|
|
| Fig. 1: Tokamak magnetic field directions and current. (Source: Wikimedia Commonos) |
Commercial fusion energy remains a major scientific and engineering challenge despite decades of progress in magnetic confinement devices such as tokamaks. Achieving net power requires heating plasma to temperatures above 100 million K and maintaining stable magnetic confinement to drive heating and avoid material damage. However, the extreme pressure and current gradients required drive plasma instabilities, making real-time control a central requirement for tokamak operation.
Tokamaks confine plasma using strong magnetic fields in a toroidal vacuum chamber. External coils generate a toroidal magnetic field that guides charged particles around the device, while a large plasma current produces a poloidal magnetic field that twists these field lines into helical paths, improving confinement away from the vessel walls. The plasma current, typically driven inductively by a central solenoid, also provides ohmic heating that raises the plasma temperature. A schematic is shown in Fig. 1.
ITER is designed to sustain plasma currents up to 15 MA with thermal energies on the order of hundreds of MJs, confined by several Tesla of magnetic fields. In these conditions plasma behaves as a highly energetic magnetized fluid. The enormous pressure and current gradients at highly confined states make the plasma intrinsically unstable, giving rise to a range of magnetohydrodynamic instabilities that grow on millisecond or faster timescales, drastically hindering the primary goal of maintaining long duration pulses, and potentially causing substantial damage to the machine. [1,2]
In a tokamak the plasma is treated as a conducting fluid. Equilibrium requires a balance between the outward plasma pressure and the confining magnetic forces generated by the toroidal field and the plasma current. This balance defines stability limits that hence depend on plasma pressure, currents, magnetic shear, and shape of confining field lines. When limits are exceeded, in highly-confined states, small perturbations in the plasma or magnetic field rapidly grow, converting stored magnetic and thermal energy into large-scale motion and distortion of the plasma column.
Several classes of instabilities arise when this force balance is pushed beyond its limits. Current-driven kink and resistive wall modes occur when the plasma current or its profile makes the confining field unable to resist bending and deformation of the plasma column. Pressure-driven and ballooning modes occur when steep pressure gradients, combined with unfavorable magnetic curvature, allow parts of the plasma to bulge outward along field lines. Reconnection-driven modes like tearing modes occur when the magnetic topology can change at resonant surfaces, forming islands that link the hot core to cooler regions and increase radial transport. At the edge, steep pedestal gradients trigger periodic expulsions of particles and energy called Edge-Localized Modes. [2,3]
|
||||||||||||
| Table 1: Typical instability and disruption timescales. [2,3] |
Once the plasma becomes unstable, magnetic forces rearrange the conducting plasma on Alfven timescales, which is the time it takes disturbances to propagate along B field lines, leading to a loss of the closed flux surface typically in microseconds to milliseconds, as summarized in Table 1. [2,3]
If an instability is not controlled it can trigger a disruption, destroying the magnetically confined state and leading to a rapid loss of the energetic plasma. Disruptions typically proceed in two stages: a thermal quench, where the plasma temperature collapses as energy streams along newly formed open lines, leading to a rise in resistivity and a subsequent current quench, unfolding in the order of milliseconds. ITER scale devices have longer current quenches (several 10s of ms) due to their massive currents and inductances. As a consequence of a disruption, energy is deposited on plasma-facing components within milliseconds, producing intense heat loads, large electromagnetic forces on the vessel structure, and beams of relativistic runaway electrons with energies of tens of MeV. Preventing or mitigating such events is therefore a central requirement for sustained tokamak operation. [1,4]
Tokamak stability is maintained through real-time feedback control systems that monitor the plasma state and adjust actuators. These operate continuously during a discharge and must respond on timescales comparable to the growth of magnetohydrodynamic instabilities.
Diagnostics: Plasma position, current, and shape are measured using magnetic probes and flux loops distributed around the vessel, typically sampled every 0.1-1 ms in advanced systems. Additional diagnostics such as interferometers and Thomson scattering provide measurements of density and temperature profiles, although magnetic measurements form the primary inputs for fast control loops. [5,6]
Equilibrium reconstruction: Sensor measurements are used to reconstruct plasma equilibrium by resolution of the Grad-Shafranov equation, which describes the balance between plasma pressure and magnetic forces in a 2D axisymmetric tokamak. Real-time codes estimate plasma boundary, current distribution, and key stability parameters with typical update times on the order of 1 - 10 ms on the most advanced compute setups. [5-7]
Control algorithms: Feedback controllers then compute corrections required to maintain the desired plasma position and shape. Most systems use classical control approaches such as proportional-integral-derivative (PID) or model-based controllers operating at loop frequencies of roughly 1-10 kHz, corresponding to control cycle times of 0.1-1 ms. [5,7,8]
Actuators: Finally, commands are applied through external poloidal field coils that adjust the magnetic configuration and stabilize the plasma column, vertical stabilization occurs in the order of 1 ms for advanced systems whereas shape control is often slower, in the order of 10 ms. Additional actuators, such as neutral beam injection, radio-frequency heating, and pellet injection modify plasma current, pressure profiles, or fuel density, although with slower effects compared to magnetic coils, often ranging in the 10s of ms. [5,7,8]
This control pipeline achieves effective response times in the order of 10 ms. This time is comparable, if not greater, to the characteristic growth of instabilities and related disruptions. As a result, the control margin between instability onset and corrective action is very small, emphasizing the difficulty to maintain highly confined and energetic plasma regimes for prolonged fusion.
