AI is used to control the nuclear fusion created by DeepMind and EPFL

DeepMind, a company (owned by Google) specializingArtificial intelligence, Is currently collaborating with researchers at the Swiss Federal Institute of Technology Lausanne (EPFL) at the Swiss Plasma Center (CPS). In an article published in NaturalThey explain how a synthetic intelligence could simplify the process of controlling plasma as a result of nuclear fusion.

The artificial intelligence in question has actually shown the ability to enter plasma into a tocama and control it to take it in different configurations. If the word “Tokamak” reminds you of a kind of horn, you may need a little explanation in this regard. In fact, it is the donut-shaped structure used in experimental research on nuclear fusion containing plasma.

Nuclear fusion is being actively studied by scientists, during which two “light” atoms collide at high speeds to merge. This phenomenon releases a large amount of energy, without the drawbacks associated today with radioactive waste and the nuclear runway hazard. It actually differs from nuclear fission in terms of the ability of atoms (mainly uranium) to split in two, a phenomenon currently used in our nuclear power plants.

Therefore nuclear fusion releases a very large amount of energy (more important than fission). To give an idea, the same phenomenon that works in our sun or other stars! When scientists create nuclear fusion, they obtain what is known as the “plasma” of hydrogen, which generally reaches higher temperatures than the Sun. Controlling this energy is not as easy as you might expect. However, to extract any energy from it and use it firmly one day, it is necessary to maintain this plasma.

If one cannot maintain it, the reaction cannot be maintained either. In a star, it (reaction) occurs naturally, it is the force of gravity, which takes the place of the “straight jacket” for energy. On Earth, to control it, scientists use coils to create a magnetic field. However, this popular plasma is highly unstable, and its state is constantly changing: so scientists need to perform more complex calculations and simulations to learn how to adjust the magnetic field. ” For example, a control system integrates several magnetic coils of Tokamak and adjusts their voltage thousands of times per second without touching the walls of the plasma vessel, causing heat loss and damage. “, To explain Deep mind.

AI “sculpting” plasma

Despite many years of research, all of these adjustments and calculations are complex and time consuming. In addition, as DeepMind explains, scientists have limited access to research structures such as tokamaks, and many are willing to conduct experiments.

” Our simulator is based on more than 20 years of research and is constantly updated Federico Felici, CPS scientist and co-author of the study, explains. A statement From EPFL. ” However, lengthy calculations are still needed to determine the exact value of each variable in the control system. This is where our collaborative research project with DeepMind comes into play. “.

The team actually used an artificial intelligence to facilitate this whole computational process. During the experiments, the plasma contained and was able to “sculpt” 19 magnetic coils, which gave it different shapes, giving it different shapes. The researchers based their use of the neural network on a “reinforcement learning” basis. As a reminder, a neural network is a system inspired by the activity of biological neurons, but is well suited to statistical methods. Artificial intelligence “feeds” data to extract logical connections, and processes them to deliver results.

Artificial intelligence therefore began with “learning” by observing how the systems of the coils affect the shape and properties of the plasma. Scientists were then able to set goals by telling her to recreate those patterns herself. Plasma was first contained by conventional methods and then handed over to good care of the AI. She showed the ability to deal with it in both simulation and real situations. No adjustment was made between the simulation and the direct test. It was possible for them to set goals about the shape of the plasma, its consistency, or its position in the tank.

Focus on the goal to be achieved

According to scientists, this approach Enables the intuitive system of performance goals, emphasizing what needs to be achieved, rather than how to achieve it “It therefore greatly simplifies calculations. It also allows the use of a single controller for all coils, instead of one for one coil.

AI, for example, was able to recreate the “snowflake” structure, the shape of which allows for better distribution of the energy released into the tank. A “drop” formation, i.e. the isolation of two different plasmas in a single tank or even a “D” shape, was planned to be created for use on a large scale Tokamak under construction in France, ITER. In fact, this application of AI should be more effective in large furnaces because the calculations can be more complicated.

“ This is a promising new direction for plasma controller design that will accelerate fusion science, explore new structures and enable future Tokamak development. “, We welcome the researchers in the study.

Source: Natural