Publications

Plasma confinement state classification in fusion power plants: Profile reflectometer and ensemble diagnostics

Clark, R., et al. · 2026

As Fusion Pilot Plants (FPPs) are increasingly viewed as within reach, many engineering challenges remain. Not many diagnostics are expected to be available in a reactor environment. We expand on previous ECE-based confinement mode classification by developing a Profile Reflectometer (PR) based classifier with 97% test accuracy, and an ensemble model combining ECE and PR models achieving 99% test accuracy.

Plasma confinement state classification via FPP relevant microwave diagnostics

Clark, R., et al. · Plasma Physics and Controlled Fusion, 2025

A parsimonious and robust machine learning approach for identifying plasma confinement states in fusion power plants, where reliable L-mode/H-mode identification is critical for safe and efficient operation. Using only electron cyclotron emission (ECE) signals with radial basis function features and a gradient boosting classifier, the framework achieves 96% test accuracy — demonstrating that state-of-the-art performance is attainable from a restricted diagnostic set.

Robustness by design: Interface contracts for AI control in high-stakes physical systems

Glukhov, V., et al. · 2025

Fusion power plants demand levels of robustness unattainable by purely data-driven or purely physics-based control. We argue that hybrid architectures' reliability depends on explicit interface contracts — formal specifications of how heterogeneous modules exchange information, share responsibility, and recover from uncertainty or model failure. Using confinement-state identification in fusion plasmas as an illustration, we show that robustness in AI-driven control arises from disciplined interface design rather than algorithmic novelty.

Reconstruction-free magnetic control of DIII-D plasma with deep reinforcement learning

Subbotin, G., et al. · Nuclear Fusion, 2025

The first application of deep reinforcement learning (RL) for magnetic plasma control on the DIII-D tokamak, using the Soft Actor-Critic algorithm. This approach eliminates the need for equilibrium reconstruction, enabling high-speed control execution. NSFsim — a 2D Grad-Shafranov solver with circuit equations and 1D transport — is used to train the agent. RL-based controllers demonstrated robust magnetic control in experimental application at DIII-D, reaching target parameters from the first discharge without additional tuning.

Summary report from the mini-conference [APS DPP 2024] on digital twins for fusion research

Schissel, D. P., et al. · Physics of Plasmas, 2025

Overview of the mini-conference on Digital Twins for Fusion Research held at APS DPP 2024 in Atlanta, GA. Presentations showcased a rapidly growing ecosystem of physics-based simulations, data assimilation strategies, and AI methods moving digital twins from concept to operational reality. Participants emphasized the need for robust uncertainty quantification, standardized data formats, and open interfaces merging legacy codes, large-scale simulations, and real-time diagnostics.

Reconstructing the plasma boundary with a reduced set of diagnostics

Stokolesov, M., et al. · Journal of Plasma Physics, 2025

Feasibility study for reconstructing the last closed flux surface (LCFS) in DIII-D using neural networks trained on reduced input feature sets. A model trained solely on coil currents achieved a mean point displacement of 0.04 m; adding plasma current and loop voltage reduced error to 0.03 m. Results demonstrate the potential of ML to perform effectively in data-limited environments expected in Fusion Power Plants.

Validation of NSFsim as a Grad-Shafranov equilibrium solver at DIII-D

Clark, R., et al. · Fusion Engineering and Design, 2025

NSFsim — a free boundary equilibrium and transport solver based on DINA — is validated against DIII-D for five plasma shapes: Lower Single Null, Upper Single Null, Double Null, Inner Wall Limited, and Negative Triangularity. NSFsim results are compared against measured signals, EFIT magnetic profile fits, and the GSevolve simulator, confirming accurate reproduction of plasma shape, poloidal flux distribution, and simulated diagnostic signals.

Electromagnetic system conceptual design for a negative triangularity tokamak

Guizzo, S., et al. · 2025

Conceptual design of a compact NT device (R₀ = 1 m, a = 0.27 m, Bₜ = 3 T, Iₚ = 0.75 MA) using TokaMaker for electromagnetic system development. Eight poloidal field coils achieve a wide range of plasma geometries with −0.7 < δ < −0.3 and 1.5 < κ < 1.9. Passive stabilizing plates reduce vertical instability growth rates by ~75%. The design demonstrates that key capabilities of a dedicated NT experiment are realizable with existing copper magnet technologies.

Machine learning approaches to plasma state mode classification via reactor relevant diagnostics at DIII-D

Clark, R., et al. · APS DPP, 2025

Successful L-H mode classification using supervised learning models trained exclusively on reactor-relevant diagnostics from DIII-D: ECE, Visible Filter scopes, and the Radial Interferometer-Polarimeter (RIP). These represent a reduced diagnostic set expected to remain viable in fusion pilot plants. Gradient Boosting Classifiers on ECE temperature features achieve strong performance; Visible Filter and RIP spectrograms further enhance classifier accuracy via ELM and LH transition signatures.

Integrated tokamak control using NSFsim simulator

Subbotin, G., et al. · APS DPP, 2025

Demonstration of integrated tokamak control within the NSFsim simulation environment, combining equilibrium solving, transport, and RL-based magnetic control into a unified workflow for scenario development and controller training on DIII-D.

