Egemen Kolemen

A centralized platform for fusion-related machine learning workflows.

A centralized platform for fusion-related machine learning workflows.

Abstract We present the design and application of a general algorithm for Prediction And Control using MAchiNe learning ( PACMAN ) in DIII-D. Machine learning (ML)-based predictors and controllers have shown great promise in achieving regimes in which traditional controllers fail, such as tearing mode (TM) free scenarios, ELM-free scenarios and stable advanced tokamak conditions. The architecture presented here was deployed on DIII-D to facilitate the end-to-end implementation of advanced control experiments, from diagnostic processing to final actuation commands. This paper describes the detailed design of the algorithm and explains the motivation behind each design point. We also describe several successful ML control experiments in DIII-D using this algorithm, including a reinforcement learning controller targeting advanced non-inductive plasmas, a wide-pedestal quiescent H-mode ELM predictor, an Alfvén Eigenmode controller, a Model Predictive Control plasma profile controller and a state-machine TM predictor-controller. There is also discussion on guiding principles for real-time ML controller design and implementation.

We present a fast, spectrally (exponentially) accurate, automatically differentiable bounce-averaging algorithm that is used to simplify kinetic models. Using this algorithm, implemented in the DESC stellarator optimisation suite, we can perform efficient optimisation of many objectives to improve stellarator performance, such as the effective ripple $\epsilon _{\mathrm{eff}}$ metric for the neoclassical transport coefficient in the low collisionality regime and proxies for energetic particle confinement. For the first time, we optimise a finite-beta stellarator to directly reduce neoclassical ripple transport using reverse-mode differentiation. This ensures the computational cost of differentiation is independent of the number of controllable parameters.

A real-time safe multi-threading library was developed on the DIII-D plasma control system to optimize the real-time TORBEAM and real-time STRIDE physics codes. These physics codes are crucial for future fusion power plant operation as they provide information about electron cyclotron wave propagation and heating as well as inform about ideal plasma stability limits. The multi-threading library uses a single Manager thread to coordinate Worker threads and distribute computational work evenly to speed up real-time computations. The flexible nature of the library allows it to work with multiple physics codes and with different numbers of total Worker thread counts. The real-time TORBEAM code executed consistently in under 20 ms while the real-time STRIDE code computes in 100 ms. The multi-threading library developed in this work can be applied to other real-time physics-based codes that will be crucial for the next generation of fusion devices. • Developed multi-threading library for DIII-D real-time plasma control system. • Utilized library for efficient deployment of real-time TORBEAM ray tracing code. • Controlled gyrotrons mirror angles to track dynamic ECH targets. • Calculated MHD stability of DIII-D plasmas in real-time with STRIDE code.

Artificial intelligence models have demonstrated strong predictive capabilities for various instabilities in fusion devices such as Tokamaks, including tearing modes (TM), edge localized modes, and disruptive events, but their opaque nature raises concerns about safety and trustworthiness when applied to fusion power plants. Here, we present a physics-based interpretation framework using a TM prediction model as a demonstration that is validated through a dedicated DIII-D TM avoidance experiment. By applying Shapley analysis, we identify how profiles such as rotation, temperature, and density contribute to the model's prediction of TM stability. Our analysis shows that in our experimental scenario, core electron temperature and rotation peaking play the primary role in TM stability, while density changes have smaller effects on stability. We show that off-axis ion temperature stabilizes TMs, suggesting that off-axis neutral beam heating can further stabilize this scenario. This work presents a generalizable ML-based event prediction methodology, from training to physics-driven interpretation, bridging the gap between physics understanding and opaque ML models.

Abstract While artificial intelligence (AI) has been promising for fusion control, its inherent black-box nature will make compliant implementation in regulatory environments a challenge. This study implements and validates a real-time AI-enabled linear and interpretable control system for successful divertor detachment control with the DIII-D lower divertor camera. Using D 2 gas, we demonstrate successful feedback divertor detachment control with a mean absolute difference of 2% from the target for both detachment and reattachment. This automatic training and linear processing framework can be extended to any image-based diagnostic for future fusion reactors.