Several physical and engineering constraints anchor these response times. Magnetic diagnostics provide a sparse set of measurements, requiring real-time reconstruction to infer the full plasma state. Grad-Shafranov equilibrium calculation introduces additional computational delay, while noise and incomplete measurements limit how decisively controllers respond. Additionally, the response speed of magnetic coils is constrained by coil inductance and power supply capability.
Advancements in control systems focus on adopting highly parallelized computing hardware to run equilibrium reconstruction codes faster and reduce latency for distributed control, aiming to stay below the 1 ms target. [5] Future developments in diagnostics include irradiation-hardened optical and microwave sensors to provide high-resolution, high-frequency interior plasma measurements, reducing reliance on sparse magnetic data to speed equilibrium reconstruction. [2] Future actuator architectures adopt multivariable decoupling controllers and frequency separation techniques to split faster stabilization commands to low-inductance in-vessel coils and slower shape commands to ex-vessel coils, increasing response bandwidth without saturating power supplies. [6]
Recent research explored Reinforcement Learning (RL) as a new avenue to reduce control latency in tokamaks. Instead of reconstructing plasma equilibrium and computing corrections through physics-based control pipeline, an RL policy learns a direct mapping from diagnostic signals to actuator commands. The policy is trained using a combination of simulation and experimental data, with reward functions and early-termination constraints designed to penalize trajectories that approach known instability and hardware limits. In principle this approach could bypass the 1 ms equilibrium reconstruction step and generate coil voltage commands directly from magnetic sensor data at control loop frequencies of 10 kHz (0.1 ms). Experiments on the TCV tokamak demonstrated a trained RL policy could control 19 magnetic coils to regulate plasma shape and position during short experimental discharges, while work on DIII-D showed similar policies achieve reconstruction-free magnetic control using raw diagnostic signals. [9,10]
Despite promising demonstrations, machine learning usage remains experimental and has not replaced classical control architectures in operational devices. Tokamak control is safety-critical system where incorrect commands could trigger disruptions and damage components. As a result, RL controllers are constrained by hardware interlocks and conventional feedback systems that enforce current and voltage limits. Additionally, these policies must still rely on the inherently slow actuation hardware, which remains the control bottleneck. In practice, ML approaches are currently investigated as potential supervisory tools or fast approximations to existing control algorithms rather than as standalone replacements. [9,10]
Tokamaks inherently operate near stability limits where small perturbations grow rapidly into disruptions of plasma and machine damage, with characteristic timescales ranging from hundreds of microseconds to a few milliseconds. Maintaining confinement requires feedback control systems capable of reacting on comparable timescales while managing large magnetic and thermal energies. Sensing limitations, equilibrium reconstruction latency, and actuator response constraints leave a narrow control margin between instability onset and corrective action. Continued advances in diagnostics, computing hardware, and emerging approaches such as machine learning aim to reduce this gap, but reliable disruption avoidance and control of highly confined plasmas remain a central engineering challenge for fusion reactors.
© Francesco Marchioni. The author warrants that the work is the author's own and that Stanford University provided no input other than typesetting and referencing guidelines. The author grants permission to copy, distribute and display this work in unaltered form, with attribution to the author, for noncommercial purposes only. All other rights, including commercial rights, are reserved to the author.
[1] ITER Physics Basis Editors et al., "Chapter 1: Overview and Summary," Nucl. Fusion 39, 2137 (1999).
[2] M. Lehnen et al., "Disruptions in ITER and Strategies For Their Control and Mitigation," J. Nucl. Mater. 463, 3948 (2015).
[3] T. C. Hender et al., "Chapter 3: MHD Stability, Operational Limits and Disruptions," Nucl. Fusion 47, S128 (2007).
[4] A. W. Leonard, "Edge-Localized Modes in Tokamaks," Phys. Plasmas 21, 090501 (2014).
[5] C. Galperti et al., "Overview of the TCV Digital Real-Time Plasma Control System and its Applications," Fusion Eng. Des. 208, 114640 (2024).
[6] G. De Tommasi et al., "Current, Position and Shape Control in Tokamaks," Fusion Sci. Technol. 58, 486 (2011).
[7] M. L. Walker et al., "Next-Generation Plasma Control in the DIII-D Tokamak," Fusion Eng. Des. 66-68, 749 (2003).
[8] G. De Tommasi, "Plasma Magnetic Control in Tokamak Devices," J. Fusion energy 38, 406 (2019).
[9] J. Degrave et al., "Magnetic Control of Tokamak Plasmas Through Deep Reinforcement Learning," Nature 602, 414 (2022).
[10] G. F. Subbotin et al., "Demonstration of Reconstruction-Free Static Magnetic Control of DIII-D Plasma With Deep Reinforcement Learning," Nucl. Fusion 66, 026040 (2026).
