Fusion and AI Technology

Fusion and AI Technology

Along with the previously mentioned micro-nuclear power generation, the fusion of nuclear fusion technology and AI technology has become a hot topic in recent years.

The following article appeared in ZDNET Japan on 2020.02.18. “UK-based DeepMind, an Alphabet subsidiary, announced on February 16 local time that it has developed an artificial intelligence (AI) that it hopes will lead to the realization of a magnetic control device to maintain stable plasma conditions that are unstable at ultra-high temperatures. This is a new step toward the development of fusion power generation technology.

The device to which this AI has been applied is called a tokamak-type device, in which a group of high-power magnetic coils are arranged in a doughnut shape to confine and control plasma whose temperature is equivalent to that of the central core of the sun. By stably confining the plasma, molecular hydrogen can be fused with each other. Fusion is being investigated as a sustainable way to generate electricity.

In a new paper published in Nature, researchers from DeepMind and the Swiss Plasma Center (SPC) at the Swiss Federal Institute of Technology Lausanne (EPFL) detail the development of a suite of DeepMind AI algorithms that can control the shape of the plasma inside a containment vessel.

SPC has a vacuum vessel to configure a “Variable Configuration Tokamak” (TCV) to test the confinement of the plasma for fusion; SPC is able to determine the appropriate values for each variable (such as the voltage to be applied) for the control system to confine the plasma. SPC needed a reliable method to reliably determine the appropriate value of each variable (e.g., applied voltage) for the control system to confine the plasma.

According to SPC, it already has a simulator that provides sufficient insight, but it is still necessary to perform time-consuming computational processes in determining the appropriate values for each variable in the control system.

Setting the right values in the control system will confine the plasma and prevent it from colliding with the walls of the vacuum vessel that makes up the TCV and destroying its condition.

The researchers note that they have created “an innovative architecture in the design of tokamak magnetic controllers that autonomously learns how to control all control coil groups. This architecture reduces the effort required to create new plasma configurations, they said.

According to the SPC blog, DeepMind’s AI, trained on the simulator, can generate and maintain specific plasma configurations. These include “more advanced configurations that create shapes such as negative triangles and ‘snowflakes’ along with traditional elongated shapes,” as detailed in the paper.

DeepMind and SPC have also successfully run the algorithm on a real TCV owned by SPC. They also demonstrated a setup that maintains a stable “droplet” of two independent plasmas within the containment vessel.

While DeepMind has not published the results of this research at this time, it will eventually be publicly linked.」

The Nature paper listed above is “Magnetic control of tokamak plasmas through deep reinforcement learning”; Abstruct says, “Fusion using magnetic confinement, especially tokamak type, is a promising path to sustainable energy, especially the tokamak type, as a promising path to sustainable energy. The central challenge is to form and sustain a high-temperature plasma in the tokamak vessel. This requires high-dimensional, high-frequency closed-loop control using magnetic actuator coils, which further complicates the requirements to accommodate various plasma shapes. We present an unpublished architecture for autonomously learning control coils in the design of tokamak magnetic controllers. This architecture satisfies the physical and operational constraints while meeting specified control goals at a high level. This approach offers unprecedented flexibility and versatility and can significantly reduce the design effort required to create new plasma geometries. We have successfully generated and controlled a wide variety of plasma configurations in tokamakà configuration variables1,2 ranging from elongated conventional shapes to sophisticated shapes such as negative triangles and snowflakes. Our approach has successfully tracked the position, current, and shape of these configurations precisely. We have also demonstrated a “droplet” in TCV that sustains two plasma configurations simultaneously. This is a remarkable advance in tokamak feedback control and demonstrates the potential of reinforcement learning to accelerate research in the fusion regime, as well as one of the most challenging real-world systems where reinforcement learning has been applied.” The first is the “plasma”.

Plasma is a state in which the molecules that make up a gas are ionized, split into cations and electrons, and set in motion. One of the most famous plasmas in practical use is plasma CVD, which is used to deposit semiconductors used in ICs, solar cells, and other devices. In this process, the semiconductor material gas (such as SiH4 gas for Si-based materials) is brought to a low pressure, passed through a chamber, and decomposed by applying high frequency (commonly 13.56 MHz, the ISM (industrial frequency) used in microwave ovens, etc.) at the electrodes.

In this case, the plasma itself does not have much temperature, as the low-pitched plasma is mostly composed of electrically neutral molecules. Plasma stability can also be ensured by optimizing the impedance of the electrodes (high-frequency electrodes are often coil-shaped, so they can be adjusted with a variable capacitor).

In nuclear fusion, on the other hand, the plasma must be completely ionized into ions and electrons to reach ultra-high temperatures, so a special electromagnetic field (called a tokamak, which is a donut-shaped vacuum vessel with coils wound around it) is used.

The challenge is that the electromagnetic field generated by the plasma confined by the magnetic field generated by the tokamak, which is fully ionized and becomes a flow of ions and electrons, interferes with the confining magnetic field of the torus, causing the plasma to be unstable. Therefore, the technical miso of tokamaks is to control not only one coil but also multiple coils (TF coil (to group plasma into a doughnut shape), PF coil (to push plasma inward), CS coil (to induce plasma current in toroidal direction), etc.) to create stable conditions The technical miso is probably to create a stable state.

In these papers, it seems that the parameters were learned by using the reinforcement learning technique described in this blog to stabilize this plasma. The configuration of the system is presented as follows.

Basically, a model generated by learning simulation data is used to control the plasma by feeding back the impedance and current of the coils or the values of sensors installed in the furnace (optical sensors to measure the shape and temperature of the plasma?) The model is then controlled by feeding back the impedance and current of the coils, or the values of sensors installed in the furnace (optical sensors measuring plasma shape and temperature?

Just considering sensing methods, there are many possible approaches, such as using a general CCD to capture the shape of the plasma and extracting features through image processing, or using an imaging device with a slightly wider bandwidth to analyze the spectrum of the plasma, or installing multiple antennas to measure the electromagnetic wave patterns emitted, etc. Various approaches are possible. There are also various approaches to learning, such as reinforcement learning by DeepMind, which uses different hypotheses for the model, and sparse modeling and sub-modular optimization for multi-variable analysis. In addition, to stabilize the changing plasma, various techniques such as real-time sensor data processing and online learning can be applied to this very interesting task.