Abstract We report the experimental results of controlling the pedestal-top electron density by applying resonant magnetic perturbation (RMP) with in-vessel control coils and the main gas puff in the 2024-2025 KSTAR experimental campaign. The density is reconstructed using a parameterized <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" overflow="scroll"> <mml:mrow> <mml:msub> <mml:mi>ψ</mml:mi> <mml:mrow> <mml:mtext mathvariant="italic">N</mml:mtext> </mml:mrow> </mml:msub> </mml:mrow> </mml:math> grid and five channels of line-averaged density measured by the two-colored interferometer (TCI). The reconstruction procedure is accelerated by deploying a multi-layer perceptron to run in approximately 120 µ s and is sufficiently fast for real-time control. A proportional-integral controller was adopted, with the controller gains estimated from the system identification procedure. The experimental results demonstrate that the developed controller can follow a dynamic target while exclusively using both actuators. The absolute percentage errors between the electron density at <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" overflow="scroll"> <mml:mrow> <mml:msub> <mml:mi>ψ</mml:mi> <mml:mtext mathvariant="italic">N</mml:mtext> </mml:msub> <mml:mo>=</mml:mo> <mml:mn>0.89</mml:mn> </mml:mrow> </mml:math> and the target were approximately 1.5% median and a 2.5% average, respectively. The developed controller can even lower the density by using the pump-out mechanism under RMP, and it can follow a more dynamic range of density targets than a single actuator controller. The developed controller will enable experimental scenario exploration within a shot by dynamically setting the density target or maintaining a constant electron density within a discharge.

To better understand the dependence of the magnetic field structure in the plasma edge on the plasma boundary shape, in the context of X-point and island divertor designs, we define and develop a class of stellarators called umbilic stellarators. These equilibria are characterised by a single continuous high-curvature edge on the plasma boundary that goes around multiple times toroidally before meeting itself. We develop a technique that allows us to simultaneously optimise the plasma boundary along with a curve lying on the boundary on which we impose a high curvature while imposing omnigenity – a property of the magnetic field that ensures trapped particle confinement throughout the plasma volume. We find that umbilic stellarators naturally tend to favour piecewise omnigenity instead of omnigenity with a specific helicity. After generating omnigenous umbilic stellarators, we design coil sets for some of them and explore the fieldline structure in the edge and its sensitivity to small fluctuations in the plasma. Finally, using single-stage optimisation, we simultaneously modify the plasma and coil shape and propose an experiment to modify an existing tokamak to a finite- $\beta$ stellarator using this technique and explore a potentially simpler way to convert a limited tokamak into a diverted stellarator.

Achieving reliable real-time control of tokamak plasmas is essential for sustaining high-performance operation in next-generation fusion reactors. A major challenge is the accurate and timely prediction of edge-localized modes (ELMs), especially in high-confinement regimes such as wide-pedestal quiescent H-mode. We present a hardware-accelerated machine learning (ML) inference system integrated into the RTSTAB processing node of the DIII-D real-time diagnostic and control infrastructure. The system uses an AMD/Xilinx KCU1500 FPGA to enable ultra low latency plasma state classification and ELM forecasting. Input features come from real-time Beam Emission Spectroscopy (BES), and the ML model is implemented as a dense neural network using the SLAC Neural Network Library (SNL). A key capability is SNL dynamic parameter loading, which allows on-the-fly updates of neural network weights and biases without hardware resynthesis. This enables multiple classification tasks on a single FPGA design and supports adaptive control strategies that respond to evolving plasma conditions. By decoupling inference from fixed-weight configurations, the system supports continuous model refinement and seamless task switching during live operation. The SNL-based inference engine is fully integrated with the FPGA in the DIII-D RTSTAB Plasma Control System (PCS), improving ELM avoidance, confinement, and operational stability. These results show the feasibility of embedding dynamically reconfigurable FPGA-based ML inference into real-time fusion diagnostic pipelines, providing a scalable and resilient path toward intelligent and autonomous plasma control in future magnetic confinement fusion devices.

Understanding and controlling plasma behavior in fusion devices requires integrating information across multiple diagnostics and actuator systems, each capturing different aspects of the complex plasma state. Traditional approaches often analyze these modalities in isolation, missing important cross-modal dependencies that can provide deeper insight into plasma dynamics and control. We propose a self-supervised learning approach based on a large-scale masked autoencoder architecture that is designed to uncover a unified representation of plasma state and control inputs across multiple diagnostic modalities.The learned representation can be used to identify subtle patterns across modalities that may be invisible to individual diagnostic analysis, potentially revealing new physics insights about plasma behavior under different control strategies. The model’s capacity to reconstruct masked diagnostic data also demonstrates its potential for robust operation during diagnostic failure and for achieving enhanced resolution beyond the limits of single diagnostic models.[1] Furthermore, the intermediate feature space provides a new holistic representation of the plasma state which can be leveraged for transfer learning of downstream tasks such as mode identification, instability prediction, scenario design and optimal control synthesis. [1] Jalalvand, et al., 2024. “Multimodal Super-Resolution: Discovering hidden physics and its application to fusion plasmas” *This material is based upon work supported by the U.S. Department of Energy, Office of Science, Office of Fusion Energy Sciences, using the DIII-D National Fusion Facility, a DOE Office of Science user facility, under Award(s) DE-FC02-04ER54698 and DE-SC0024527. The authors also gratefully acknowledge financial support from the Princeton Laboratory for Artificial Intelligence under Award 2025-97.

