betaFeedback

WorksPhysics and Astronomy

Real-time virtual circuits for plasma shape control via neural network emulators: dynamic validation in closed-loop simulations

Abstract

Reliable confinement and stable performance of tokamak fusion plasmas require accurate real-time magnetic shape control. A promising route to reduced latency and increased flexibility in plasma control systems (PCS) is to emulate physics-based controllers using neural networks (NNs). In prior work, we have demonstrated that virtual circuits (VCs), which define the poloidal field coil current vectors able to modify each plasma shape parameter independently, can be accurately emulated with NN models trained on a large library of simulated Grad-Shafranov equilibria. This enables magnetic controllers to accurately adapt to evolving plasma equilibria using real-time VCs (RTVCs), in contrast to pre-set VC schedules whose performance degrades upon departure from their reference equilibria. Here, we investigate the performance and robustness of these RTVCs in closed-loop simulations using the FreeGSNKE Pulse Design Tool (FPDT): a framework that couples the FreeGSNKE evolutive equilibrium solver with a virtual PCS. The FPDT models the coupling between controllers, plasma current and shape response, and actuator constraints. Using the RTVCs within the FPDT, we demonstrate effective in-silico control of MAST Upgrade (MAST-U) plasma scenarios and show that the emulators are robust in the presence of input measurement uncertainty and under different update frequencies. These results establish the viability of RTVCs for closed-loop plasma shape control, representing a key step toward real-time deployment in the MAST-U PCS.

Introduction

Accurate real-time magnetic control of tokamak fusion plasma position and shape is essential for maintaining reliable confinement and enabling access to high-performance operating regimes and advanced divertor configurations [1, 2]. To this end, the plasma control system (PCS) regulates deviations in geometrical shape parameters (e.g. midplane radii, X-point locations, strikepoints—typically under the assumption of toroidal symmetry) from prescribed reference waveforms by computing the required adjustments to poloidal field (PF) coil currents. These adjustments are enacted in real-time via PF coil voltage actuation, with sub-millisecond latency [3, 4]. Here, our focus is on calculating the adjustments in PF coil currents using virtual circuits (VCs) [5]. A VC matrix decouples the shape control problem around a specific plasma equilibrium, i.e. it determines the PF coil current changes required to move each shape parameter independently, by one physical unit.

Building a VC matrix is not currently feasible in real time as it requires a linearisation of the Grad-Shafranov (GS) equilibrium [6, 7] to obtain the sensitivities of the shape parameters with respect to the PF coil currents. In classical plasma scenario design, VCs for a small number of target GS equilibria are calculated offline and used throughout a plasma discharge in a pre-set, phase-segmented manner. This has been demonstrated successfully on MAST Upgrade (MAST-U) [8], JET [9, 10], EAST [11], JT-60SA [12], and TCV [13], as well as used for designing shape controllers for ITER [14]. Since a VC’s validity is only local, its accuracy rapidly degrades as the plasma equilibrium evolves away from the reference state, causing the pre-calculated VCs to lose their effectiveness in producing orthogonal plasma responses. This can lead to coil current adjustments interfering with one another, potentially driving unnecessary stress on the PF coil power supplies.

Prior work has demonstrated VCs can be accurately emulated using feedforward neural networks (NNs), based on real-time tokamak diagnostic measurements [15, 16]. With infer ence times on the order of milliseconds, NN emulators enable state-aware decoupling of plasma shape control by providing real-time VCs (RTVCs) that are continuously updated with the evolving plasma state. This can improve shape controller responsiveness and accuracy, particularly for complex shape displacements that would otherwise require segmented control strategies with a large number of pre-set time-phases and pre-computed VCs at the corresponding equilibrium waypoints. This contribution complements an increasing body of work related to plasma control using NNs [17, 18, 19, 20] and reinforcement learning [21, 22, 23, 24].