Fusion Twin Platform: an innovative tool for fusion research and education

Zhurba, A., et al. · FEC, 2025

The Fusion Twin Platform (FTP) at fusiontwin.io is a free web-based tool democratizing access to advanced tokamak simulations. FTP allows researchers, educators, and students to use pre-built digital replicas of tokamaks, enabling precise simulations, exploration of ML models, visualization of plasma dynamics, and flexible data management — powered by the NSFsim free boundary equilibrium and transport solver.

Reconstructing the plasma boundary with a reduced set of diagnostics

Stokolesov, M., et al. · FEC, 2025

Neural network models for LCFS reconstruction in DIII-D, ranging from well-posed tasks with extensive diagnostics to ill-posed tasks relying on minimal data. The model trained exclusively on coil currents achieved a mean point displacement of 4.5 × 10⁻² m, demonstrating ML capability in weak data environments anticipated in FPPs due to blanket and shielding constraints.

Fusion Twin Platform: an innovative tool for fusion research and education

Zhurba, A., et al. · EPS, 2025

Presentation of the Fusion Twin Platform at EPS 2025 — a free web-based tool enabling precise tokamak simulations, ML model exploration, and plasma dynamics visualization, powered by NSFsim and designed to enhance collaborative research and fusion education.

Design and implementation of a reinforcement learning-based plasma shape controller at DIII-D

Subbotin, G., et al. · EPS, 2025

A novel RL-based controller for plasma shape and position control developed within the NSFsim environment. NSFsim, validated as a DIII-D Grad-Shafranov solver, provides simulated diagnostics as agent inputs and simulated actuator controls as outputs. The RL agent was successfully applied to DIII-D, controlling plasma shape and position during piggyback experiments.

NSFsim code for machine design and scenario development

Nurgaliev, M., et al. · EPS, 2025

NSFsim is a free-boundary Grad-Shafranov and 1D transport solver inheriting numerical approaches from the DINA code. Recent upgrades have expanded its application capabilities while improving performance and flexibility. Currently used for tokamak performance prediction, scenario development, ML applications, and device optimization — enabling estimation of plasma parameter evolution, coil current dynamics, and induced currents in passive structures.

Developing a state-oriented plasma control system

Kachkin, A., et al. · OSSFE, 2025

A plasma-state-oriented control system for FPPs built around: (1) definition of finite target robust plasma states optimizing fusion performance and stability; (2) real-time plasma state reconstruction; (3) anticipating state transitions through predictive modeling; (4) coordinated actuation to guide plasma toward desired states while avoiding events that could lead to machine damage. The proposed machine-agnostic approach is applicable to any magnetic confinement device.

Integrated modeling for equilibrium, scenarios, and disruption processes in tokamaks with DINA and NSFsim

Khairutdinov, E., et al. · IMEG 17, 2025

Novel tools for tokamak design and control

Subbotin, G. · PhDiaFusion, 2025

Plasma control online meetup — May 6th, 2025

Zhurba, A.; Subbotin, G.; Mele, A.; Dubbioso, S.; Glukhov, V. · 2025

Four presentations: "Controlling fusion plasma with reinforcement learning" (Subbotin, NSF) · "Application of Model Predictive Control to plasma shape in TCV" (Mele, SPC/EPFL) · "Data-driven and model-free approaches for magnetic control" (Dubbioso, Consorzio CREATE) · "Multi-Objective Optimization and Control of Plasma States" (Glukhov, NSF).

Magnetic control of tokamak plasma through deep reinforcement learning with privileged information

Sorokin, D., et al. · AI4X, 2025

NSFsim validation as a DIII-D plasma equilibrium simulator

Clark, R., et al. · APS DPP, 2024

Plasma shape must be considered for any Fusion Pilot Plant for its impact on plasma stability and core confinement. NSFsim, developed based on DINA to solve the Grad-Shafranov equation and related transport equations, is validated against DIII-D measurements and EFIT reconstructions.

Validation and verification of synthetic magnetic diagnostics based on free boundary equilibrium solver for DIII-D plasma control system

Nurgaliev, M., et al. · EPS, 2024

Validation of simulation models for magnetic probes and flux loops for future RL simulation environment applications. Recent results demonstrated the possibility of using RL algorithms to develop efficient controllers operating on magnetic diagnostic signals directly — without equilibrium reconstruction. NSFsim, a free-boundary equilibrium and transport solver based on DINA, is used for modeling.

Electromagnetic system conceptual design for a negative triangularity tokamak

Subbotin, G., et al. · EPS, 2024

Next Step Fusion's design of an optimized small-to-medium NT tokamak with sophisticated control capabilities, targeting high-β regimes with high current, density, and energy confinement. Negative triangularity was selected as the target shape for the Next Step Fusion device (NTT) due to its ability to achieve high-performance plasma while posing meaningful challenges for plasma control.

Magnetic feedback control of DIII-D tokamak via deep reinforcement learning

Granovskiy, A., et al. · EEML, 2024

Demonstration of the fundamental capability of magnetically confining plasma on the DIII-D tokamak with a feedback control policy learned via deep reinforcement learning — addressing the key challenge of confining a tens-of-millions K plasma in a precise shape in the presence of instabilities developing on sub-ms timescales.