Understanding and controlling plasma behavior in fusion devices requires integrating information across multiple diagnostics and actuator systems, each capturing different aspects of the complex plasma state. Traditional approaches often analyze these modalities in isolation, missing important cross-modal dependencies that can provide deeper insight into plasma dynamics and control. We propose a self-supervised learning approach based on a large-scale masked autoencoder architecture that is designed to uncover a unified representation of plasma state and control inputs across multiple diagnostic modalities.The learned representation can be used to identify subtle patterns across modalities that may be invisible to individual diagnostic analysis, potentially revealing new physics insights about plasma behavior under different control strategies. The model’s capacity to reconstruct masked diagnostic data also demonstrates its potential for robust operation during diagnostic failure and for achieving enhanced resolution beyond the limits of single diagnostic models.[1] Furthermore, the intermediate feature space provides a new holistic representation of the plasma state which can be leveraged for transfer learning of downstream tasks such as mode identification, instability prediction, scenario design and optimal control synthesis. [1] Jalalvand, et al., 2024. “Multimodal Super-Resolution: Discovering hidden physics and its application to fusion plasmas” *This material is based upon work supported by the U.S. Department of Energy, Office of Science, Office of Fusion Energy Sciences, using the DIII-D National Fusion Facility, a DOE Office of Science user facility, under Award(s) DE-FC02-04ER54698 and DE-SC0024527. The authors also gratefully acknowledge financial support from the Princeton Laboratory for Artificial Intelligence under Award 2025-97.

Abstract One of the common tasks required for designing new plasma scenarios or evaluating capabilities of a tokamak is to design the desired equilibria using a Grad-Shafranov (GS) equilibrium solver. However, most standard equilibrium solvers are time-independent and do not include dynamic effects such as plasma current flux consumption, induced vessel currents, or voltage constraints. Another class of tools, plasma equilibrium evolution simulators, do include time-dependent effects. These are generally structured to solve the forward problem of evolving the plasma equilibrium given feedback-controlled voltages. In this work, we introduce GSPulse, a novel algorithm for equilibrium trajectory optimization, that is more akin to a pulse planner than a pulse simulator. GSPulse includes time-dependent effects and solves the inverse problem: given a user-specified set of target equilibrium shapes, as well as limits on the coil currents and voltages, the optimizer returns trajectories of the voltages, currents, and achievable equilibria. This task is useful for scoping performance of a tokamak and exploring the space of achievable pulses. The computed equilibria satisfy both Grad-Shafranov force balance and axisymmetric circuit dynamics. The optimization is performed by restructuring the free-boundary equilibrium evolution equations into a form where it is computationally efficient to optimize the entire dynamic sequence. GSPulse can solve for hundreds of equilibria simultaneously within a few minutes. GSPulse has been validated against NSTX-U and MAST-U experiments and against SPARC feedback control simulations, and is being used to perform scenario design for SPARC. The computed trajectories can be used as feedforward inputs that are connected to the feedback controller to inform and improve feedback performance. The code for GSPulse is available open-source at github.com/jwai-cfs/GSPulse_public .

Liquid metal (LM) plasma facing components (PFCs) are an essential component of conceptual thermonuclear fusion reactors due to their self healing and low recycling properties. Lithium is the primary choice for LMPFCs, as it is highly soluble for hydrogen species and has a low melting point. The retrieval of tritium from the Li+H/D/T compound is an unavoidable step in the development of economically feasible fusion reactors. Inductively heated distillation stacks are the leading choice for retrieving trapped tritium and returning clarified lithium back to LMPFCs. Separation of precipitated LiH crystals from the lithium utilizing existing fields will enhance the distillation yield of tritium, minimize the volume of lithium inventory, and reduce the net MHD drag from pumping LM out of the coil region. We have constructed and experimented with a modular, jxB enhanced, hydrocyclone running in a steady state, closed loop configuration, dubbed the Magneto Cyclone Centrifuge (MC 2 ). Galinstan liquid metal was amalgamated with copper dust and fed to the MC 2 while aligned axially in the magnetic field of the PPPL's LMX-U facility. Separation efficiency of the device was determined using x-ray fluorescence, which measured the concentration of copper in Galinstan sampled during operation. Efficiency and consistency increased with radially applied currents, improving device performance where magnetic fields prohibit the fast, high pressure traditionally required at the inlets of conventional hydrocyclones.

