🔬 physics

Real-time virtual circuits for plasma shape control via neural network emulators: experimental demonstration on MAST Upgrade

This paper reports the first experimental demonstration on MAST Upgrade of real-time virtual circuits for plasma shape control, which utilize neural network emulators to dynamically update control parameters online, thereby replacing conventional offline pre-set schedules with an automated, scenario-independent workflow.

Original authors: Nicola C. Amorisco, Kamran Pentland, Adriano Agnello, George K. Holt, Alasdair Ross, Matthew J. Marshall, Edward Jones, Graham J. McArdle, Charles Vincent, Timothy Nunn, Martin Kochan, Pedro Cavestany

Published 2026-08-31

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Nicola C. Amorisco, Kamran Pentland, Adriano Agnello, George K. Holt, Alasdair Ross, Matthew J. Marshall, Edward Jones, Graham J. McArdle, Charles Vincent, Timothy Nunn, Martin Kochan, Pedro Cavestany, Aran Garrod, Stanislas Pamela, James Buchanan

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 the heart of a fusion experiment, a superheated cloud of gas called plasma must be held in a precise, invisible cage. This cage is not made of metal, but of powerful magnetic fields generated by massive coils surrounding the vessel. If the plasma drifts even slightly from its intended shape, it can touch the walls, cool down instantly, and ruin the experiment. Keeping this shape stable is the job of a sophisticated control system that acts like a constant, invisible hand, adjusting the magnetic coils thousands of times every second to counteract the plasma's natural tendency to wobble or stretch. For decades, scientists have managed this task by using pre-calculated maps. Before a machine starts, experts design a schedule of adjustments based on a few specific, ideal shapes the plasma might take. During the experiment, the computer simply looks up the next step in the schedule and applies it. This works well when the plasma behaves exactly as predicted, but if the plasma suddenly changes in an unexpected way, the old map no longer fits, and the control system can struggle to keep up.

A team of researchers at the MAST Upgrade facility in the United Kingdom has now demonstrated a new way to handle this challenge. Instead of relying on a fixed schedule prepared long before the experiment, they have taught a computer to calculate the necessary adjustments in real time, as the plasma is actually moving. They used a type of artificial intelligence, known as a neural network, which acts as a fast, digital twin of the plasma. This digital twin learns how the plasma responds to changes in the magnetic coils based on a vast library of theoretical shapes. During the experiment, the system feeds the current state of the plasma into this digital twin, which instantly predicts how the shape will change in response to a tweak in the coils. The computer then uses this prediction to generate a fresh set of instructions for the magnetic coils, effectively updating the control map every few milliseconds. This approach allows the machine to adapt to the plasma's behavior as it evolves, rather than trying to force the plasma to follow a rigid, pre-written path.

The researchers tested this new method on the MAST Upgrade tokamak, a large doughnut-shaped machine designed to study fusion energy. They ran a series of experiments that grew increasingly difficult, starting with a simple, steady plasma shape and moving toward complex, rapidly changing configurations. In the first tests, the system successfully maintained a standard plasma shape, proving that the new real-time method could work alongside the existing control hardware without causing instability. The team then introduced deliberate, wobbly changes to the shape, asking the system to shift the plasma's core and move the point where the plasma touches the vessel wall. The real-time system tracked these movements with high precision, adjusting the magnetic coils to follow the new targets smoothly. This showed that the artificial intelligence could handle dynamic instructions, not just static ones.

The most demanding tests involved pushing the plasma into a highly stretched shape that is notoriously difficult to control on this machine. In one experiment, the system guided the plasma through a transition to a more elongated form while simultaneously sweeping the contact point along the wall. The real-time controller managed to keep the plasma stable throughout this evolution, a task that would have been extremely hard to plan with a fixed schedule because the plasma's behavior changes so drastically during the shift. In a final, exploratory test, the researchers asked the system to control seven different aspects of the plasma's shape all at once. While the system kept the plasma under control, the researchers observed that the adjustments became less smooth and required larger changes in the magnetic coils. This indicated that when too many variables are controlled simultaneously, the math behind the scenes can become unstable, a known challenge that the team plans to address in future work.

The results of these experiments confirm that it is possible to replace the old, pre-set schedules with a system that learns and adapts on the fly. The researchers found that they could use the same trained artificial intelligence model across all these different scenarios without needing to retrain it for each specific experiment. This suggests a future where the preparation for fusion experiments could be much simpler, as scientists would no longer need to spend weeks manually designing control schedules for every new type of plasma shape. While the system is not yet perfect and requires further refinement to handle the most complex situations, the successful tests prove that the core idea works. By giving the control system the ability to see the plasma's immediate response and adjust instantly, the researchers have taken a significant step toward making fusion experiments more robust and easier to run.

