Source-linked AI summary
Real-time virtual circuits for plasma shape control via neural network emulators: experimental demonstration on MAST Upgrade
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
TL;DR
Tokamak shape control traditionally depends on offline, phase-scheduled virtual circuits that can become mismatched as the plasma evolves. This paper uses neural-network emulators to generate local virtual circuits online within the existing MAST-U architecture, experimentally demonstrating shape-control tasks across distinct scenarios without scenario-specific retraining. The implementation establishes experimental feasibility while identifying state-representation, inverse-conditioning, and operational-envelope boundaries.
Problem
Offline, phase-scheduled VCs can become mismatched as the plasma evolves or departs from planned trajectories, while PCS deployment also requires interpretability and operational familiarity.
Method
Neural-network emulators predict plasma shape response from the instantaneous state, whose Jacobians are used to update VCs inside the existing MAST-U PCS.
Results
RTVCs controlled MAST-U plasmas across constant-shape, perturbed, divertor-leg-motion, strongly evolving, and seven-parameter experiments using the same implementation and emulators.
Takeaways & Limitations
RTVCs provide an experimentally feasible extension of conventional plasma shape control that removes manually prepared phased VC schedules across distinct scenarios.
Takeaways & Limitations
The present emulator omits passive-structure currents and uses prescribed rather than real-time plasma-profile parameters; wider operation also requires a richer state and better-conditioned inverse.
Abstract
from arXiv · showhide
Conventional plasma shape control in tokamaks relies on virtual circuits (VCs) that are computed offline from linearisations around a small, tailored number of reference equilibria, and deployed as expertly prepared schedules during the discharge. Here, we report on the first experimental deployment of real-time VCs. We replace pre-set look up tables with VCs updated in real time using surrogates of the plasma response. Both the existing control architecture and the interpretability of VC-based control are retained. Previous work showed that neural network emulators can produce accurate VCs, and validated their performance in closed-loop shape control simulations. Here, we report their first experimental validation on MAST Upgrade (MAST-U). Dedicated experiments spanning different scenarios, including prescribed shape perturbations, feedback-driven divertor-leg motion, and strongly evolving plasma configurations, show that real-time VCs can realise plasma shape control tasks within the MAST-U plasma control system. These results establish the experimental feasibility of real-time linearisations as a practical extension of conventional plasma shape control in tokamaks. The present implementation demonstrates a central step towards a simpler control workflow, in which manually constructed, phased VC schedules are replaced by VCs generated automatically online from a trained surrogate model, without scenario-specific retraining.
1 Introduction
RTVCs use neural-network emulators to update local plasma-response linearisations online, while preserving the established virtual-circuit control structure. Experiments on MAST-U provide the first experimental demonstration across distinct scenarios without scenario-specific retraining or manually prepared VC schedules.
- AI plasma-control methods must preserve interpretability and operational familiarity, not only control robustness.
- Conventional VCs are replaced by emulator-derived local linearisations that retain explicit feedback on reconstructed plasma-shape error.
- Broadly trained emulators can provide real-time VCs across different plasma conditions without scenario-specific retraining or manually prepared VC schedules.
- Earlier work trained shape emulators from broad equilibrium libraries, differentiated them to construct VCs, and validated them in static and closed-loop simulation.
- The MAST-U experiments span conventional shapes, prescribed perturbations, divertor-leg motion, and strongly evolving configurations.
2 Experimental methodology
RTVCs generate virtual circuits from neural-network predictions of the instantaneous plasma state and integrate them into the existing MAST-U PCS. The deployed workflow uses finite-difference Jacobians, pseudoinversion, and prior software and non-actuating validation.
- Conventional VCs linearise controlled shape parameters around reference equilibria and invert the Jacobian to obtain PF-coil current perturbations.
- Offline MAST-U linearisations are sampled along planned trajectories and manually assigned to pulse phases, becoming mismatched as the plasma evolves away from references.
- Neural-network emulators predict seven controlled shape parameters from PF-coil currents, total plasma current, and prescribed profile parameters.
- At each update, the PCS receives an emulator-derived Jacobian, computes its pseudoinverse, and replaces selected entries of the conventional VC matrix while retaining downstream control and protection handling.
- Finite differences evaluate the Jacobian because automatic differentiation is unavailable in the real-time inference engine.
- The workflow was validated through software tests, playback, emulator comparisons, and non-actuating piggyback operation before closed-loop experiments.
3 Experimental demonstration
MAST-U experiments tested RTVCs from routine shape control through prescribed perturbations, divertor-leg motion, dynamic evolution, and seven-parameter stress testing. RTVCs achieved the requested shape tasks, while conditioning and model-state limitations constrained some behavior.
- 3.1 Routine plasma control: Routine shot 53996 reached its programmed duration while RTVC-controlled shape parameters remained close to requested waveforms.
- 3.2 Prescribed shape perturbations and divertor-leg motion: Shot 54000 successfully tracked prescribed trapezoidal perturbations in Rout and ZX and enacted the requested divertor-leg sweep before termination at 563 ms.
- 3.2 Prescribed shape perturbations and divertor-leg motion: The 563 ms termination followed a Langmuir probe protection trip and was not attributed to RTVC operation.
- 3.3 Dynamic plasma shape evolution: Shot 54002 completed the requested transition toward higher elongation and a near-Super-X configuration before shape control was lost after interaction with the nose.
- 3.3 Dynamic plasma shape evolution: Offline EFIT++ reconstructions independently corroborated the substantial evolution in triangularity, elongation, and divertor-leg position.
- 3.4 Simultaneous control of seven shape parameters: In the seven-parameter stress test, transient departures occurred around 0.50 s and 0.77 s, followed by less smooth tracking before the 1 s ramp-down.
4 Discussion and outlook
Experiments established that real-time, emulator-derived linearisations can control MAST-U plasmas across distinct tasks using the existing architecture and no scenario-specific retraining. The discussion identifies conditioning and emulator-state approximations as boundaries, while pointing toward simpler online VC preparation.
- Experimental feasibility: RTVC controlled MAST-U plasmas across constant-shape, perturbation, divertor-leg, and strongly evolving scenarios without scenario-specific retraining.The same implementation and emulators were used while retaining the established shape-control architecture.
- Experimental feasibility: Shot 54168 extended RTVC testing to simultaneous active feedback control of all seven available shape parameters.This configuration was deliberately outside normal MAST-U practice because several shape parameters have strongly correlated PF-coil responses.
- Model scope: The emulator omitted passive-structure currents and used prescribed rather than real-time plasma-current-density-profile parameters.These approximations did not preclude control in the tested states but may be inadequate during rapidly evolving phases.
- Conditioning: Near-collinearity of the Rnose and Rs columns around 0.50 s amplified pseudoinverse VC drives and increased PF-coil current requests.The resulting amplification was associated with greater sensitivity to cross-coupling, measurement error, and model mismatch, although causality for less-smooth tracking was not established.
- Outlook: A single emulator can generate evolving local linearisations online, reducing bespoke phased VC schedules and scenario-specific retraining.Routine use still requires a richer real-time plasma state, a better-conditioned inverse, and validation across a wider operational envelope.