Fusion diagnostics provide abundant information on the state of a tokamak, which can be used for studying physics or for control. Critical information can be found in resonant modes such as magnetic modes. We rely on experts to identify these modes and their consequences, or by exploiting geometric properties of the tokamak for generating structures. Recent advances in computer vision and acoustic modelling have shown the potential for this to be self-learned. We apply these principles directly on individual diagnostics to achieve preliminary results in automatic mode detection and learning. Two methods will be explored using Mirnov coil and CO2 interferometer diagnostics. First, the eigenvalues of visual similarities will be taken with deep self-supervised feature extraction [1]. Second, an internally represented codebook of modes will be used for reconstruction using a foundation model autoencoder [2]. These methods seek to encode semantic modal information, create an automatic mode detection scheme for shot analysis, and advance knowledge about new types of modes or interactions between these modes. [1] K. E. J. Olofsson et al., Array magnetics modal analysis for the DIII-D tokamak based on localized time-series modelling, Plasma Phys. Control. Fusion 56, 095012 (2014). [2] A. Jalalvand, M. Curie, S. Kim, P. Steiner, J. Seo, Q. Hu, A. O. Nelson, and E. Kolemen, Diag2Diag: Multimodal Super-Resolution Diagnostics for Physics Discovery with Application to Fusion, arXiv:2405.05908.

Divertorlets is a liquid metal divertor concept that divides the target into small, recirculating sections to reduce plasma exposure time and overheating. To explore the full parameter design space, various designs are rapidly prototyped with 3D-printed PLA and tested with galinstan in the LMX-U facility at PPPL. Promising designs are then manufactured via 3D printing using liquid lithium compatible materials such as stainless steel and tungsten for testing under reactor-like conditions. Operation is driven by applied currents and magnetic fields, with research focused on optimizing the current density distribution to control the resulting velocity and temperature profiles. Key considerations for reactor performance include reducing surface fluctuations and splashing, utilizing stabilizing meshes, and managing thermoelectric MHD effects from the surface heat flux. This integrated approach of iterative design and material progression advances the divertorlets concept toward a robust solution for a fusion reactor.

Fusion diagnostics provide abundant information on the state of a tokamak, which can be used for studying physics or for control. Critical information can be found in resonant modes such as magnetic modes. We rely on experts to identify these modes and their consequences, or by exploiting geometric properties of the tokamak for generating structures. Recent advances in computer vision and acoustic modelling have shown the potential for this to be self-learned. We apply these principles directly on individual diagnostics to achieve preliminary results in automatic mode detection and learning. Two methods will be explored using Mirnov coil and CO2 interferometer diagnostics. First, the eigenvalues of visual similarities will be taken with deep self-supervised feature extraction [1]. Second, an internally represented codebook of modes will be used for reconstruction using a foundation model autoencoder [2]. These methods seek to encode semantic modal information, create an automatic mode detection scheme for shot analysis, and advance knowledge about new types of modes or interactions between these modes. [1] K. E. J. Olofsson et al., Array magnetics modal analysis for the DIII-D tokamak based on localized time-series modelling, Plasma Phys. Control. Fusion 56, 095012 (2014). [2] A. Jalalvand, M. Curie, S. Kim, P. Steiner, J. Seo, Q. Hu, A. O. Nelson, and E. Kolemen, Diag2Diag: Multimodal Super-Resolution Diagnostics for Physics Discovery with Application to Fusion, arXiv:2405.05908.

Divertorlets is a liquid metal divertor concept that divides the target into small, recirculating sections to reduce plasma exposure time and overheating. To explore the full parameter design space, various designs are rapidly prototyped with 3D-printed PLA and tested with galinstan in the LMX-U facility at PPPL. Promising designs are then manufactured via 3D printing using liquid lithium compatible materials such as stainless steel and tungsten for testing under reactor-like conditions. Operation is driven by applied currents and magnetic fields, with research focused on optimizing the current density distribution to control the resulting velocity and temperature profiles. Key considerations for reactor performance include reducing surface fluctuations and splashing, utilizing stabilizing meshes, and managing thermoelectric MHD effects from the surface heat flux. This integrated approach of iterative design and material progression advances the divertorlets concept toward a robust solution for a fusion reactor.