Technical Summary: Real-Time Virtual Circuits for Plasma Shape Control via Neural Network Emulators on MAST Upgrade

Problem Statement
Conventional plasma shape control in tokamaks relies on Virtual Circuits (VCs)—vectors of poloidal field (PF) coil current perturbations derived from offline linearizations of the plasma response around a limited set of reference equilibria. These VCs are typically pre-computed and deployed as fixed, expertly prepared schedules during specific phases of a discharge. A significant limitation of this approach is that as the plasma evolves or deviates unexpectedly from the planned trajectory, the fixed VCs become mismatched with the instantaneous plasma response. This degradation can compromise the orthogonality between shape-control channels and reduce control effectiveness. While Artificial Intelligence (AI) has been applied to plasma control in various ways (e.g., direct inverse mapping or reinforcement learning policies), routine deployment often faces challenges regarding interpretability, operational familiarity, and the high cost of retraining controllers for new scenarios.

Methodology
This work introduces and experimentally validates Real-Time Virtual Circuits (RTVCs) on the MAST Upgrade (MAST-U) tokamak. The core methodology replaces static, pre-set look-up tables with VCs that are updated in real-time using neural network (NN) emulators of the plasma response.

Neural Network Emulators: The system utilizes NN emulators trained on a large library of synthetic Grad-Shafranov equilibria generated by the FreeGSNKE solver. These emulators predict seven key shape parameters (inner/outer midplane radii, X-point position, squareness gap, strike-point position, and nose position) based on a reduced description of the plasma state: the currents in the twelve active PF coils, total plasma current, and prescribed current-density profile parameters.
Real-Time Linearization: At each update cycle (5 ms to 5.6 ms), the emulator predicts the shape parameters. The local shape-response Jacobian matrix is computed via finite differences of the emulator outputs with respect to the PF coil currents.
VC Generation: The Jacobian is inverted (using a Moore-Penrose pseudoinverse) to generate the local VCs. These VCs define the PF coil current perturbations required to change a specific shape parameter while suppressing changes in others.
Integration: The RTVC system is integrated into the existing MAST-U Plasma Control System (PCS) via a dedicated real-time inference server. Crucially, the downstream controller, gains, and safety limits remain unchanged. The system allows for the selection of either conventional or RTVC-derived VCs by shape parameter and pulse phase, preserving the established control architecture and interpretability.

Key Contributions and Experimental Demonstration
The paper reports the first experimental deployment of RTVCs, validating the framework across four distinct discharges (Shots 53996, 54000, 54002, and 54168) with increasing complexity:

Routine Control (Shot 53996): Demonstrated compatibility with standard MAST-U scenarios. RTVC successfully controlled the core shape ( $R_{out}, R_{in}, Z_X$ ) while a conventional VC schedule controlled the divertor nose ( $R_{nose}$ ), proving that real-time and conventional VCs can operate simultaneously.
Prescribed Perturbations and Divertor Motion (Shot 54000): Validated the ability to enact deliberate shape changes. The system successfully tracked trapezoidal perturbations in the core shape and executed a feedback-driven sweep of the divertor leg toward a Super-X configuration.
Dynamic Evolution (Shot 54002): Addressed a highly dynamic scenario involving a transition to a high-elongation, Super-X leg configuration. RTVC successfully managed the evolving plasma shape, completing the transition before a loss of control occurred due to an unmodeled interaction with the plasma nose. Offline EFIT++ reconstructions confirmed the substantial evolution in triangularity and elongation driven by the RTVC.
Simultaneous Multi-Parameter Control (Shot 54168): An exploratory stress test where all seven available shape parameters were assigned active feedback control simultaneously. While the discharge reached its programmed ramp-down, the tracking was less smooth than in previous shots, revealing limitations in the conditioning of the inversion when all parameters are active.

Results and Limitations
The experiments established that RTVCs can realize plasma shape-control tasks across diverse scenarios using the same trained models without scenario-specific retraining. The system successfully maintained shape control during constant shapes, perturbations, and dynamic transitions.

However, the study identified specific limitations:

Model Approximations: The current implementation does not account for currents induced in passive structures or real-time plasma profile evolution, relying instead on prescribed parameters. While sufficient for the tested states, these approximations may be insufficient for rapidly evolving phases.
Inversion Conditioning: In Shot 54168, attempting to control all seven parameters simultaneously led to a poorly conditioned Jacobian matrix. Specifically, near-degeneracy between the $R_{nose}$ and $R_s$ columns caused the pseudoinverse to amplify shape corrections into disproportionately large PF coil current requests. This resulted in less smooth tracking and highlighted the need for damped or Tikhonov-regularized inversion strategies in future developments.

Significance
The paper claims that this work demonstrates the experimental feasibility of integrating real-time, learned equilibrium models into an existing tokamak control architecture. The primary significance lies in establishing a pathway toward a simpler control workflow where manually constructed, phased VC schedules are replaced by VCs generated automatically online from a single, broadly trained surrogate model. This approach retains the interpretability and robustness of conventional VC-based control while enabling the controller to adapt to evolving plasma responses without the need for manual re-tuning or retraining for every new scenario. The authors position this as a central step toward more autonomous and adaptable plasma control, distinct from approaches that replace the feedback controller entirely.

Technical Summary: Real-Time Virtual Circuits for Plasma Shape Control via Neural Network Emulators on MAST Upgrade

More like this