The research group has taught AI to magnetically generate high-power plasma jets used in fusion research-but wait! Clean up the EMP and driver. This is definitely a good thing, not a terrible weapon to use against humanity in the upcoming Robocalypse.
This project is a collaboration between Google’s DeepMind and École Polytechnique Fédérale de Lausanne (EPFL). This began a few years before the former AI researcher and the latter fusion researcher met at a hackathon in London. EPFL’s Federico Felici explained the problems his lab had with tokamak plasma maintenance.
Such a general complaint! However, it struck a chord with DeepMind, and the two set about to work.
Fusion research is done in a variety of ways, but all contain plasmas formed at incredibly high temperatures (hundreds of millions of degrees). Sounds dangerous, but tokamak is a way to control it and allow you to closely observe the fusion activity taking place within it. This is basically a torus or donut, where the superheated plasma travels in a circle and its path is carefully restricted by a magnetic field.
To be clear, this isn`t a fusion reactor of the sort you listen approximately giving almost limitless easy energy; it doesn`t produce energy, and if it unexpectedly started, you wouldn`t need to be everywhere nearby. It`s a studies device for checking out and watching how those risky however promising approaches may be managed and used for good.
In particular, the Swiss Plasma Center’s “variable configuration” tokamak not only allows the trapping of the plasma ring, but also allows researchers to control its shape and path. By adjusting the magnetic parameters thousands of times per second, all factors that can affect the properties of the ring can be made wider, thinner, denser or more diffuse.
Of course, the exact setting of the machine’s magnetic field must be determined in advance. Insufficient improvisation can cause serious damage. The settings are configured using a powerful tokamak and plasma simulator that the team has been updating over the years.
The team trained the machine learning system to first predict the plasma pattern produced by a particular set of settings, then work backwards from the desired plasma pattern to identify the settings that would produce it.
This architecture meets specified control goals at a high level while meeting physical and operational constraints. This approach demonstrates unprecedented flexibility and generality in the specifications in question, significantly reducing engineering effort to create new plasma configurations. We successfully manufacture and control a diverse set of plasma profiles on the variable-profile Tokamak, including oblong and regular shapes, as well as advanced configurations, such as negative triangle configurations. and configure “snowflake”.
And here are some examples of different shapes and configurations the model was able to produce:
This is an important task. Because experimenting with such a plasma requires a great many (possibly millions) small adjustments, not to mention using the plasma to power it, and configure everything manually. Because you can’t do that.
If a theory requires two threads, one is 22% larger than the other, it can take weeks or months to figure out the theoretical parameters to generate those theoretical parameters using “traditional” methods (which are obviously already extremely complex numerical simulations.
It can also be crucial for safety, as no human being able to improvise the settings for a second or two can prevent timing anomalies. But the AI can change settings in real time to avoid damage.