In this paper, we validate the performance and test the robustness of the RTVCs for shape control by deploying the NN models presented in [16] within closed-loop plasma equilibrium simulations using the open-source FreeGSNKE Pulse Design Tool (FPDT) [25, 26]. Previously validated against MAST Upgrade (MAST-U) discharges, the coupled evolutive equilibrium solver and virtual PCS within the FPDT provide an ideal test bed for assessing the RTVCs in silico, in preparation for deployment. This ‘dynamic validation’ complements previous work on purely ‘static validation’ (more details in Section II), by accounting for the evolving currents in passive conducting structures and the interaction with other controllers in the PCS, both of which are neglected in static tests. It also enables monitoring of the plasma and coil responses to frequently changing RTVCs over time, providing a more realistic assessment of their impact on closed-loop plasma control.

In Section II, we briefly summarise the process by which the RTVCs are generated and how they were statically validated. We then explain, in Section III, how the RTVCs are seamlessly integrated into the virtual PCS within the FPDT. In Section IV, we perform robustness tests and verify that RTVCs can successfully control MAST-U plasma discharges in silico.

Emulated virtual circuits

In the following, we adopt cylindrical coordinates (R,ϕ,Z)(R,\phi,Z) under toroidal (ϕ)(\phi) axisymmetry, with RR denoting the major radius and ZZ the vertical distance from the tokamak midplane.

Definitions and notation

For any given GS equilibrium, we describe its core shape and divertor leg characteristics with a set of geometric shape parameters P∈RmP \in\mathbb{R}^{m}. Plasma equilibria are fully defined by the currents in any active and passively conducting metals Im=(Iact,Ipas)I_m=(I_{\mathrm{act}},I_{\mathrm{pas}}); the total plasma current IpI_p; and the plasma current density profile parameters θ\theta. Therefore, we have P=P(Im,Ip,θ)P=P(I_m,I_p,\theta). Note, however, that the passive currents IpasI_{\mathrm{pas}} are not actively controlled, and only the active PF coil currents Iact∈RnI_{\mathrm{act}}\in\mathbb{R}^{n} (m≤nm\leq n) can in fact be modulated. For this work, PP encapsulates the inner and outer midplane radii (Rin,Rout)(R_{\mathrm{in}},R_{\mathrm{out}}); the lower X-point position (RX,ZX)(R_X,Z_X); the “gap” between the plasma core and a specific point on the wall of the main chamber SgapS_{\mathrm{gap}} (i.e. squareness); the lower outboard strikepoint RsR_s; and the radial intersection of lower divertor leg and a virtual line crossing the divertor entrance RnoseR_{\mathrm{nose}} (refer to Figure 1 and [16] for more details).

Poloidal cross-section of a MAST-U equilibrium

Figure 1. Poloidal cross-section of a MAST-U equilibrium (shot 53278, t=0.5t=0.5s) equilibrium generated using FreeGSNKE. Displayed are the contours of total poloidal flux (yellow/green), the separatrix (red), and the shape parameters (labelled black dots). Also shown are the PF coils (blue), vessel structures (grey), and first wall outline (black).

The VC matrix V\mathcal{V} is defined as the pseudoinverse of the sensitivity matrix SS, which itself is the Jacobian of shape parameters with respect to PF coil currents1 for a specific GS equilibrium:

V=(STS)−1ST,S=∂P∂Iact∣(Im,Ip,θ).(1)\mathcal{V}=(S^{T}S)^{-1}S^{T}, \qquad S=\left.\frac{\partial P}{\partial I_{\mathrm{act}}}\right|_{(I_m,I_p,\theta)}. \tag*{(1)}

The columns of V\mathcal{V} represent the vectors of PF coil current changes required to modify a shape parameter by one unit, without altering the others. The validity of VCs is only local: a VC matrix V\mathcal{V} will not achieve this decoupling when applied to different equilibria, with progressively degrading performance as the equilibria diverge from one another. In addition, the aforementioned parameters are in fact time-dependent but we will drop the additional notation for brevity.

