Reinforcement Learning Helps Tokamaks Avoid Damage During Plasma Shutdown

A hybrid AI model trained on just five high-performance plasma shots has demonstrated a new way for fusion reactors to shut down safely, offering a smarter, disruption-free approach to fusion control in next-generation machines like ITER.

Plasma Light Ball on black background.

Study: Learning plasma dynamics and robust rampdown trajectories with predict-first experiments at TCV. Image Credit: mollyw/Shutterstock.com

In a study that was recently published in Nature, researchers outlined a new method to predict plasma behavior during the critical shutdown phase, or "rampdown," in tokamak fusion reactors.

By combining a physics-based model with machine learning into a neural state-space model (NSSM), the team achieved highly accurate predictions of plasma dynamics and instabilities. This approach, validated on the Tokamak à Configuration Variable (TCV) tokamak in Switzerland, enabled the design of control trajectories that significantly improved rampdown safety, demonstrating the power of scientific machine learning (SciML) to address key challenges in fusion energy.

Notably, the experiments involved a limited number of high-performance pulses, underscoring the efficiency of the modeling approach despite constrained data availability.

Background

Tokamaks are machines designed to replicate the power of the sun on Earth. They use immensely powerful magnetic fields to confine a superheated plasma and force atomic nuclei to fuse, releasing vast amounts of energy.

The ultimate goal is to harness this fusion process to provide a safe, clean, and virtually limitless source of electricity. However, maintaining control over a plasma that can reach temperatures over 100 million degrees Celsius is a monumental challenge.

One of the most delicate operations is the "rampdown," the process of safely reducing and terminating the plasma current at the end of an experiment. If not managed carefully, the rampdown itself can destabilize the plasma, leading to sudden "disruptions" that can unleash intense heat fluxes and cause significant damage to the reactor's interior walls.

As fusion devices scale up to power-plant dimensions, where plasma energies will be vastly higher, developing reliable methods to manage rampdowns is paramount to ensuring the viability and economic feasibility of fusion power.

The 2020 high-performance campaign at the Joint European Torus (JET) highlighted this challenge, where a modest increase in plasma current led to a sharp rise in disruption rates during rampdowns, illustrating the real-world consequences of poorly optimized shutdown strategies.

The Rampdown Challenge and a SciML Solution

In experimental tokamaks, plasmas are pulsed, meaning they are created, sustained for a brief period, and then deliberately turned off. When a plasma becomes unstable during its main phase, or when a planned experiment concludes, operators must initiate a rampdown to reduce the plasma current to zero. However, this process is fraught with risk.

As the plasma's energy and current are dialed down, the conditions that keep it stable change rapidly, often pushing the plasma closer to its operational limits. This can exacerbate instabilities, leading to a disruption, an uncontrolled collapse of the plasma column.

The researchers recognized that a purely data-driven machine learning approach would be insufficient. While a neural network could theoretically learn to predict plasma behavior, it would require an immense amount of data, a major obstacle given that each experimental pulse in a tokamak is expensive, and data, especially from high-performance plasmas, is limited. Instead, the team opted to develop a hybrid model that integrates machine learning with established physics principles.

They created an NSSM that uses a neural network to learn the complex, non-linear dynamics of the plasma from real experimental data, but within a framework guided by the fundamental laws of plasma physics. This synergistic approach allowed the model to learn efficiently and accurately from a surprisingly modest dataset of just 311 plasma pulses from the TCV tokamak in Switzerland, with only five of those pulses in the high-performance regime most relevant to future reactors.

The model was further fine-tuned to capture behavior near operational limits, including constraints on Greenwald fraction, poloidal beta, and vertical instability growth rate—key parameters influencing disruption risk.

From Prediction to Control and Experimental Validation

For the model to be useful in a practical setting, it needed to be translated into actionable commands for the tokamak's control system. The team achieved this by combining their predictive model with reinforcement learning (RL), a type of machine learning where an algorithm learns to make a sequence of decisions to achieve a goal.