Unlike their axisymmetric counterparts, stellarators do not inherently guarantee particle confinement due to their non-axisymmetric nature. To make stellarators viable candidates for magnetic confinement fusion reactors, researchers are exploring the quasi-symmetric magnetic fields – a subset of omnigeneous fields– as optimization targets. Traditional approaches rely on surrogate metrics such as effective ripple to assess confinement without explicitly simulating particle trajectories. However, recent advances in GPU computing enable the simultaneous integration of thousands of particle trajectories, making direct trajectory-based optimization feasible. In this work, we implement the guiding center trajectory model in the DESC library. DESC's native GPU support and direct access to magnetic field data, without interpolation, enable efficient optimization of 3D MHD equilibria. To ensure numerical differentiability, we employ gradient clipping to exclude non-confined trajectories from backpropagation. This trajectory-based approach, long limited by computational resources, opens new opportunities to directly optimize for particle confinement and deepen our understanding of confinement physics in stellarators.

Liquid metal (LM) plasma facing components (PFCs) are an essential component of conceptual thermonuclear fusion reactors due to their self healing and low recycling properties. Lithium is the primary choice for LMPFCs, as it is highly soluble for hydrogen species and has a low melting point. The retrieval of tritium from the Li+H/D/T compound is an unavoidable step in the development of economically feasible fusion reactors. Inductively heated distillation stacks are the leading choice for retrieving trapped tritium and returning clarified lithium back to LMPFCs. Separation of precipitated LiH crystals from the lithium utilizing existing fields will enhance the distillation yield of tritium, minimize the volume of lithium inventory, and reduce the net MHD drag from pumping LM out of the coil region. We have constructed and experimented with a modular, jxB enhanced, hydrocyclone running in a steady state, closed loop configuration, dubbed the Magneto Cyclone Centrifuge (MC 2 ). Galinstan liquid metal was amalgamated with copper dust and fed to the MC 2 while aligned axially in the magnetic field of the PPPL's LMX-U facility. Separation efficiency of the device was determined using x-ray fluorescence, which measured the concentration of copper in Galinstan sampled during operation. Efficiency and consistency increased with radially applied currents, improving device performance where magnetic fields prohibit the fast, high pressure traditionally required at the inlets of conventional hydrocyclones.

Unlike their axisymmetric counterparts, stellarators do not inherently guarantee particle confinement due to their non-axisymmetric nature. To make stellarators viable candidates for magnetic confinement fusion reactors, researchers are exploring the quasi-symmetric magnetic fields – a subset of omnigeneous fields– as optimization targets. Traditional approaches rely on surrogate metrics such as effective ripple to assess confinement without explicitly simulating particle trajectories. However, recent advances in GPU computing enable the simultaneous integration of thousands of particle trajectories, making direct trajectory-based optimization feasible. In this work, we implement the guiding center trajectory model in the DESC library. DESC's native GPU support and direct access to magnetic field data, without interpolation, enable efficient optimization of 3D MHD equilibria. To ensure numerical differentiability, we employ gradient clipping to exclude non-confined trajectories from backpropagation. This trajectory-based approach, long limited by computational resources, opens new opportunities to directly optimize for particle confinement and deepen our understanding of confinement physics in stellarators.

We present new capabilities in DESC [2,3,4,5] for the 3D stellarator equilibrium experimental reconstruction problem. The 3D equilibrium reconstruction problem conventionally requires many expensive 3D equilibrium solves in order to acquire the derivative information necessary for matching the synthetic diagnostic signals to the measured signals [1]. DESC's automatic differentiation enables methods that use fewer solves per reconstruction iteration, resulting in more efficient, faster optimization. Results will be shown using these capabilities to perform reconstruction and compare DESC to other reconstruction codes and literature. [1] Hanson et. al., NF (2009). [2] Dudt, D. & Kolemen, E. PoP (2020). [3] Panici, D. et al. JPP (2023). [4] Conlin, R. et al. JPP (2023). [5] Dudt, D. et al. JPP (2023).

We present new capabilities in DESC [2,3,4,5] for the 3D stellarator equilibrium experimental reconstruction problem. The 3D equilibrium reconstruction problem conventionally requires many expensive 3D equilibrium solves in order to acquire the derivative information necessary for matching the synthetic diagnostic signals to the measured signals [1]. DESC's automatic differentiation enables methods that use fewer solves per reconstruction iteration, resulting in more efficient, faster optimization. Results will be shown using these capabilities to perform reconstruction and compare DESC to other reconstruction codes and literature. [1] Hanson et. al., NF (2009). [2] Dudt, D. & Kolemen, E. PoP (2020). [3] Panici, D. et al. JPP (2023). [4] Conlin, R. et al. JPP (2023). [5] Dudt, D. et al. JPP (2023).