We should note that, alternatively, V\mathcal{V} can be computed via singular value decomposition of SS rather than the standard pseudoinverse, which has been shown to reduce cross-coupling between PF coil circuits at the cost of shape accuracy [8, 9, 10, 11, 12, 13, 14]; here, however, we adopt the standard pseudoinverse formulation in (1) and leave this comparison to future work.

Neural network emulation and static validation

Here, we provide a brief description of the NN emulators, including the GS library generation, NN training, and static validation, but we refer the reader to [16] for further details.

Feedforward NNs are used to learn a mapping from (Iact,Ip,θ)∈R17(I_{\rm act}, I_{\rm p}, \theta) \in\mathbb{R}^{17} to P∈R7\mathbf{P} \in\mathbb{R}^{7} (where Iact∈R12I_{\rm act} \in\mathbb{R}^{12}), which we differentiate to derive sensitivity matrices and corresponding VCs. Compared to emulating the full Jacobian SS, this significantly reduces the dimensionality of the output space, consequently reducing both minimum model size and training dataset size. Similarly, we also drop the dependence on IpasI_{\rm pas} for two reasons. Firstly, IpasI_{\rm pas} itself has limited effect on SS, and second, we expect real time deployment of the emulators to take place before being able to accurately estimate IpasI_{\rm pas} in real time on MAST-U [8].

In order to build the mapping above we used a library of synthetic plasma equilibria. Starting from experimentally observed “seed” equilibria from previous MAST-U campaigns, we used the random walk Markov Chain Monte Carlo method presented in [27] to explore the input space and build the training dataset. This resulted in a dataset comprised of approximately 900k FreeGSNKE GS solutions spanning the operational space of MAST-U, including both experimentally observed and yet-to-be observed plasma configurations. The trained models typically contain 4-5 hidden layers and 128-256 nodes per layer, for a total of 1–5 × 10510^{5} trainable weights. We retain the top eight performing models, which we use to produce ensemble estimates of the Jacobian SS using finite differences.

Static validation of the finite-difference-built RTVCs was carried out in [16] to assess the accuracy and orthogonality of the plasma response, i.e. verifying that a requested change in one shape parameter would not change the others. This involved systematically comparing an RTVC-requested 5 mm shape parameter change with the observed change obtained when applying either a finite-difference-based VC from FreeGSNKE or an auto-differentiated VC from RTVC. Sufficient accuracy was demonstrated using the finite-difference-based RTVCs (which we will use here).

Integration with the FreeGSNKE Pulse Design Tool

The FPDT is a simulation framework that enables a closed-loop interaction between the FreeGSNKE evolutive plasma simulator and a virtual PCS. Inspired by the MAST-U PCS, the virtual PCS used by the FPDT regulates plasma current, shape parameters, and vertical plasma position, simultaneously enforcing safety limits on coil currents and voltages2. At each time step, the PCS takes in equilibrium-based measurements of plasma current, shape parameters, and vertical position, returning voltage commands to be applied to the active PF coils. Together with pre-prescribed plasma profiles θ\theta and a plasma resistivity value, these voltages are passed to the simulator and used to advance the plasma and metal currents forward in time.

Focusing on shape control at simulation time tt, the Shape Controller provides a requested rate of change in P\mathbf{P}, computed from the measured shape parameters using a proportional-integral-derivative (PID) controller. These requests are used by the Virtual Circuits Controller to calculate the requested rate of change of the active PF currents:

[ddtIact]req=V(t)[ddtP]req.(2)\left[\frac{\mathrm{d}}{\mathrm{d}t} I_{\rm act}\right]_{\rm req} = \mathcal{V}(t)\left[\frac{\mathrm{d}}{\mathrm{d}t}\mathbf{P}\right]_{\rm req}. \tag*{(2)}

When using the RTVCs, the VC matrix V(t)\mathcal{V}(t) is obtained within the FPDT by querying the NN emulators, which are provided with the latest ‘measurements’ for the inputs (Iact,Ip,θ)(I_{\rm act}, I_{\rm p}, \theta) from the simulator. No additional changes to the control algorithm are required in order to do this. Also note that there are separate plasma current and vertical position controllers running alongside the shape controller—see [26 Section 2.3] for more details.

