Construction of the ITER experimental fusion reactor in Cadarache (France) is expected to be completed in 2027. Meanwhile, smaller reactors such as the DIII-D reactor at the National Fusion Center in San Diego, USA, are being used to study how to optimize the plasma in ITER. It is published in Nature an intelligent control algorithm based on deep reinforcement learning to avoid plasma instabilities using the neoclassical tearing mode (NTM). The instability associated with the 2/1 poloidal/toroidal resonance was controlled in the DIII-D state analogous to the ITER reference scenario (IBS); In this state, a fusion power of 500 MW is produced in ITER with a gain of Q=10 for 300 seconds. These types of studies could accelerate the operation of ITER to successfully achieve its goals as soon as possible.
The advantage of AI control is that action is taken before instability occurs, rather than trying to mitigate its effect once it occurs. To do this, it is necessary to predict in real time when it will happen, which borders on impossibility using theoretical models (plasma simulations require supercomputers). Reinforcement learning reveals patterns in experimental diagnostic data (obtained by magnetic, Thomson scattering, and charge-exchange recombination spectroscopy, as shown in the figure at left) that can detect tear instability 2/1,300 milliseconds before it occurs. Using a feedback loop (image right) with a delay of 25 milliseconds, the actuators, which heat the plasma with beams of neutral atoms and radio frequency waves, are controlled using electron-cyclotron resonance to prevent the occurrence of instabilities (image right). center).
The NTM instability is one of the most important in tokamak operation; However, there are many others whose intelligent control will also need to be studied in DIII-D. In addition, the learning was based on historical experimental data from DIII-D, which will need to be replicated in other fusion reactors and with future ITER data. The new controller is a proof of concept and much work needs to be done, but all indications are that intelligent control will be used in all future commercial fusion reactors. The article is Jaemin Seo, SangKyeun Kim, …, Egemen Kolemen, “Avoiding fusion plasma tearing instability with deep learning”, Nature 626: 746-751 (21 February 2024), doi: https://doi.org/10.1038/ s41586-024-07024-9. On an informative level, I recommend the press release “Engineers Use AI to Argue Fusion Power for the Grid,” Princeton Plasma Physics Laboratory (PPPL), February 21, 2024.
Deuterium plasma in fusion reactors is subject to a large number of magnetohydrodynamic instabilities. In the case of ITER, it will take about ten years (between 2027 and 2037) to study them in detail before we can move on to tritium injection fusion studies. Many of these instabilities lead to disruption of the plasma, causing rapid energy loss and sudden termination of the discharge. This lost plasma energy is dissipated in the tokamak walls, coils, etc., which can cause damage. To minimize them, it is necessary to limit the maximum density, pressure and current of the plasma. Its intelligent control will allow these limits to be released, guaranteeing optimal plasma pressure and maximizing the production of useful energy.
In the new paper, intelligent control was studied experimentally using the DIII-D small fusion reactor, the largest in the US; This tokamak has a main radius of 1.67 m, a minor radius of 0.67 m, a maximum toroidal magnetic field of 2.2 T and a heating power of 23 MW. These values can be compared with the UK’s JET (2.96 m, 1.25 m, 3.45 T and 38 MW) or the future ITER (6.20 m, 2.00 m, 12 T and 320 MW). These small experimental reactors are very useful for many plasma studies while ITER is not working. When ITER injects its first plasma, many of these small reactors will begin to reach the end of their operational lives and end up being dismantled. For example, the British JET (Common European torus) ceased operations in December 2023; There is a scientific mobilization to keep it running, but the UK government is turning a deaf ear.
The reinforcement learning algorithm resembles the obstacle avoidance algorithms used in mobile robotics (tear instability would be an obstacle to be avoided). The stability of the m/n crack is associated with the resonance between the poloidal and toroidal modes with man numbers, respectively; The most relevant are 1/1 resistance detachment (RTM) and 2/1 neoclassical (NTM). The latter limits the plasma pressure, because its origin is the so-called z current bootstrapwhich occurs spontaneously with a pressure gradient.
The architecture of the deep artificial neural network (DNN) used is shown in the figure. The input contains experimental diagnostic data (one-dimensional curves associated with magnetic spectroscopy, Thomson scattering and charge-exchange recombination) and the state of the actuators (plasma heating). The output is the normalized plasma pressure (bN) a T tearability (tearability). A reward function that uses a threshold was used during training. to tearability, which is set to 0.2, 0.5 or 0.7 in this work. The reward function is R(bN, T; to) = bN > 0 a Tto, in order to maximize the plasma pressure (necessary for efficient fusion energy production). To avoid tearing, a negative reward is penalized. R(bN, T; to) = to − T < 0, you are to > T. Future studies that consider other plasma instabilities will need to use a more complex reward function that takes other factors into account.
The basic function of the intelligent control algorithm is shown in this figure (obtained through computer simulations). Without an intelligent control algorithm (black curve on the left), when the pressure exceeds a certain limit, NTM instability and subsequent disruption of the plasma occur. Thanks to the intelligent control (blue curve), the pressure gradient and tearability are controlled, which oscillate between certain edges, avoiding any instability (which can also occur when the pressure drops too low, due to the presence of too negative a pressure gradient).
This figure shows the real data of three experimental discharges: 193266 (yellow color) is a stable discharge as a reference with a standard feedback control algorithm, 193273 (black color) is an unstable discharge as a reference also with a standard control algorithm, and 193280 (blue color) is a charge controlled by an artificial intelligence. Plasma current (in MA, megaamperes), plasma power (in MW, megawatts), poloidal triangularity (a dimensionless geometric factor that measures the triangular shape of the poloidal part of the plasma), magnetic fluctuations (in G , gauss), and normalized plasma pressure are shown. In stable download 193266, the conventional control algorithm focused on bN= 1.7 while in unstable download 193273 was the target bN= 23; too much because at 2.6 seconds there was a tearing instability that led to a break at 3.1 seconds. Intelligent control (p to = 0.5) ensures that the discharge 193280 avoids instability by reaching high pressures higher than those achieved in a stable discharge.
Threshold to the breakability must be sufficient to prevent instability. This figure shows that for to = 0.2 (black curve) the control algorithm maintained low tear until 5 s, but became unstable and ended up aborting at 5.5 s. Further analysis showed that the AI correctly predicted the instability, but because the beam power was too low, the action mechanism could not reduce it further and could not prevent it. For to = 0.5 (blue curve) and to = 0.7 (red curve), the intelligent control algorithm successfully avoids instability. As shown by the normalized pressure curve, the best result was obtained for to= 0.5. The authors of the article specify that they did not attempt to determine the optimal value of this threshold, which will be the subject of future studies.
In short, a very interesting paper that shows us that the current state of the art in plasma control in (experimental) fusion reactors involves the use of artificial intelligence and deep learning techniques based on artificial neural networks. Only one instability has been checked, but there seem to be no obstacles to extending this check to many others. The use of intelligent control in ITER will tell us whether this technology will eventually become the standard in future experimental fusion reactors (which will arrive in the last quarter of this century). But the 21st century is the century of artificial intelligence, I have no doubt about that.