Abstract Omnigenity is a property of a magnetic field which ensures confinement of trapped particles. It is a necessary requirement for any high-performance stellarator. After creating an omnigenous equilibrium, one must also ensure reduced transport resulting from kinetic and magnetohydrodynamic (MHD) instabilities. To this end, we leverage the GPU-accelerated DESC optimization suite, which is used to design stable, finite- β omnigenous equilibria with poloidal, toroidal, and helical symmetry, achieving Mercier, ideal ballooning, and as a consequence, improved kinetic ballooning stability. We discover stellarators with second stability, a regime of large pressure gradient where an equilibrium becomes ideal ballooning stable, and demonstrate and explore both using theory and gyrokinetic simulations the connection between ideal and kinetic ballooning stability.

Abstract High fidelity kinetic equilibria are crucial for tokamak modeling and analysis. Manual workflows for constructing kinetic equilibria are time consuming and subject to user error, motivating development of automated equilibrium reconstruction tools to provide accurate and consistent reconstructions for downstream physics analysis. These automated tools also provide access to kinetic equilibria at large database scales, which enables the quantification of general uncertainties arising from equilibrium reconstruction techniques. In this paper, we compare a large database of DIII-D kinetic equilibria generated manually by physics experts to equilibria from automated kinetic reconstruction tools, assessing the impact of reconstruction method on equilibrium parameters and resulting magnetohydrodynamic stability calculations. We find agreement among scalar parameters, whereas profile quantities, such as the bootstrap current, show larger disagreements. We analyze ideal kink and classical tearing stability with DCON and STRIDE respectively, finding that the kink stability calculation is generally more robust than the tearing index <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" overflow="scroll"> <mml:mrow> <mml:msup> <mml:mi mathvariant="normal">Δ</mml:mi> <mml:mi>′</mml:mi> </mml:msup> </mml:mrow> </mml:math> calculation. We find that in 90% of cases, both kink stability classifications are unchanged between the manual expert and automated kinetic equilibria.

Understanding complex physical systems often requires integrating data from multiple diagnostics, each with limited resolution or coverage. We present a machine learning framework that reconstructs synthetic high-temporal-resolution data for a target diagnostic using information from other diagnostics, without direct target measurements during the inference. This multimodal super-resolution technique improves diagnostic robustness and enables monitoring even in case of measurement failures or degradation. Applied to fusion plasmas, our method targets edge-localized modes (ELMs), which can damage plasma-facing materials. By reconstructing super-resolution Thomson Scattering data from complementary diagnostics, we uncover fine-scale plasma dynamics and validate the role of resonant magnetic perturbations (RMPs) in ELM suppression through magnetic island formation. The approach provides new observation supporting the plasma profile flattening due to these islands. Our results demonstrate the framework’s ability to generate high-fidelity synthetic diagnostics, offering a powerful tool for ELM control development in future reactors like ITER. The approach is broadly transferable to other domains facing sparse, incomplete, or degraded diagnostic data, opening new avenues for discovery. Sensor failures and limited resolution challenge many complex systems. Here, authors develop a multimodal AI method to generate super-resolution of a sensor using other available sensors in the system, revealing hidden dynamics in fusion plasmas and enabling cost-effective, high-resolution diagnostics.

Data-driven acceleration of scientific computing workflows has been a high-profile aim of machine learning (ML) for science, with numerical simulation of transient partial differential equations (PDEs) being one of the main applications. The focus thus far has been on methods that require classical simulations to train, which when combined with the data-hungriness and optimization challenges of neural networks has caused difficulties in demonstrating a convincing advantage against strong classical baselines. We consider an alternative paradigm in which the learner uses a classical solver's own data to accelerate it, enabling a one-shot speedup of the simulation. Concretely, since transient PDEs often require solving a sequence of related linear systems, the feedback from repeated calls to a linear solver such as preconditioned conjugate gradient (PCG) can be used by a bandit algorithm to online-learn an adaptive sequence of solver configurations (e.g. preconditioners). The method we develop, PCGBandit, is implemented directly on top of the popular open source software OpenFOAM, which we use to show its effectiveness on a set of fluid and magnetohydrodynamics (MHD) problems.