Closed-loop simulations using emulated virtual circuits

We now perform closed-loop simulations using the RTVCs within the FPDT on a MAST-U-like discharges. We use a past discharge as a reference and adopt the original PCS configuration for the shot as input to the FPDT. In order to test different plasma shapes, however, we modify the shape controller settings (e.g. the feedback reference waveforms).

Simulations focus on the flat-top phase, using fixed timestep dt=0.2\mathrm{d}t = 0.2 ms, with the PCS called every 0.1 ms. In order to match the update frequency of the real MAST-U PCS, we call the virtual PCS twice per simulator time step, with identical observations. The voltage commands from the two calls are averaged before being applied to the FreeGSNKE time stepper. The initial equilibrium and plasma profiles θ\theta are built using magnetics-only EFIT++ reconstruction data, while the resistivity profile is determined offline. In Figure 2 and Figure 4 we use green shading to indicate when a particular shape parameter is under feedback control and no shading to indicate when no control is applied. Note that the inversion in (1) is performed using only the shape parameters selected to be controlled during simulation. All simulation results use piecewise linear evolution of the dynamics (circuits and plasma equations), coupled with a fully nonlinear solution of GS—refer to [26 Section 2.2.3].

FPDT-simulated evolution of shape parameters $P$ using RTVCs for two plasma scenarios

Figure 2. FPDT-simulated evolution of shape parameters PP using RTVCs for two plasma scenarios: scenario 1 (blue) and scenario 2 (pink). Solid, dashed, and dotted lines correspond to using constant VCs, RTVCs with γ=20\gamma= 20 ms, and RTVCs with γ=2\gamma= 2 ms, respectively. The corresponding feedback reference waveforms are also shown (dashed black for scenario 1 and dashed brown for scenario 2) with green shading indicating when feedback control in on. The respective grey lines show the effect of including measurement uncertainty within the simulations (see main text for further details).

Evolution of shape parameters under feedback control

Figure 4. FPDT-simulated evolution of shape parameters PP using the RTVCs, updated every γ=5\gamma= 5 ms. The simulator uses the same θ\theta as in the reference shot (solid blue), as in shot 53098 (dashed pink), and as in shot 53097 (dotted dark blue). Also shown are the feedback (dashed black) reference waveforms used in the virtual PCS and green background shading, indicating that feedback control is switched on.

Responsiveness to VC update frequency and measurement uncertainty

We start by focusing on two different scenarios, both based on the same MAST-U reference shot 53152: a Super-X divertor, Ohmically heated, double-null discharge with Ip≈750I_{\rm p} \approx750 kA. Scenario 1 retains the typical Super X divertor target waveforms of the original setup. This scenario tests the performance of the controllers during both the conventional (t<0.3t < 0.3s) and Super X phases (t>0.4t > 0.4s), as well as during the dynamic transition between the two (0.3s<t<0.4s0.3\mathrm{s} < t < 0.4\mathrm{s}). Scenario 2 introduces further changes to the plasma core, thereby exploring the performance of the emulators in a wider range of configurations. Target waveforms for both scenarios are displayed as dashed black and brown lines in Figure 2. We do not modify any additional controller settings, and use, for example the same PID gains adopted in the original discharge.

