FreeGSNKE is a Python-based, open-source suite for solving a range of static and dynamic Grad–Shafranov equilibrium problems in tokamak fusion plasmas. It employs finite-difference discretisation with a Newton–Krylov nonlinear solver and is designed to integrate with modern machine learning and artificial intelligence libraries. FreeGSNKE supports static forward and inverse equilibrium calculations, as well as time-evolving simulations through a coupling between plasma equilibria and currents in poloidal field coils and passive conducting structures. The framework enables virtual circuit construction for plasma shape control and estimation of vertical growth rates for vertical displacement event modelling. The codebase is tokamak-agnostic, well documented, and includes worked examples. The solvers have been validated against MAST-U equilibria, with ongoing development including a JAX-based implementation of the core solvers and a coupling with the current diffusion equation. Recent work has focused on coupling FreeGSNKE's evolutive solver with a newly developed modular and customisable virtual Plasma Control System (PCS). Using user-defined waveforms and control parameters, the PCS employs feedback and feedforward techniques to regulate plasma current, position, and shape throughout a simulation, while enforcing machine safety limits on coil currents and voltages. This combined framework, termed the FreeGSNKE Pulse Design Tool (FPDT), enables predictive modelling and virtual testing of advanced plasma configurations, new control schemes, and the impact of diagnostic uncertainties on plasma evolution and performance. This significantly reduces reliance on costly and iterative physical experiments. Although inspired by the MAST-U PCS architecture, the virtual PCS is inherently machine-independent, supporting the implementation of diverse control algorithms tailored to specific tokamaks. The FPDT has been validated by re-simulating previously executed MAST-U discharges and comparing nonlinear and piecewise-linear simulation modes. Under modest uncertainties in the initial equilibrium, current density profiles, and plasma resistivity, the FPDT achieves strong quantitative agreement with target control waveforms and real-time MAST-U measurements. With this extension to the FreeGSNKE open-source suite, we aim to foster greater reproducibility and collaboration in plasma modelling and control research.

I’ll cover continuing progress on FreeGSNKE, a fully open-source Python code for simulating both static and evolutive tokamak plasma equilibria. Alongside existing static forward/inverse and evolutive forward solvers, the FreeGSNKE ecosystem now also includes capabilities for calculating virtual circuits and vertical growth rates. Recent work has focused on equipping FreeGSNKE with a Plasma Control System (PCS) to enable plasma position, shape, and current control. While inspired by the controllers used on the MAST Upgrade tokamak, the framework can be customised to any axisymmetric machine. The resulting FreeGSNKE Pulse Design Tool (FPDT) supports in-silico scenario design, control design, and control validation. We aim to open source this tool as part of the FreeGSNKE ecosystem.

The plasma control system (PCS) of the MAST Upgrade (MAST-U) tokamak has been in operational use during the first five experimental campaigns. We present an overview of the recent improvements to the capabilities of the control system, key experimental results that were enabled by the improved control capabilities and the recent hardware platform upgrade that paves the way to building on these capabilities for future experimental campaigns. Versatile and reliable core and divertor shape controllers and a magnetic reconstruction step for gap feedback control have been developed collaboratively and enabled the development of plasma scenarios for the exploitation of conventional and alternative divertor configurations and studies of plasma exhaust power load mitigation. Improvements in control of the shape of the plasma boundary have enabled detailed studies of the effects of shaping on confinement and stability in the core and edge pedestal regions, and investigations of time-dependent behaviour during vertical oscillations around a connected double null topology. MAST-U achieved access to negative triangularity for the first time on a spherical tokamak, which will significantly contribute to studies into stability boundaries and ELM-free operation. Plasma electron density control further facilitated repeatable plasma scenarios and enabled reactor-relevant studies into locked mode mitigation and L-H power threshold. A collaboratively developed detachment front control was achieved and demonstrated the feasibility of controlling the spatial extent of the detached neutral buffer. Building on the unique character of MAST-U being a double-null device, recently this capability was extended to allow the simultaneous control of core density and the detachment fronts independently in the lower and upper divertor chambers, which opens avenues for studying up-down exhaust symmetries. To support future capabilities, significant infrastructure upgrades have been undertaken. The PCS hardware and software infrastructure have been upgraded to a 64-bit architecture during the 5 th experimental campaign, increasing the number of available processing cores from 4 to 32. The hardware has built-in support for low-latency data transfer over User Datagram Protocol (UDP) based protocols to enable communication over a Real-time Data Network built from commercial, off-the-shelf components. These enhancements will enable the integration of machine learning for shape control, continued exploration of advanced control techniques for density and detachment control and will provide capacity for the growing control capability demands of the MAST-U experimental programme, such as plasma current profile control involving Electron Bernstein Wave (EBW) heating that is scheduled to be installed in the upcoming engineering break.

