Demonstration of remote, real-time predictive control of fusion plasma

Background

Fusion energy development is advancing as a promising solution to global energy challenges. Among candidate approaches, magnetic confinement — which traps super-hot plasma using strong magnetic fields — is one of the most developed. Turning fusion into a practical power source requires the ability to predict plasma behavior while controlling it. A promising method is digital-twin control, where a numerically reconstructed plasma runs in parallel to the experiment and navigates the actual device operation.

Our group has integrated data assimilation — a mathematical technique that continuously adjusts the model with real-time observations — into this framework, building a control technology that can estimate optimal actuation in real time with improved model fidelity.

Digital-twin control shows its full potential when multiple state variables (e.g., temperature and density) must be simultaneously optimized in real time. It can also estimate non-measured physical quantities, which is particularly important for future prototype fusion reactors where diagnostics will be limited. Achieving this requires exclusive access to supercomputer-class calculation resources to keep massively parallel predictions running alongside the experiment. Enabling the use of off-site supercomputers would, in turn, greatly broaden accessibility and deployment. The key challenge has therefore been whether one can exclusively leverage a remote supercomputer over a high-quality long-distance network with sufficiently low latency and run the observe → assimilate → predict → control cycle in real time.

Results

Data assimilation reduces discrepancies between simulation and reality by optimizing large-scale models using observations; it is widely used in numerical weather prediction. We developed ASTI (Assimilation System for Toroidal plasma Integrated simulation) as a data-assimilation system for fusion plasmas, and — by adding a control function — made digital-twin control of fusion plasmas possible. By this method, the model is adapted to measurements in real time, numerous prediction branches are run in parallel under varying conditions, and current optimal control inputs are derived probabilistically. We previously applied this system to the LHD, a world-leading superconducting plasma device with many actuators (particle fueling, heating, etc.) and advanced diagnostics.

In the present study, we have implemented real-time control while executing predictions on a remote supercomputer. The Plasma Simulator in Rokkasho (QST/NIFS joint system) and LHD in Toki have been connected over SINET6. The connection has also benefited from access infrastructure and operational know-how developed through EU-Japan Broader Approach (BA) activities, including the ITER Remote Experimentation Centre initiative coordinated by QST. With exclusive use of more than 20,000 CPU cores—providing ~200-fold larger parallelism than our prior setup — and minimized latency, we achieved real-time control across a ~2,000 km round-trip network path (Figure 1).

Furthermore, we have demonstrated for the first time worldwide, remote, real-time multivariable predictive control — simultaneous control of plasma temperature and density — in LHD, using a data-assimilation-driven digital twin. Concretely, electron density and temperature profiles measured in LHD have been streamed to the predictive model running in the Plasma Simulator in Rokkasho for sequential assimilation. From massively parallel predictions, optimal control inputs have been computed and applied to the heating and fueling systems. By continuously operating the observe → assimilate → predict → control cycle in real time, we have maintained model accuracy while steering the temperature and density spatial structures toward their targets simultaneously.

Significance and Outlook

The control system developed here represents an implementation step toward reactor-relevant predictive control where measured and non-measured quantities must be considered together. By exclusively leveraging a remote supercomputer over a high-speed and low-latency network for real-time operation, the approach enables both more computationally intensive models and flexible scaling to multiple devices. Demonstrating control over a supercomputer around a 2,000 km round-trip away suggests the feasibility of nationwide remote operations — for example, orchestrating devices across Japan from the Tokyo metropolitan area — supporting the pathway to practical fusion power.

Glossary

*1 Large Helical Device (LHD) (Figure 2): One of the world’s largest superconducting helical plasma experimental devices, located at the National Institute for Fusion Science (NIFS) in Toki, Gifu, Japan.

*2 Plasma Simulator (Figure 3): A state-of-the-art supercomputer jointly procured by NIFS and QST and operated at QST’s Rokkasho Fusion Energy Research Center (Aomori, Japan) since July 2025, with a peak theoretical performance of 40.4 PFLOPS. Connected to SINET6, it comprises multiple subsystems with CPUs and Graphics Processing Units (GPUs) for large-scale simulation, AI development, and is broadly used by universities and research institutes in Japan.