For each scenario in Figure 2, we compare simulation results using two different update times γ\gamma. The update time defines the rate at which the NN models are queried for a new VC. Equivalently, it defines how long an emulated VC is used (i.e. over [t,t+γ)[t,t+\gamma) in (2)) before being re-calculated using the latest measurements3. Preliminary integration tests with the MAST-U PCS lead us to believe that the current hardware will permit γ≈1\gamma\approx1-5 ms and therefore, here, we have explored a baseline case γ=2\gamma= 2 ms and a conservative case γ=20\gamma= 20 ms. We find the emulator performance to be robust across both choices of γ\gamma: even in the conservative case, the RTVCs perform as expected. The controlled parameters remain within a centimetre or so of the desired feedback reference waveforms, an acceptable level of accuracy for MAST-U [8], while the uncontrolled parameters evolve freely. These results also highlight the ease with which one can modify the feedback references between different scenarios and simply simulate, without having to pre-calculate and schedule new VCs manually. Figure 3 illustrates the evolution of the two plasma separatrices side by side.

Plasma separatrices at three different times for scenario 1 and scenario 2

Figure 3. Plasma separatrices at three different times for scenario 1 (dotted blue) and scenario 2 (dotted pink) during the (γ=2\gamma= 2 ms) FPDT simulations in Figure 2. Also shown are the corresponding constant VC simulations (solid lines).

Also shown in Figure 2 and Figure 3 are simulations that use constant VCs (solid lines), constructed around the initial equilibrium at t=0.2t = 0.2 s. Both sets of simulations track the shape change requests successfully, with slightly slower responses for RinR_{\mathrm{in}} and RoutR_{\mathrm{out}} (scenario 2) and RsR_{\mathrm{s}} (scenario 1), though these are minor. We reiterate that this work aims to demonstrate the successful use of RTVCs in closed-loop simulations, without claiming improved shape-control performance. Future work will investigate whether RTVCs can improve control accuracy, reduce coil-current requirements, and/or enable plasma shapes that have previously proved challenging to achieve.

The two solid grey lines, one for each scenario, in Figure 2 show the impact measurement uncertainty has on the simulations (when γ=2\gamma= 2 ms). Here, we have added Gaussian noise4 to the “measurements” of IactI_{\mathrm{act}}, IpI_{\mathrm{p}}, and PP generated by the simulator at each time step. The effect on the simulation here is twofold, as the uncertainty changes the requests generated by the PID controllers in the virtual PCS and also the VCs generated by the NN emulators. In both cases, we see that both the virtual PCS and the emulators perform very well with the uncertainty and track the corresponding unperturbed simulations. These results demonstrate that the RTVCs are ready for the uncertain conditions that deployment in a real-world PCS will yield.

Robustness against plasma profiles uncertainty

In the absence of real-time inference for the plasma profiles θ\theta, the RTVCs will need to be provided with time dependent profile prescriptions derived from a reference shot with a similar ramp up, plasma heating profile, and fuelling strategy. This experiment aims to evaluate robustness to discrepancies between these estimated plasma profiles prescribed to the emulators and the actual real time plasma conditions. In order to mimic the deployment of RTVCs during a real discharge, we use one set of plasma profiles in the simulator and a different set in the RTVCs, which allows us to quantify this residual variability in closed-loop simulation.

In Figure 4, we show the evolution of the shape parameters when all seven parameters in PP are under feedback control using the RTVCs, updated every γ=5\gamma= 5 ms. The reference is MAST-U shot 53278, a conventional divertor, lower single-null discharge with Ip≈750I_p \approx750 kA, heated by ∼3.2\sim3.2 MW neutral beam injection (NBI) power.

In each FPDT simulation, the RTVCs use the same profile parameters θ\theta from the reference shot, while the simulator uses either: the reference profiles; profiles from shot 53098; or profiles from shot 53097. Shot 53098 had slightly more NBI heating (∼3.4\sim3.4 MW), whereas shot 53097 had significantly less (∼1.8\sim1.8 MW) due to the loss of one of the two beam systems. During each simulation, the alternative profiles used in the simulator are “linearly blended” into the existing profiles between 300-350 ms, with the new profiles in full effect thereafter5. In all cases, we observe that the shape parameters remain robust to variations in the profile parameters. As expected, the core shape parameters exhibit the largest deviations between simulations, since profile evolution has relatively little influence on the divertor leg geometry. Even in the case of a NBI system failure, the emulators are able to maintain robust plasma shape control.