Reliable position and shape control in tokamak plasmas requires accurate real-time regulation of several strongly coupled shape parameters. The control vectors disentangling these couplings, hereafter referred to as Virtual Circuits (VCs), enable independent shape parameter control for a specific equilibrium. These VCs are conventionally computed in advance of an experiment using a small number of reference Grad–Shafranov equilibria and used to control the plasma across different pre-set time intervals. While effective near the reference equilibrium, this approach can lead to degraded performance when the plasma departs from the target trajectory and complicates the design of robust control strategies for rapidly evolving plasma configurations. This work presents the construction of emulators for VCs that satisfy the requirements for use in real-time control on MAST Upgrade (MAST-U). We develop an extensive library of several million simulated magnetic equilibria, covering a substantial portion of the MAST-U operational space, and use it to train neural network emulators to predict shape parameters given plant parameters. These are differentiable functions, whose Jacobians can be computed with millisecond latency, providing accurate VCs suited for real-time shape control. We perform extensive verification of the emulated VCs by applying small, independent shifts to a diverse set of several thousand equilibria and comparing the resulting target displacements with those obtained using finite-difference Jacobians of Grad-Shafranov solutions. The neural-network-based approach delivers high accuracy and orthogonality across the full equilibrium space. This work establishes the physical validity of emulated VCs as a scalable and general alternative to pre-set schedules of pre-computed VCs.

Spherical tokamaks are an economically attractive concept for a fusion power plant, as they operate at high $\beta_N$. The low aspect ratio of spherical tokamaks naturally facilitates operation of highly elongated ($\kappa \approx 2.5$) plasmas which leads to high non-inductive bootstrap fractions. Operation at high elongation can pose additional vertical stability challenges, however coupling high triangularity ($\delta$) and high elongation naturally facilitates operation at high $\beta_p$, which is stabilizing for low toroidal mode number $n$ MHD instabilities. High triangularity has a strong stabilizing effect on ideal ballooning modes (IBMs) which impose a limit on the maximum achievable $\beta_N$. This work analyses the MHD stability of the recently developed high elongation MAST-U plasma scenario, achieving a plasma current flat top shaping target of $\kappa = 2.5$ and a standard "medium" elongation scenario at $\kappa = 2$. A triangularity scan of $\delta = 0.4$–$0.55$ was performed in both the medium and high $\kappa$ cases. At both medium and high elongation $\beta_N \approx 2.5$–$3$ is achieved with $\beta_e \approx 5\%$. Low $n$ MHD activity is reduced in the early plasma current flat top in the high $\kappa$ case due to a slow $\beta_N$ evolution compared to the medium $\kappa$ scenario. Finally, the impact of the plasma current ramp rate in the high $\kappa$ scenario is assessed. The high $\kappa$ fast ramp rate scenario exhibits a strongly reverse shear $q$ profile which leads to an elevated $q_{min}$, delaying the onset of a confinement degrading 2/1 tearing instability. The magnetic shear $\hat{s} = \frac{r}{q}\frac{dq}{dr}$ evolution is compared in the medium and high $\kappa$ cases and calculation of the ideal ballooning stability (IBM) is performed using the Pyrokinetics workflow.

The deployment of advanced, AI-enabled control algorithms in tokamak experiments requires robust integration with existing plasma control system (PCS) architectures and extensive pre-experimental validation. In this contribution, we describe the implementation of neural-network-emulated virtual circuits for plasma shape control within the MAST Upgrade (MAST-U) PCS environment. The neural network models are transformed into a real-time-suitable data structure and deployed via a C++ inference server that interfaces with the General Atomics Plasma Control System (GAPCS), a real-time control framework. An integration layer transfers real-time GAPCS signals into the neural network input space, executes inference within real-time constraints, and returns the resulting Jacobian coefficients to the control system. These coefficients are then used to assemble and invert the virtual circuit matrices and compute updated coil current requests for real-time actuation. We discuss the overall software architecture and data pathways underpinning this integration, the timing and latency constraints encountered, and the challenges of interfacing modern AI components with real-time control infrastructure. Emphasis is placed on the validation workflow and best practices adopted to ensure confidence in the solution prior to experimental deployment. This work demonstrates a practical pathway for introducing AI-based control components into operational fusion control systems, with direct relevance for upcoming MAST-U experiments and future devices.