Controlling nuclear fusion experiments with AI, another transformative technology from DeepMind

Recently, the DeepMind team came up with another surprise. In collaboration with the Swiss Plasma Center (SPC), it has developed a new approach to plasma magnetic control based on deep reinforcement learning and applied it for the first time to real-world plasma control, a transformative technique that opens up new avenues for advancing fusion research.

It seems that DeepMind has gotten so far in developing AI that it has even taught you to control nuclear fusion. Yeah, that’s not a joke. Just like it sounds.

Clean energy will play an important role in the technology of the future. Companies are finally embracing sustainability, whether it’s through repairable laptops or sneakers with cute snakes on them.

But when it comes to the energy that underlies the use of virtually every technology you come into contact with, we need to think a little further ahead. Nuclear fusion, despite its scary name, has the potential to be a great source of clean energy in the future, and we’re one step closer to being able to control it easily.

Wired reports that Google-backed artificial intelligence company DeepMind has trained an AI to help control the plasma in nuclear fusion. Nuclear fusion is the fusion of atomic nuclei that releases enormous amounts of energy.

It’s what happens inside a star, and although it’s an incredibly hot plasma, it’s considered a good source of clean energy fed by seawater. It’s just a matter of controlling it, and thankfully putting it in the hands of an AI isn’t as scary as it sounds – at least we think it is. Or don’t we think that? Doesn’t anyone read Asimov’s, Scott’s, The Singularity, or anything else? Amazon recently mentioned something fishy in its terms of service. And now this?

How AI learned to control fusion

Physicists are using a device called a Tokamak to contain and control these reactions. It is shaped like a doughnut and uses magnetic coils that can be controlled to shape and control the plasma. It is not particularly risky for the scientists involved, but this magnetic field must keep the plasma away from the walls of the tokamak to avoid damaging the device and slowing down the reaction.

The process is very cumbersome, making it difficult to test new methods to increase cleanliness or energy from the process. An AI trained specifically for plasma control could be the solution.

First, DeepMind trained the AI in simulations where it could change settings and receive computerized feedback on how the plasma should react. It also tried different shapes and ran other tests.

Using a deep reinforcement learning system, it was able to independently figure out how to perform the tests and then replicate the same level of control in a real tokamak.

One of the most difficult aspects of controlling plasma is that it is constantly changing. This made it difficult to train the AI, but in the end, it seems to work well.

Ambrogio Fasoli, director of the Swiss Plasma Center at the École Polytechnique Fédérale in Lausanne, Switzerland, says the breakthrough is a “big step,” with the hope that it can be expanded into a much larger tokamak and start providing the clean energy we need for the high-tech PC games of the future. Can you imagine that? PCs powered by fusion energy? That sounds a lot like Fallout to me. And we all know how that turned out, don’t we?