Human decision making in computer optimization yields hot plasma in alt-fusion.

If there is one thing I hate, it's optimization. Computers don't actually understand what they are optimizing. And that creates problems for everything from bicycles to nuclear fusion.

The process goes something like this: you have a mathematical model of a bicycle. You want your bike to perform better, but there are so many things that can be changed, so you can't imagine finding the best configuration on your own. So you write a script. The script will vary the configuration of the bicycle and evaluate whether it is improved. After many iterations, you have the perfect bike, right?

No, you don't. What you didn't imagine was that the computer would remove the seat. Or that it would place, for no apparent reason, a third wheel between the (now removed) saddle and handlebars. Even worse, the stupid machine has got the chain passing through a bit of solid steel.

The problem you've just discovered is that it is hard to encode what "best" and "better" mean for all circumstances. When confronted with a problem like this—and many incomprehensible bike parts—my stubborn side comes out. Surely, I can programatically encode an all-encompassing definition of "best" and "better," right? One threatened divorce later, I give up the pursuit. At least until the next optimization problem sucks me in.

While my example is a little overwrought, the problem is very real. Many types of optimization involve far too many options for a person to tune the design, but they also have no good definition of improvement that allows a computer to do it for you. This was the problem facing scientists at Tri Alpha Energy and Google when they wanted to optimize their alternative fusion device.

I'll discuss the fusion bit later, but, in general, what you want to achieve in a fusion device is a plasma that is hot, stable, and has a particular density. If the plasma is too cold, it won't be energetic enough to allow fusion. If the density is too low, fusion will be inefficient. And, if the plasma is unstable, it will escape confinement, possibly sending large currents through the vessel walls.

Let humans decide

That may sound simple, but judging the quality of the plasma has a subjective element to it. For instance, if the temperature drops just a tiny bit but the stability is much improved—is that better, worse, or about the same? Humans make these decisions instinctively, while computers get confused.

In the case of Tri Alpha's fusion setup, though, things move rather slowly. Each run takes about eight minutes, along with a couple of minutes to setup parameters. So, given the plasma generated during the previous run and the plasma of the present run, humans have nearly 10 minutes to decide which is better. Of course, for a human operator, that's actually not enough time to decide, generate new parameters, and input the new operating parameters into the machine.

This inspired researchers to adopt what is referred to as the "optometrist's algorithm." The idea is that, after each shot, a human pushes one of three buttons: things are getting better, things are getting worse, and things are about the same (they're big buttons, since they have to hold all the text). In anticipation, the computer builds a decision tree and prepares three different sets of machine parameters. After a run of the reactor, the computer grabs all the sensor data and presents the operator with a visualization of the plasma. A quick decision and one button push later, and the computer dumps a new set of machine parameters into the control computers for the next shot.

This set-up sounds great, but there are still too many control parameters that can be played with for any reasonable optimization. Imagine a machine with some thousand-odd analog settings that can be adjusted—that gives you some idea of the scale of the task. However, the researchers realized that not all settings are independent; settings could be grouped such that optimization could be performed using 15-30 meta-parameters.

Even with that shortcut, optimization would still be too slow. The time between shots is too small for even a computer to calculate the parameters it needs for the next shot. Instead, the computer actually needs to be calculating some six to seven shots ahead. To accomplish this, the algorithm has a decision tree that holds or is in the process of calculating settings an hour ahead of the experimental schedule.

The researchers showed that this was a pretty efficient optimization technique. They improved plasma temperatures from around 400eV to above 600eV. But, more importantly, the plasma stays hotter for much longer, increasing from about 1.8ms to well over 3ms (the measurements cut off at 3ms). The effect of this is to increase the net heating from a few hundred kilowatts to some 4MW.

So, a nice result. But what exactly are the researchers optimizing?

Ctl-Alt-Fusion


I recently presented the stellarator and discussed fusion extensively there. The stellarator and the tokamak are the only two designs that survived from a multitude of concepts that were explored in the 1950s and '60s. There were good reasons some designs ended up on the discard pile: experiments failed (repeatedly), and calculations did not give any reason for optimism.

Some of these concepts are being reinvigorated now. Calculations that were unimaginable in the '60s are now being used to recheck old results. And modern engineering allows for more flexible experimental designs, as well as a better view of what the plasma is doing.

Tri Alpha Energy is in this game with what's called a colliding beam fusion reactor. As its name suggests, the basic premise is to accelerate two beams of particles at each other. Fusion will occur for some teeny-tiny fraction of the collisions (and only a tiny portion of the particles collide). At its most basic, this is an inefficient design.

The twist on the idea—the thing that raises the hopes of replacing some of those "tinys" with "larges"—is that, with the right set of magnetic fields, the collision between the two plasmas generates a kind of smoke-ring of plasma. The ring of plasma then has ample opportunity to collide and fuse if it is hot enough. The confined plasma is heated by injecting more hot particles, hopefully reaching temperatures that will allow fusion.

In a second difference from stellarators and tokamaks, Tri Alpha Energy is avoiding hydrogen-hydrogen fusion (or, more precisely deuterium-tritium fusion). Instead, they are attempting to fuse hydrogen with boron. This creates an unstable carbon atom plus some highly energetic radiation. Heat and electricity is generated by absorbing the radiation. The advantage of this reaction is that there are no neutrons, which means the plasma does not irradiate the wall vessels, which would create low-level, short-lived radioactive elements.

Except, that's not quite true. The carbon atom decays to three alpha particles (a helium nucleus). One of the alpha particles can then fuse with boron to make nitrogen and release, yes, a neutron. So, yes, fewer neutrons, but, probably not no neutrons.

That said, Tri Alpha Energy's concept reactor is a very good piece of engineering. The company's experimental work is really nice. And its researchers worked hard to come up with plausible designs that answer many of the deficiencies in the original concept. I hope to see and hear more from the in the near future.