Abstract Quasi-symmetry is a guiding principle to modern stellarator optimization for improved plasma confinement. However, the robustness of optimized configurations, which can be crucial for maintaining performance under diverse engineering constraints and practical limitations, has received relatively little attention. Here we present various case studies on this robustness, by investigating variations in <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" overflow="scroll"> <mml:mrow> <mml:mn>1</mml:mn> <mml:mrow> <mml:mo>/</mml:mo> </mml:mrow> <mml:mi>ν</mml:mi> </mml:mrow> </mml:math> neoclassical transport when plasma configurations are smoothly altered across distinct optimized targets. These targets, optimized from different families—quasi-axisymmetric, quasi-helical, and quasi-isodynamic—are approximately matched in major radius as part of a flexible stellarator design. Our study shows that an optimized target does not always represent a local minimum in transport and that the robustness of a local minimum when present can vary significantly. Furthermore, there are configurations which belong to no established families but have transport levels as low as those of optimized targets. These results highlight the importance of conducting extended searches with key parametric variations around optimized configurations, to ensure its robustness as well as flexibility if desired.

Abstract This paper reports on the extension of our amplitude-based resonant magnetic perturbation (RMP) edge localized mode (ELM) controller to support phasing control (relative toroidal phases of RMP waveforms between rows of coils), multiple toroidal mode numbers, and new ‘jump’ and ‘probing’ strategies, all deployed on KSTAR and DIII-D. By treating the control algorithm as device-independent and adjusting only the real-time interfaces to sensors and power supplies, we have confirmed that the same finite state machine—based feedback logic can be ported between machines with minor modifications. In experiments using n = 2 RMPs on KSTAR and n = 3 on DIII-D, the controller successfully modulated RMP amplitudes in real time to sustain ELM suppression while minimizing confinement degradation. Phasing control broadened the suppression window, as it permitted the system to avoid locked-mode regions and safely access ELM-free conditions. A rotating RMP phasing scheme, integrated into the same framework, distributes divertor heat loads more uniformly, making it a promising strategy for protecting plasma-facing components during long discharges. New ‘jump’ and ‘probing’ techniques demonstrate the possibility for the controller to preempt imminent ELMs and refine the minimum required RMP amplitude without returning to ELMy conditions. Taken together, these upgrades enable extended ELM-free operation while mitigating confinement degradation, providing a practical framework for real-time ELM control in future high-performance tokamaks.

One of the common tasks required for designing new plasma scenarios or evaluating capabilities of a tokamak is to design the desired equilibria using a Grad-Shafranov (GS) equilibrium solver. However, most standard equilibrium solvers are time-independent and do not include dynamic effects such as plasma current flux consumption, induced vessel currents, or voltage constraints. Another class of tools, plasma equilibrium evolution simulators, do include time-dependent effects. These are generally structured to solve the forward problem of evolving the plasma equilibrium given feedback-controlled voltages. In this work, we introduce GSPulse, a novel algorithm for equilibrium trajectory optimization, that is more akin to a pulse planner than a pulse simulator. GSPulse includes time-dependent effects and solves the inverse problem: given a user-specified set of target equilibrium shapes, as well as limits on the coil currents and voltages, the optimizer returns trajectories of the voltages, currents, and achievable equilibria. This task is useful for scoping performance of a tokamak and exploring the space of achievable pulses. The computed equilibria satisfy both Grad-Shafranov force balance and axisymmetric circuit dynamics. The optimization is performed by restructuring the free-boundary equilibrium evolution (FBEE) equations into a form where it is computationally efficient to optimize the entire dynamic sequence. GSPulse can solve for hundreds of equilibria simultaneously within a few minutes. GSPulse has been validated against NSTX-U and MAST-U experiments and against SPARC feedback control simulations, and is being used to perform scenario design for SPARC. The computed trajectories can be used as feedforward inputs to inform and improve feedback performance. The code for GSPulse is available open-source at https://github.com/jwai-cfs/GSPulse_public.

We report the experimental results of controlling the pedestal-top electron density by applying resonant magnetic perturbation with the in-vessel control coils and the main gas puff in the 2024-2025 KSTAR experimental campaign. The density is reconstructed using a parametrized psi_N grid and the five channels of the line-averaged density measured by a two-colored interferometer. The reconstruction procedure is accelerated by deploying a multi-layer perceptron to run in about 120 microseconds and is fast enough for real-time control. A proportional-integration controller is adopted, with the controller gains being estimated from the system identification processes. The experimental results show that the developed controller can follow a dynamic target while exclusively using both actuators. The absolute percentage errors between the electron density at psi_N=0.89 and the target are approximately 1.5% median and a 2.5% average value. The developed controller can even lower the density by using the pump-out mechanism under RMP, and it can follow a more dynamic target than a single actuator controller. The developed controller will enable experimental scenario exploration within a shot by dynamically setting the density target or maintaining a constant electron density within a discharge.