In this context, the "agent" was the control algorithm, the "environment" was the simulated plasma, and the "goal" was to find the optimal path to ramp down the plasma current while avoiding the boundaries of instability. The model was parallelized to run thousands of simulations simultaneously, accounting for various uncertainties in the plasma's behavior. This included sources of real-world variability, such as control system errors that had previously caused vertical displacement events (VDEs) in earlier runs.

Through this process, the RL algorithm learned which adjustments to the tokamak's magnets and other controls would guide the plasma safely to a full shutdown.

The true test came when this entire pipeline, prediction followed by trajectory optimization, was implemented in real experiments on the TCV tokamak. Despite a small sample size (just nine shots in the main rampdown experiment) the results showed statistically significant improvements in key performance metrics. The model-designed trajectories successfully ramped down plasma pulses more reliably and, in some cases, even faster than previous standard methods, all while avoiding the instabilities that lead to disruptions.

Specifically, the RL-designed strategies led to improvements in plasma current (Ip) and stored energy (Wtot) at the moment of termination, compared to baseline trajectories.

In a particularly compelling demonstration of the model's predictive power, the team conducted a "predict-first" experiment. The NSSM was used to design a rampdown for a plasma with a current 20 % higher than the baseline data it was trained on. The model successfully made this small extrapolation, and the subsequent experiment confirmed its accuracy, proving its robustness and ability to handle scenarios not explicitly in its training set.

Conclusion

This research is a big step forward for fusion energy. The team built a smart system that combines physics and machine learning to predict how the plasma will behave during shutdown—even with very little data.

It doesn’t just predict; it also figures out the safest way to turn off the plasma.

This is a major breakthrough for making future fusion reactors more reliable and better at handling the unpredictable nature of plasma—one of the hottest and most powerful substances in the universe.

Journal Reference

Wang et al. (2025). Learning plasma dynamics and robust rampdown trajectories with predict-first experiments at TCV. Nature Communications, 16(1). DOI:10.1038/s41467-025-63917-x. https://www.nature.com/articles/s41467-025-63917-x

Disclaimer: The views expressed here are those of the author expressed in their private capacity and do not necessarily represent the views of AZoM.com Limited T/A AZoNetwork the owner and operator of this website. This disclaimer forms part of the Terms and conditions of use of this website.

Citations

Please use one of the following formats to cite this article in your essay, paper or report:

  • APA

    Nandi, Soham. (2025, October 17). Reinforcement Learning Helps Tokamaks Avoid Damage During Plasma Shutdown. AZoRobotics. Retrieved on October 17, 2025 from https://www.azorobotics.com/News.aspx?newsID=16212.

  • MLA

    Nandi, Soham. "Reinforcement Learning Helps Tokamaks Avoid Damage During Plasma Shutdown". AZoRobotics. 17 October 2025. <https://www.azorobotics.com/News.aspx?newsID=16212>.

  • Chicago

    Nandi, Soham. "Reinforcement Learning Helps Tokamaks Avoid Damage During Plasma Shutdown". AZoRobotics. https://www.azorobotics.com/News.aspx?newsID=16212. (accessed October 17, 2025).

  • Harvard

    Nandi, Soham. 2025. Reinforcement Learning Helps Tokamaks Avoid Damage During Plasma Shutdown. AZoRobotics, viewed 17 October 2025, https://www.azorobotics.com/News.aspx?newsID=16212.

Tell Us What You Think

Do you have a review, update or anything you would like to add to this news story?

Leave your feedback
Your comment type
Submit

Sign in to keep reading

We're committed to providing free access to quality science. By registering and providing insight into your preferences you're joining a community of over 1m science interested individuals and help us to provide you with insightful content whilst keeping our service free.

or

While we only use edited and approved content for Azthena answers, it may on occasions provide incorrect responses. Please confirm any data provided with the related suppliers or authors. We do not provide medical advice, if you search for medical information you must always consult a medical professional before acting on any information provided.

Your questions, but not your email details will be shared with OpenAI and retained for 30 days in accordance with their privacy principles.

Please do not ask questions that use sensitive or confidential information.

Read the full Terms & Conditions.