Discussion

In this paper, we have presented an in silico validation of RTVCs for tokamak plasma shape control, demonstrating that they can be embedded within a (virtual) PCS framework and interact seamlessly with conventional, non-AI-based controllers. Performance is maintained across different updates times, consistent with conservative real-time inference delays, and remains robust under variations in plasma profile parameters, measurement noise, and evolving passive structure currents not seen during training.

Overall, these results show that RTVCs are suitable for real-time magnetic shape control. They are expected to reduce reliance on intricate combinations of feedforward/feedback scenario design, minimising the need for expert offline VC computations, thereby enabling greater automation within plasma control. This validation marks a step towards experimental verification and the eventual deployment of RTVCs into an operational PCS.

Acknowledgment

This work was funded by the Fusion Computing Lab collaboration (between UKAEA and STFC Hartree) and part funded by the EPSRC Energy Programme (EP/W008639/1). For the purpose of open access, the authors have applied a Creative Commons Attribution (CC BY) licence to any author accepted manuscript version arising from this submission. To obtain further information, please contact [email protected].

References

  1. [1]W. Morris et al. “MAST Upgrade divertor facility: A test bed for novel divertor solutions.” IEEE Trans- actions on Plasma Science, 46(5):1217–1226, 2018. doi:10.1109/TPS.2018.2815283.DOI
  2. [2]M. Lennholm et al. “Controlling a new plasma regime.” Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences, 9 October 2024; 382 (2280): 20230403. doi:10.1098/rsta.2023.0403.DOI
  3. [3]B. G. Penaflor et al. “Worldwide collaborative efforts in plasma control software development.” Fusion Engineering and Design, 83(2):176–180, 2008. doi:10.1016/j.fusengdes.2007.11.012.DOI
  4. [4]M. Ariola and A. Pironti. “Magnetic Control of Tokamak Plasmas.” Advances in Industrial Control, Springer International Publishing, 2016. doi:10.1007/978-3-319-29890-0.DOI
  5. [5]G. McArdle, L. Pangione, and M. Kochan. “The MAST Upgrade plasma control system.” Fusion Engineering and Design, 159:111764, 2020. doi:10.1016/j.fusengdes.2020.111764.DOI
  6. [6]H. Grad and H. Rubin. “Hydromagnetic equilibria and force-free fields.” In Proceedings of the Second United Nations International Conference on the Peaceful Uses of Atomic Energy: Theoretical and experimental aspects of controlled nuclear fusion, volume 31, page 400. United Nations, 1958.DOI
  7. [7]V. D. Shafranov. “On magnetohydrodynamical equilibrium configurations.” Soviet Journal of Experimental and Theoretical Physics, 6:545, 1958.
  8. [8]Anand, H. et al. “Real-time plasma equilibrium reconstruction and shape control for the MAST Upgrade tokamak.” Nuclear Fusion, 64 086051, 2024. doi:10.1088/1741-4326/ad5c80.DOI
  9. [9]R. Albanese et al. “Design, implementation and test of the XSC extreme shape controller in JET”. Fusion Engineering and Design, 74(1):627-632, 2005. doi:10.1016/j.fusengdes.2005.06.290.DOI
  10. [10]M. Ariola et al. “Integrated Plasma Shape and Boundary Flux Control on JET Tokamak”. Fusion Science and Technology, 53(3):789-805, 2008. doi:10.13182/FST08-A1735.DOI
