🔬 physics

On the feasibility of model-based feedback control of vertical instability growth rate using out-vessel coils in ARC-like scenarios

This paper proposes and validates a model-based feedback controller that utilizes out-vessel poloidal field coils to directly regulate the vertical instability growth rate in high-elongation ARC-like tokamaks, achieving successful stabilization in 83% of simulated scenarios despite the absence of in-vessel coils.

Original authors: Arunav Kumar, Cesar Clauser, Theodore Golfinopoulos, Jon C. Hillesheim

Published 2026-08-25

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Arunav Kumar, Cesar Clauser, Theodore Golfinopoulos, Jon C. Hillesheim

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Inside a fusion reactor, the superheated gas known as plasma behaves like a charged fluid that must be held in a perfect, invisible bottle made of magnetic fields. To generate enough energy to power a city, this plasma needs to be stretched tall and thin, a shape that boosts its performance but also makes it inherently unstable. Like a pencil balanced on its tip, a tall plasma column wants to tip over sideways. If it does, the discharge ends instantly, and the sudden release of energy can damage the reactor's inner walls. For decades, scientists have stabilized these tall columns using coils placed very close to the plasma, inside the reactor vessel. However, in future power plants, the intense heat and radiation will likely force engineers to place these stabilizing coils outside the vessel, further away from the plasma. This distance weakens the magnetic grip, making the task of holding the plasma upright significantly harder.

Researchers at the Massachusetts Institute of Technology and Commonwealth Fusion Systems have explored whether a new type of computer controller could solve this problem. They focused on a specific design for a future fusion power plant called ARC, which relies entirely on these distant, external coils. Instead of simply trying to keep the plasma in a fixed position, their new system directly monitors how fast the plasma is starting to tip over. This rate of tipping, known as the growth rate, is the true measure of stability. By watching this number in real time, the controller can act before the plasma moves far enough to cause trouble. The team built a digital brain for this system that combines a machine learning model with a sophisticated optimization algorithm. The machine learning part acts as a fast expert, predicting the stability of the plasma based on its current shape and internal pressure, while the optimization algorithm decides exactly how much voltage to send to each external coil to keep the plasma steady.

To test this idea, the researchers ran thousands of computer simulations of the ARC reactor operating under various conditions. They did not build a physical machine; instead, they created a detailed virtual environment that mimics the physics of the plasma and the electrical circuits of the coils. In these simulations, they introduced different challenges, such as changing the plasma's shape, reducing the power available to the coils, or simulating sudden internal jolts. The results were promising but revealed clear limits. In about half of the scenarios, the controller successfully kept the plasma stable, maintaining the desired growth rate with high precision. In another third of the cases, it managed to keep the plasma from crashing, though with some difficulty or slight deviations from the ideal target. However, in the remaining cases, the system lost control. These failures occurred when the plasma was so unstable that the external coils simply could not push back hard enough, or when the plasma changed so drastically that the computer's predictions became unreliable.

The study highlights that while controlling a tall, unstable plasma from a distance is possible, it requires a very specific approach. The researchers found that simply trying to hold the plasma in a fixed spot is not enough; the controller must actively manage the speed at which the instability grows. Their machine learning model was crucial because it could account for the complex, shifting internal structure of the plasma, something simpler models often miss. By feeding the controller real-time data about the plasma's internal pressure and current distribution, the system could make smarter decisions about which coils to use. The simulations showed that the controller works best when the plasma is not too unstable to begin with. When the instability grew too fast, exceeding a certain threshold, the external coils ran out of electrical power to counteract it, and the plasma tipped over.

Another key finding was that the system is fragile when multiple problems happen at once. If the voltage available to the coils is slightly reduced and the computer's sensors are slightly delayed, the system might still work. But if both issues occur together, the controller fails. This suggests that future power plants will need to be designed with significant safety margins, ensuring that the coils have plenty of power and the sensors are extremely fast. The researchers also noted that their system focused solely on preventing the plasma from tipping over; it did not actively control the detailed shape of the plasma's edge. In their simulations, the shape remained stable only because the initial conditions were set up perfectly. In a real power plant, a separate layer of control would be needed to maintain the plasma's shape over long periods, working alongside this stability controller.

Ultimately, this work provides a roadmap for how future fusion reactors might be controlled. It proves that using distant coils is feasible, provided the controller is smart enough to predict the plasma's behavior and fast enough to react. The study does not claim to have solved the problem of fusion energy, nor does it suggest that this specific controller is ready for a real reactor. Instead, it identifies the boundaries of what is possible with current technology. It shows that while the physics allows for control from the outside, the margin for error is narrow. The path forward involves refining these digital controllers, testing them on existing reactors, and ensuring that the hardware can handle the strict timing and power requirements. If these hurdles are cleared, the vision of a fusion power plant that can safely harness the power of the stars, even with its stabilizing magnets placed far away, moves one step closer to reality.

Technical Summary: Feasibility of Model-Based Feedback Control of Vertical Instability Growth Rate in ARC-like Scenarios

Problem Statement
High-elongation tokamak plasmas, essential for achieving high fusion performance, are inherently vertically unstable to the axisymmetric $n=0$ mode. In current research devices, this instability is typically stabilized using in-vessel coils that offer short coupling distances and high bandwidth. However, in future fusion power plants like the ARC design, in-vessel coils may be absent due to blanket and vacuum vessel constraints, leaving only distant out-vessel poloidal field (PF) coils for stabilization. The increased stand-off distance weakens magnetic coupling and increases the required control effort. Furthermore, conventional proportional-derivative (PD) control, which regulates the plasma current centroid position, does not directly regulate the stability margin (growth rate, $\gamma_{gr}$ ). A plasma can satisfy a position reference while operating with a dangerously high growth rate, as centroid displacement does not immediately reflect changes in elongation, internal inductance, or pressure profiles that drive instability. The challenge is to design a controller that directly regulates $\gamma_{gr}$ using only out-vessel actuators, while managing the multivariable coupling between vertical stability, double-null (DN) flux balance, and plasma shape.

