and no greenhouse gas emissions.

That would be nice.

That would be nice.

And it's also challenging.

So obviously anything that has that high of a payoff.

People have been trying and they've been trying for about 70 years.

And what happens is the plasma gets unstable.

So so the reaction happens inside of this plasma, which is, say, a swarm of charged particles at 10 million degrees or hotter, and they don't want to stay in place.

Right, so that way get excited.

Yeah, things get a little hot, they want to get out.

And what happens is things become unstable.

They get out and then it cools down in the reaction stops, and we end up putting in more energy just to keep the reaction going when we get out.

That's what's been happening for 70 years.

Okay, so there's this company in Foothill Ranch Ta Technologies that is now on their fifth generation plasma generator way.

We've been working with them since 2015 to try to accelerate progress.

So if this ends up being this gigantic machine, it's like think of this long cigar where there's a hot plasma in the middle, okay, and the goal is to use the magnets and neutral beams and all these other technologies to keep that plasma confines.

So this is an experimental reactor.

It's I mean, it's it's not.

The fusion reaction is not what's happening.

What's your head?

What's happening here is you're seeing how can tightly we confined this plasma as we ramp up the temperature.

And so they're doing experiments every 20 minutes.

And then every 20 minutes they pushed down data way.

Pull it in and we give back a three dimensional image off what their plasma looks like.

Okay, nice.

You might think.

Well, why can't I just look at the plasma, you know?

And the point is, First of all, it's well, it is very thin, and it's incredibly hot.

And, you know, you could consider, like, Why can't I just look at a light bulb and tell you what's inside of it?

Well, you only have a two dimensional view.

You're on the boundary, you get a two dimensional view of what goes on.

But you want to look at a three dimensional objects, and moreover, it's so hot that you can only have a limited number of use into its.

So what you end up with things are like you shine 14 lasers through the center and you measure the phase shift of those lasers.

Okay?

And this gives you the density along each of the lines.

But from that, you want to reconstruct the density everywhere.

So it's a very under determined problem.

So I've gone on for a little.

This is great.

So it's helping clarified for me.

Hopefully for you, too.

So but the idea then is that so?

This thing, Fusion reactor.

Every 20 minutes you're doing an experiment.

You're pulling that into code that you've written, intensive low on.

Then that code is generating this treaty image, right?

Yeah.

Now, that's not the typical use of tensorflow.

No, no, not at all.

So this really isn't deep learning.

This is an inverse problem.

This is great.

So it just shows the flexibility of a real key difference is in deep learning.

Often you have many, many, many examples, many, many labels.

So you have the input output, and you could learn a very complex, functional relationship.

So here we actually don't have the labels that the labels would be.

This is what the actual plasma looks like.

But we're the ones telling people with a plasma.

Looks like we don't have those.

We do have a precise model for how the measurement works.

And so from that it's you gonna wait?

Think the naive thing would be just to invert the equation.

So if this is what the plasma waas, this is what the measurement is.

Well, if we know the measurement, let's invert the equation, Get the plasma.

Problem is, there's many plasmas that could have given the same measurement.

So we're gonna give them a distribution over possible plasmas suing Obey ze.

An inverse problem.

Okay, in our graph is modeling, measurement, physics and some physical assumptions rather than an arbitrary function.

So then the output of this does it become a situation where, because of running experiment every 20 minutes, that they could now optimize their future experiments based on the results that you're giving them well, they can understand their experiments.

Based on these results, there's actually ah, coincide and efforts by Googlers to help optimize experimental design.

And there was a paper published in that Ted Vaults is was one of the names Michael DeKosky.

We'll put a link to the paper and descriptions.

So then people who read it for themselves.

Fascinating stuff.

It's like, How did you get involved in?

Well, so I have been a Google for about four years, and I was looking around internally for teams that we're doing something with reducing greenhouse gas emissions.

And so then this popped up and I was like, Oh, this is great.

And then it's an inverse problem, which is what I did for my PhD and my post, doc.

So I was like, Okay, this is like a perfect fit when I started working with Nice Is your role like coding?

The way so I wrote is writing a code and also coming up with statistical models.

So I've been writing code specific for our team and then also the same time ever writing code for the tensorflow distributions in tensorflow Probability libraries.

So these air sort of form the core building blocks like we have a normal distribution object or the lacrosse distribution and so on and so forth.

And these are it's an object that allows you to you can produce samples, you get a PdF and so on, and then we have a method of transforming one distribution to another.

With this thing called the ejector.

So way actually did an episode of coffee with on the distribution of a P I, which is Yeah.

So check it out on the Google Developers Channel.

It was it was mind bending because just the kind of things that it offers to you as an A p I baj actors and all that kind of thing.

So it's like as a developer, go check that out.

Distributions a p I maybe they'll be able to build something like you built for the nuclear fusion.

Thank you so much in this has been a blast.

It's all right.

I never expected to come here to learn a little bit about your fusion today.

So it's like a very pleasant surprise.

Thank you so much.

If you've any questions for me or have any questions for Ian, just please leave them in the comments below.

And don't forget to hit that subscribe.

Thank you so much.