Machine learning algorithms often struggle to control complex real-world systems. In the case of nuclear fusion, these challenges are exacerbated, as the dynamics are notoriously complex, data is poor, hardware is subject to failures, and experiments often affect dynamics beyond the experiment's duration. Existing tools like reinforcement learning, supervised learning, and Bayesian optimization address some of these challenges but fail to provide a comprehensive solution. To overcome these limitations, we present a multi-scale Bayesian optimization approach that integrates a high-frequency data-driven dynamics model with a low-frequency Gaussian process. By updating the Gaussian process between experiments, the method rapidly adapts to new data, refining the predictions of the less reliable dynamical model. We validate our approach by controlling tearing instabilities in the DIII-D nuclear fusion plant. Offline testing on historical data shows that our method significantly outperforms several baselines. Results on live experiments on the DIII-D tokamak, conducted under high-performance plasma scenarios prone to instabilities, shows a 50% success rate, marking a 117% improvement over historical outcomes.

The near-axis description of optimised stellarator fields has proven to be a powerful tool both for design and understanding of this magnetic confinement concept. The description consists of an asymptotic model of the equilibrium in the distance from its centermost axis, and is thus only approximate. Any practical application therefore requires the eventual construction of a global equilibrium. This paper presents a novel way of constructing global equilibria using the \texttt{DESC} code that guarantees the correct asymptotic behaviour imposed by a given near-axis construction. The theoretical underpinnings of this construction are carefully presented, and benchmarking examples provided. This opens the door to an efficient coupling of the near-axis framework and that of global equilibria for future optimisation efforts.

Zernike polynomials serve as an orthogonal basis on the unit disc, and have proven to be effective in optics simulations, astrophysics, and more recently in plasma simulations. Unlike Bessel functions, Zernike polynomials are inherently finite and smooth at the disc center (r=0), ensuring continuous differentiability along the axis. This property makes them particularly suitable for simulations, requiring no additional handling at the origin. We developed ZERNIPAX, an open-source Python package capable of utilizing CPU/GPUs, leveraging Google's JAX package and available on GitHub as well as the Python software repository PyPI. Our implementation of the recursion relation between Jacobi polynomials significantly improves computation time compared to alternative methods by use of parallel computing while still performing more accurately for high-mode numbers.

High fidelity kinetic equilibria are crucial for tokamak modeling and analysis. Manual workflows for constructing kinetic equilibria are time consuming and subject to user error, motivating development of several automated equilibrium reconstruction tools to provide accurate and consistent reconstructions for downstream physics analysis. These automated tools also provide access to kinetic equilibria at large database scales, which enables the quantification of general uncertainties with sufficient statistics arising from equilibrium reconstruction techniques. In this paper, we compare a large database of DIII-D kinetic equilibria generated manually by physics experts to equilibria from the CAKE and JAKE automated kinetic reconstruction tools, assessing the impact of reconstruction method on equilibrium parameters and resulting magnetohydrodynamic (MHD) stability calculations. We find good agreement among scalar parameters, whereas profile quantities, such as the bootstrap current, show substantial disagreement. We analyze ideal kink and classical tearing stability with DCON and STRIDE respectively, finding that the $δW$ calculation is generally more robust than $Δ^\prime$. We find that in $90\%$ of cases, both $δW$ stability classifications are unchanged between the manual expert and CAKE equilibria.

Abstract We have developed TorbeamNN: a machine learning surrogate model for the TORBEAM ray tracing code to predict electron cyclotron heating (ECH) and current drive locations in tokamak plasmas. TorbeamNN provides more than a 100 times speed-up compared to the highly optimized and simplified real-time implementation of TORBEAM without any reduction in accuracy compared to the offline, full fidelity TORBEAM code. The model was trained using KSTAR ECH mirror geometries and works for both O-mode and X-mode absorption. The TorbeamNN predictions have been validated both offline and real-time in experiment. TorbeamNN has been utilized to track an ECH absorption vertical position target in dynamic KSTAR plasmas as well as under varying toroidal mirror angles and with a minimal average tracking error of 0.5 cm.

We have developed TorbeamNN: a machine learning surrogate model for the TORBEAM ray tracing code to predict electron cyclotron heating and current drive locations in tokamak plasmas. TorbeamNN provides more than a 100 times speed-up compared to the highly optimized and simplified real-time implementation of TORBEAM without any reduction in accuracy compared to the offline, full fidelity TORBEAM code. The model was trained using KSTAR electron cyclotron heating (ECH) mirror geometries and works for both O-mode and X-mode absorption. The TorbeamNN predictions have been validated both offline and real-time in experiment. TorbeamNN has been utilized to track an ECH absorption vertical position target in dynamic KSTAR plasmas as well as under varying toroidal mirror angles and with a minimal average tracking error of 0.5cm.