Methodology
The authors propose a model-based feedback controller that regulates the vertical instability growth rate ( $\gamma_{gr}$ ) as the primary controlled variable. The architecture consists of three integrated components:

Machine Learning Surrogate for Non-Rigid Stability:
A neural network surrogate is trained to estimate $\gamma_{gr}$ and its sensitivity to actuator currents in real-time. Unlike rigid plasma approximations, this surrogate accepts non-rigid profile inputs: the pressure gradient $P'(\psi)$ and the toroidal field function term $TT'(\psi)$ . This allows the model to resolve profile-dependent variations in instability that integral parameters (like internal inductance $\ell_i$ ) cannot distinguish. The architecture uses a dual-branch approach: a Multi-Layer Perceptron (MLP) processes scalar equilibrium parameters (e.g., $I_p$ , $\ell_i$ , coil currents), while a 1D Convolutional Neural Network (CNN) processes the profile shapes. These branches are fused to output $\gamma_{gr}$ , perturbed flux, and current density. Automatic differentiation is used on the trained surrogate to compute state-dependent actuator sensitivities ( $\partial \gamma_{gr} / \partial I_{PF}^c$ ) in real-time.
Constrained Quadratic Program (QP) Controller:
The controller employs a constrained QP to compute coil voltage commands ( $V_{PF}$ ) at each time step. The cost function minimizes the error between predicted and target values for four outputs: growth rate ( $\gamma_{gr}$ ), DN flux balance ( $\psi_r$ ), vertical position ( $z$ ), and plasma current ( $I_p$ ). The QP explicitly enforces hard constraints on coil voltages and currents, reflecting the physical limitations of out-vessel power supplies. The controller does not explicitly regulate boundary shape (elongation, triangularity) in the cost function; shape is assumed to be managed by a separate outer loop or feedforward design.
Simulation Environment:
The system is tested on the ARC V3A power plant design configuration (high-field, compact tokamak) using an offline closed-loop simulation. The simulation couples the MEQ-FGE-L equilibrium evolution code (which includes passive structure eddy currents) with models of the PF coil circuits and power supplies. The control cycle operates at 1 kHz.

Key Results
The controller was evaluated across 24 closed-loop simulation cases covering equilibrium variations, actuator degradations, and transient disturbances.

Performance Metrics: The controller achieved "Pass" (meeting all design criteria) in 50% of cases, "Marginal" (meeting criteria narrowly or with intermittent violations) in 33%, and "Fail" in 17%.
Success Cases: In the baseline scenario ( $\kappa_{sep}=1.80$ ), the controller regulated $\gamma_{gr}$ from an initial +4.2 s $^{-1}$ to a target of 5 s $^{-1}$ within 0.5 seconds, maintaining vertical excursions below 2% of the minor radius.
Failure Modes:
- Voltage Authority Limits: Cases with very high initial growth rates (e.g., $\gamma_{gr} > 10$ s $^{-1}$ due to high elongation and peaked current profiles) failed because the 500 V voltage limit was insufficient to counteract the instability.
- Compound Degradation: While individual degradations (e.g., reduced voltage or increased latency) were often manageable, combinations (e.g., low voltage + high latency) led to failure, eroding stability margins from multiple directions.
- Surrogate Extrapolation: Large transient disturbances (e.g., extreme sawtooth crashes) pushed the equilibrium outside the surrogate's training distribution, causing prediction errors and sensitivity inaccuracies that led to loss of control or sustained oscillation.
Shape Coupling: In the baseline case, the separatrix elongation remained stable ( $\pm 0.02$ ) without explicit shape control, attributed to the weak coupling between the dominant stabilizing actuator (PF5 differential mode) and shape-setting field harmonics. However, the authors note that sustained shape evolution would require a separate control layer.

Significance and Claims
The paper claims to demonstrate the feasibility of using a model-based controller to directly regulate vertical instability growth rates using only out-vessel coils in a high-elongation, double-null scenario representative of a fusion power plant.

Direct Growth Rate Regulation: The study validates that regulating $\gamma_{gr}$ directly, rather than position, provides a stability margin signal before large vertical displacements occur, extending the correction window.
Profile-Aware Surrogates: The work highlights the necessity of using profile-aware (non-rigid) ML surrogates to accurately capture instability variations driven by internal plasma profiles, which integral parameters miss.
Constraint Handling: The constrained QP approach effectively manages the multivariable coupling and actuator limits inherent to out-vessel control, outperforming unconstrained or decoupled control strategies in the tested scenarios.
Boundaries of Controllability: The study maps the boundaries of out-vessel controllability, identifying that voltage authority and the accuracy of the stability surrogate under extreme conditions are the primary limiting factors.

The authors explicitly state that this work establishes a basis for future experimental validation but does not constitute a finalized engineering solution. They emphasize that the controller does not explicitly regulate boundary shape and that a cascade architecture with a separate shape control layer is required for sustained scenario evolution. The results are based on offline simulations, and hardware-in-the-loop testing is identified as the necessary next step.

Technical Summary: Feasibility of Model-Based Feedback Control of Vertical Instability Growth Rate in ARC-like Scenarios

More like this