  11. [11]R. Albanese et al. “A MIMO architecture for integrated control of plasma shape and flux expansion for the EAST tokamak”. IEEE Conference on Control Applications, 611-616, 2016. doi:10.1109/CCA.2016.7587897.DOI
  12. [12]D. Corona et al. “Plasma shape control assessment for JT-60SA using the CREATE tools”. Fusion Engineering and Design, 146, 1773-1777, 2019. doi:10.1016/j.fusengdes.2019.03.032.DOI
  13. [13]A. Mele et al. “Design and implementation of a model-based hierarchical architecture for plasma shape control in the TCV tokamak”. Plasma Physics and Controlled Fusion, 67(6):065035, 2025. doi:10.1088/1361-6587/addeee.DOI
  14. [14]G. Ambrosino et al. “Design of the Plasma Position and Shape Control in the ITER Tokamak Using In-Vessel Coils”. IEEE Transactions on Plasma Science, 37(7):1324-1331, 2009. doi:10.1109/TPS.2009.2021476.DOI
  15. [15]P. Cavestany et al. “Real-Time Applicability of Emulated Virtual Circuits for Tokamak Plasma Shape Control”. IEEE Conference on Control Technology and Applications (CCTA), San Diego, CA, USA, 2025, pp. 826-831. doi:10.1109/CCTA53793.2025.11151371.DOI
  16. [16]A. Ross et al. “Real Time Virtual Circuits for plasma shape control via NN emulators”. In preparation. 2026.arxiv.org/abs/2605.14939
  17. [17]Bishop, C. M. (1995). Real-Time Control of a Tokamak Plasma Using Neural Networks.. Neural Computation, 7(1), 206–217. https://doi.org/10.1162/neco.1995.7.1.206
  18. [18]J.T. Wai et al. “Neural net modeling of equilibria in NSTX-U”. Nuclear Fusion, 62 086042, 2022. doi:10.1088/1741-4326/ac77e6.DOI
  19. [19]L.L. Lao et al. “Application of machine learning and artificial intelligence to extend EFIT equilibrium reconstruction”. Plasma Physics and Controlled Fusion, 64(7):074001, 2022. doi:10.1088/1361-6587/ac6fff.DOI
  20. [20]W. Rui et al. “Adaptive vertical position control system based on neural networks”. Nuclear Fusion, 66(2):026012, 2025. doi:10.1088/1741-4326/ae2693.DOI
  21. [21]J. Degrave et al. “Magnetic control of tokamak plasmas through deep reinforcement learning”. Nature 602, 414–419, 2022. doi:10.1038/s41586-021-04301-9.DOI
  22. [22]B. D. Tracey et al. “Towards practical reinforcement learning for tokamak magnetic control”. Fusion Engineering and Design, 52(9), 3698–3703, 2024. doi:10.1016/j.fusengdes.2024.114161.DOI
  23. [23]S. Kerboua-Benlarbi et al. “Magnetic control of WEST plasmas through deep reinforcement learning”. IEEE Transactions on Plasma Science 200, 114161, 2024. doi:10.1109/TPS.2024.3377811.DOI
  24. [24]A. Wang. “Enabling robust trajectory and control design for tokamaks with scientific machine learning”. PhD Thesis, Massachusetts Institute of Technology, 2026. doi: 1721.1/169933.DOI
  25. [25]Amorisо, N. C. (2024). FreeGSNKE: A python-based dynamic free-boundary toroidal plasma equilibrium solver.. Physics of Plasmas, 31(4), 042517. https://doi.org/10.1063/5.0188467
  26. [26]K. Pentland et al. “The FreeGSNKE Pulse Design Tool (FPDT): a computational framework for evolutive plasma scenario and control design”. In preparation. 2026. arXiv:2603.28513.arXiv
  27. [27]A. Agnello et al. “Emulation techniques for scenario and classical control design of tokamak plasmas”. Physics of Plasmas, 31(4):043901, 2024. doi:10.1063/5.0187822.DOI

Paper details

vlnc:W_ex86wede
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 4

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents