Plasma disruptions in tokamaks—sudden instabilities that can quench fusion reactions—pose a major challenge to sustainable energy. When magnetic fields collapse, they unleash damaging forces on reactor walls. AI-driven models are now being trained to detect subtle precursors, predicting quenches before they cascade.
Machine intelligence learns from vast plasma data. It could become the guardian of fusion stability, edging humanity closer to safe, controlled star power on Earth.
Let’s explore the physics of plasma disruptions in tokamaks and how AI could predict quench events before magnetic fields collapse, securing fusion stability.
 |
| Plasma disruption in fusion reactor |
The Physics of "Plasma Disruptions": Can AI Predict a Tokamak Quench Before the Magnetic Field Collapses?
The pursuit of commercial nuclear fusion energy is often described as the ultimate scientific "moonshot," a multi-generational effort to replicate the power of the stars within the confines of a terrestrial laboratory.
At the heart of this endeavor is the tokamak, a complex machine that uses intense magnetic fields to trap a plasma of hydrogen isotopes at temperatures exceeding 100 million degrees Celsius. However, keeping this "star in a bottle" stable is perhaps the most daunting challenge in modern physics.
The plasma is a capricious medium, prone to sudden, violent instabilities known as disruptions. These events represent a rapid loss of confinement, where the stored thermal and magnetic energy collapses in a matter of milliseconds.
As we transition from experimental devices to reactor-scale facilities like ITER, the stakes have never been higher.
A single unmitigated disruption in ITER could release forces equivalent to the weight of a jumbo jet and thermal loads that exceed the melting point of any known material by an order of magnitude.
For decades, the goal was merely to survive these events, but the emergence of artificial intelligence (AI) has shifted the paradigm.
We are now entering an era where deep learning models can "see" the precursors of a disruption hundreds of milliseconds before they occur, allowing autonomous control systems to intervene and steer the plasma back to safety.
In this article, we will explore the intricate physics of plasma disruptions and the revolutionary role AI is playing in predicting the terminal "quench" before the magnetic field—and our hopes for clean energy—collapses.
The MHD Foundation: Maintaining the Delicate Balance
The stability of a tokamak plasma is governed by the laws of magnetohydrodynamics (MHD), a theoretical framework that treats the plasma as a single-fluid conducting medium interacting with electromagnetic fields.
In a state of perfect equilibrium, the expansive pressure of the hot plasma is exactly balanced by the magnetic Lorentz force, a condition expressed by the vector equation $\nabla p = J \times B$.
Achieving this balance in a toroidal (donut-shaped) geometry is inherently difficult because the magnetic field is naturally stronger on the inboard side than the outboard side, creating a gradient that tends to push the plasma outward.
Furthermore, the Virial Theorem dictates that a magnetofluid cannot maintain equilibrium through its own internal currents alone; it requires an intricate array of external magnetic coils to provide the necessary shaping and positioning forces.
When any of these balancing forces fail, or when the plasma exceeds certain operational limits in density or pressure, the result is a disruption.
Instabilities in the MHD framework are categorized based on their drivers and timescales. Ideal MHD instabilities, such as the internal kink mode or the vertical displacement event (VDE), occur at the speed of Alfvén waves and are driven by steep pressure gradients or current density profiles.
For example, in modern tokamaks that utilize elongated (non-circular) plasma cross-sections for better confinement, the plasma is inherently unstable to vertical motions.
If the control system fails to compensate for a minor vertical shift, the instability grows exponentially, leading to a collision with the vacuum vessel wall.
Resistive MHD instabilities, by contrast, are more subtle and arise because the plasma has a finite electrical resistance. This resistivity allows magnetic field lines to break and "reconnect," forming magnetic islands that short-circuit the nested magnetic surfaces and degrade confinement.
Understanding these fundamental mechanisms is the first step in developing predictive models, as they define the "operational space" where the plasma remains stable.
Disruption Dynamics: The Violent Transition of Quenches
A plasma disruption is not a single instantaneous event but a cascading failure that proceeds through two distinct phases: the Thermal Quench (TQ) and the Current Quench (CQ).
The process typically begins with a precursor phase where MHD instabilities grow to a critical amplitude. Once the threshold is crossed, the TQ occurs. During this phase, the plasma’s stored thermal energy is suddenly released to the surrounding material surfaces.
In existing tokamaks like DIII-D, the TQ happens in less than a millisecond, causing the electron temperature ($T_e$) to plummet from millions of degrees to just a few tens of electron volts.
This collapse is often attributed to the growth of the Resistive Wall Tearing Mode (RWTM), which creates a stochastic magnetic field that allows heat to escape the core at incredible speeds.
In larger devices like ITER, the TQ is expected to last longer—roughly 70 to 100 milliseconds—due to the higher plasma volume and different vessel wall conductivities.
Immediately following the TQ is the Current Quench (CQ). Because the plasma has lost its thermal energy, it becomes highly resistive.
The massive toroidal current, which can reach 15 million amperes in ITER, can no longer be sustained and rapidly decays to zero. This decay is dangerous because it induces powerful electric currents in the conducting structures of the tokamak, such as the vacuum vessel and the blanket modules. These induced "eddy currents" and the direct-contact "halo currents" interact with the background magnetic fields to generate immense mechanical stresses.
The timescale of the CQ is a critical parameter; if it is too fast, the electromagnetic forces can be destructive, but if it is too slow, it can lead to the generation of a relativistic beam of runaway electrons. Balancing these risks is the primary objective of any disruption mitigation strategy.
Magnetic Islands and the Critical Onset of Tearing
The most common precursors to a total plasma collapse are tearing modes, specifically Neoclassical Tearing Modes (NTMs). These instabilities represent a fundamental change in the topology of the magnetic field.
When the plasma's internal pressure and current profile reach a certain state, it becomes energetically favorable for the magnetic field lines to "tear" and reconnect into "islands".
These islands are essentially bubbles of independent magnetic flux that rotate with the plasma. Because they flatten the local temperature and pressure profiles, they act as a drain on the plasma's energy, reducing the fusion performance. If left unchecked, these islands can grow large enough to overlap with other islands, leading to a global stochasticity that triggers the thermal quench.
A particularly dangerous scenario occurs when a rotating magnetic island "locks" to the vacuum vessel. Tokamaks have slight imperfections in their magnetic coils, known as error fields.
As an island grows, its rotation slows down due to electromagnetic drag against the vessel wall. When it finally stops rotating—a state called a "locked mode"—it creates a persistent, localized perturbation.
Locked modes are nearly always followed by a disruption. Research at the MAST-U tokamak has focused on using machine learning to predict the trajectory of the plasma toward this locked state.
By analyzing core density and temperature distributions, researchers have found that the probability distributions for locked and unlocked shots are well-separated, allowing AI models to provide reliable alarms with warning times of 10 milliseconds or more—enough to trigger mitigation systems.
The Halo Current Challenge: Structural Integrity Under Fire
When a disruption causes the plasma column to lose vertical stability, it often undergoes a Vertical Displacement Event (VDE).
As the plasma shifts toward the top or bottom of the vacuum vessel, it comes into direct contact with the wall. This contact allows a portion of the plasma current to flow directly through the vessel’s conducting components before returning to the plasma, creating what is known as a "halo current".
These currents are particularly problematic because they are not confined to the plasma; they flow through the structural ribs and cooling pipes of the reactor.
The interaction of these halo currents with the high-strength toroidal magnetic field ($B_t$) produces massive, asymmetric Lorentz forces that can twist and deform the entire vacuum vessel.
Physicists at the COMPASS tokamak in Prague recently conducted an extensive series of experiments to map these currents with high spatial resolution. Using arrays of electric sensors, they discovered a crucial physical limit: the local halo current density cannot exceed the local plasma particle flux to the components.
This finding is significant for ITER because it suggests that the total surface area over which halo currents pass actually increases as the total plasma current increases. This "spreading" effect could potentially reduce the local mechanical stresses on individual wall components, making the disruption less damaging than previously feared.
However, even with this mitigation by nature, the global forces remain staggering, with ITER simulations predicting loads equivalent to several hundred tons of force on the blanket modules.
Relativistic Runaway: The Peril of High-Energy Electron Beams
Perhaps the most insidious threat posed by a disruption is the generation of runaway electrons (REs). During the current quench, the sudden drop in plasma temperature increases the electrical resistance.
According to Lenz's law, the collapsing magnetic field induces a massive toroidal electric field to oppose the change in current. In the low-density environment of a post-TQ plasma, some electrons are accelerated by this electric field to nearly the speed of light.
These electrons become "decoupled" from the rest of the plasma because the collisional drag decreases as their velocity increases. This can lead to a "knock-on" avalanche effect, where a single high-energy electron collides with others, creating a massive beam of relativistic particles that can carry several mega-amperes of current.
If this runaway beam strikes the reactor wall, it acts like a high-powered laser, penetrating deep into the material. While a thermal quench might melt the surface of a beryllium wall, a runaway electron beam can cause bulk melting of the stainless steel structure or even the superconducting magnets behind it. This risk is so severe that ITER's disruption mitigation system is specifically designed to prevent RE formation by injecting large amounts of heavy gases like neon or argon to increase the plasma density and "brake" the electrons through collisions.
AI models are being trained to recognize the specific magnetic signatures of RE "seeds"—the initial population of high-energy electrons—so that the gas injectors can be fired before the avalanche begins.
Early Warning Systems: The Evolution of Disruption Prediction
The first generation of disruption predictors relied on traditional machine learning algorithms like Support Vector Machines (SVMs) and Random Forests. These models were essentially binary classifiers: they were fed a "feature vector" of plasma parameters (such as current, density, and radiation levels) and asked to determine if the state was "disruptive" or "safe".
A notable success in this area was the APODIS system at the JET tokamak, which used a two-layer SVM architecture to analyze data from multiple time windows before a potential disruption. APODIS achieved an impressive accuracy rate of over 98%, with a false alarm rate of less than 2%.
However, traditional machine learning has fundamental limitations in a fusion environment. These models are purely empirical, meaning they do not understand the underlying physics; they simply look for patterns in the data. This makes them difficult to "extrapolate" to new machines.
For example, a model trained on the small DIII-D tokamak might fail on the much larger JET or ITER because the timescales and physics regimes are different. To overcome this, researchers are now turning to Deep Learning and Recurrent Neural Networks (RNNs) that can process the entire "trajectory" of a plasma shot over time.
By training these models on multi-terabyte databases of historical experiments from different facilities, scientists are developing "cross-tokamak" predictors that can identify universal precursors of instability regardless of the machine's size.
The AI Breakthrough: Real-Time Avoidance at the DIII-D Tokamak
A paradigm shift occurred in early 2024 when a team led by Princeton University demonstrated the use of Deep Reinforcement Learning (DRL) to not just predict, but actively avoid disruptions in real-time.
Operating at the DIII-D National Fusion Facility, the AI was tasked with managing tearing mode instabilities. Unlike previous "alarm" systems, this AI was a "pilot." It was trained in a simulated environment to understand how changing the "knobs" of the tokamak—such as the plasma shape, the neutral beam injection power, and the magnetic coil currents—would affect the stability of the plasma.
During live experiments, the AI monitored the plasma and forecasted the likelihood of a tearing mode up to 300 milliseconds in advance. This is an eternity in plasma physics.
Upon detecting a burgeoning instability, the AI autonomously adjusted the beam torque and magnetic perturbations to "steer" the plasma away from the unstable regime. This successfully prevented disruptions in high-performance scenarios that were previously considered too risky to explore.
This success proves that AI can understand and control high-level physics in a way that traditional, hand-tuned control laws cannot, paving the way for autonomous "autopilots" in future commercial fusion reactors.
Scientific Machine Learning: Hybrid Models for Complex Dynamics
While pure AI models are powerful, they often lack the "common sense" of physical laws. To bridge this gap, researchers at MIT have developed Scientific Machine Learning (SciML) techniques, specifically the Neural State-Space Model (NSSM).
The NSSM is a hybrid architecture: it uses standard physical equations (0D models) to describe the basic conservation of energy and particles, but embeds "neural network nodes" to represent the complex, non-linear effects that are too difficult to simulate from first principles—such as confinement times and radiation losses.
One of the most impressive aspects of the NSSM is its sample efficiency. While traditional deep learning requires thousands of examples, the NSSM was trained on only 311 discharge experiments from the TCV tokamak.
Despite this small dataset, the model showed remarkable accuracy in predicting plasma dynamics during the high-risk "ramp-down" phase—the period at the end of a shot when the current is reduced.
The model is so fast that it can simulate 10,000 different ramp-down trajectories per second on a single GPU. This allows operators to run a "predict-first" experiment, where the AI tests millions of scenarios in seconds to find the safest possible path to shut down the reactor without a disruption.
Beyond the Plasma: AI for Superconducting Magnet Quench Protection
While the plasma is the most visible source of instability, the superconducting magnets that provide the confinement field are also subject to their own "quenches."
A magnet quench occurs when a portion of the superconducting coil loses its zero-resistance state and returns to a normal, resistive state. This transition releases the gigajoules of energy stored in the magnet as heat, potentially melting the coil or the vacuum vessel.
Because the magnets are located in a high-noise environment—surrounded by the electromagnetic chaos of the plasma—detecting a quench early is notoriously difficult.
AI is now being deployed to monitor these magnets by analyzing diagnostic data for "quench precursors"—tiny anomalies in the voltage signals that indicate a local loss of superconductivity.
The "IntelliMIK" system, developed for the EAST tokamak, uses a neural network to compensate for the complex induced voltages caused by changing magnetic fields.
By filtering out this background noise, the AI can detect a quench signal that is orders of magnitude smaller than the interference.
This provides the 2-3 seconds of warning time needed to safely discharge the magnet's energy, protecting the reactor's most expensive components from catastrophic failure.
The Final Frontier: Protecting Superconducting Magnets and ITER Scaling
As we look toward the completion of ITER, the integration of AI-based disruption prediction and avoidance is no longer a research luxury; it is a fundamental requirement for the machine's survival.
ITER's Disruption Mitigation System (DMS) will rely on Shattered Pellet Injection (SPI), where cryogenic pellets of neon and deuterium are fired into the plasma at 250 meters per second.
For SPI to be effective, the timing must be perfect—the pellets must arrive at the exact moment the disruption begins to maximize their cooling effect and minimize the formation of runaway electrons.
The success of these systems depends on "transfer learning," where AI models trained on today's smaller tokamaks like KSTAR, DIII-D, and JET are "scaled up" to ITER's dimensions.
Research at KSTAR has already demonstrated that injecting multiple pellets from different toroidal locations can more effectively radiate away the plasma's energy, a strategy that ITER will adopt.
By combining these experimental findings with AI that can process data from 3,000 magnet sensors and hundreds of plasma diagnostics, we are building a comprehensive safety net.
The goal is to reach a state where the "artificial sun" is no longer a volatile beast to be managed, but a stable, reliable source of power, steered by an AI that can anticipate a quench before the first magnetic field line even begins to tear.
The future of fusion rests on this delicate intersection of plasma physics and machine learning.
As our models become more physically grounded and our control systems more autonomous, the threat of plasma disruptions will transform from a show-stopping obstacle into a manageable engineering constraint.
In the decades to come, the "star in a bottle" will finally stay contained, fueled by the invisible intelligence of the algorithms that watch over it.
References
- Adámek, J., et al. (2022). "Physical limit to electric currents between plasma and first reactor wall during disruptions." Nuclear Fusion. https://doi.org/10.1088/1741-4326/ac5e5b
- Murari, A., et al. (2024). "A control oriented strategy of disruption prediction to avoid the configuration collapse of tokamak reactors." Nature Communications, 15(1). https://doi.org/10.1038/s41467-024-45432-1
- Hollmann, E. M., et al. (2010). "Consequences of disruptions on tokamak components and vacuum vessel." Journal of Nuclear Materials. https://doi.org/10.1016/j.jnucmat.2010.10.009
- Paccagnella, R. (2011). "Tokamak Magnetohydrodynamic Equilibrium and Stability." ITER Physics Basis. https://doi.org/10.1088/0029-5515/39/12/301
- Zohm, H. (2014). Magnetohydrodynamic Stability of Tokamaks. Wiley-VCH. https://doi.org/10.1002/9783527677337
- Wang, A. M., et al. (2025). "Learning plasma dynamics and robust rampdown trajectories with predict-first experiments at TCV." Nature Communications, 16(1). https://doi.org/10.1038/s41467-025-63917-x
- Boozer, A. H. (2015). "Physics of tokamak disruptions and their mitigation." Physics of Plasmas, 22(3). https://doi.org/10.1063/1.4913582
- Strauss, H. (2022). "Thermal quench time and resistive wall tearing modes." Physics of Plasmas, 29, 112508. https://doi.org/10.1063/5.0112658
- Yan, Q., & Hu, Y. (2025). "IntelliMIK: a novel intelligent quench detection method for fusion devices." Nuclear Fusion, 65(3), 036022. https://doi.org/10.1088/1741-4326/ad977f
- Seo, J., et al. (2024). "Avoiding fusion plasma tearing instability with deep reinforcement learning." Nature, 626, 746–751. https://doi.org/10.1038/s41586-024-07024-9
- Gambrioli, M., et al. (2024). "Locked mode prediction through machine learning algorithms in MAST-U." 50th EPS Conference on Plasma Physics, P1.082. https://doi.org/10.48550/arXiv.2405.05908
- Mantri, G., et al. (2025). "AI-driven physics-informed neural operators for predictive modelling of plasma turbulence." The European Physical Journal Plus, 140(11), 1141. https://doi.org/10.1140/epjp/s13360-025-07090-6
- Kates-Harbeck, J., et al. (2019). "Predicting disruptive instabilities in controlled fusion plasmas through deep learning." Nature, 568, 526–531. https://doi.org/10.1038/s41586-019-1116-4
- Jachmich, S., et al. (2021). "Shattered pellet injection experiments at JET." IAEA Fusion Energy Conference.(https://juser.fz-juelich.de/record/892844)
- Jalalvand, A., et al. (2025). "Multimodal super-resolution: discovering hidden physics and its application to fusion plasmas." Nature Communications, 16(1). https://doi.org/10.1038/s41467-025-72815-4
- Stabilini, E. (2025). "Machine Learning for superconducting magnets application." OSTI Technical Report. https://doi.org/10.2172/3018641
Read Here: Can a Uniformly Dense Sphere in a Vacuum Rotate on Two Axes Simultaneously?