For decades, one of fusion energy’s hardest problems has been keeping superheated plasma from blasting holes in reactor walls. A United States team working on the DIII-D tokamak now reports using artificial intelligence to steer that plasma in real time, stopping the most damaging bursts before they reach the metal. Their peer-reviewed results in Nature Communications frame the advance as a step toward protecting ITER-scale reactors so they can run hotter and longer without melting their divertors.
The Plasma Control Challenge in Tokamaks
Tokamaks promise clean fusion power by confining plasma at temperatures hotter than the Sun, but the same energy that could power the grid can also destroy the machine. At the plasma edge, devices see rapid eruptions called edge localized modes, or ELMs, that eject energy toward the divertor, the component that handles exhaust heat. According to a Government explainer from the DOE, these events can focus heat in a narrow stripe on ITER’s divertor plates, raising the local heat flux above 10 MW per square meter and creating a credible wall-melting risk if left unchecked.
That same Government analysis describes how extreme-scale turbulence simulation and AI-derived formulas are now used to predict how narrowly this heat will strike, because the heat-load width directly sets how fast materials erode. In parallel, institutional accounts from Princeton’s PPPL note that ELMs are among the “harmful edge instabilities” that can drive intense heat flux to plasma-facing components, especially at ITER-like conditions. Together, these studies frame ELM control not as a fine-tuning exercise but as a make-or-break requirement for any future power plant-scale tokamak.
How the US Team Cracked Real-Time Control
The new work from Princeton and PPPL attacks that problem with a machine-learning surrogate model that can optimize three-dimensional resonant magnetic perturbations, or RMPs, in milliseconds. In the peer-reviewed Nature Communications paper, a Primary team describes training a fast surrogate on detailed plasma calculations, then embedding it in the DIII-D control loop so the system can search for RMP settings that suppress ELMs without sacrificing performance. Instead of waiting tens of seconds for traditional codes to evaluate a single configuration, the AI controller can explore many options during one evolving discharge.
The authors report that this approach maintained ELM suppression while roughly doubling a standard fusion performance metric, the triple product of density, temperature and confinement time, compared with more conservative settings. In a separate institutional account, the Institutional summary quotes the lead Princeton researcher describing the method as providing “stability across fusion devices,” highlighting that the same surrogate-based strategy could be adapted beyond DIII-D. By keeping harmful edge instabilities at bay while pushing the plasma toward higher pressure, the group argues that AI-guided RMP control changes the trade-off between safety and performance that has limited earlier experiments.
Experimental Evidence from DIII-D
The DIII-D tokamak has become a testbed for AI-guided plasma control, and the ELM work builds on a broader effort to integrate machine learning directly into its real-time systems. A preprint from a Primary DIII team describes deep reinforcement learning controllers that adjust magnetic coils to control plasma shape and position without relying on full equilibrium reconstruction, emphasizing real-time performance and resilience across scenarios. Another Primary DIII study documents hardware-accelerated inference in the RTSTAB and PCS infrastructure, classifying plasma states and forecasting ELMs quickly enough to inform control decisions.
Beyond magnetic control, DIII-D researchers have also probed how material injection can blunt heat loads. A Primary conference abstract on powder injection in a tungsten-coated divertor configuration reports that boron nitride and boron powders can trigger detachment and cut peak parallel heat flux by up to 90%, while also affecting tungsten erosion and leakage. These results complement the AI-driven RMP experiments: while the Nature Communications work targets the transient edge instabilities that cause bursts, the powder studies show that even when heat reaches the divertor, clever conditioning can dramatically reduce how much of it hits the wall.
Broader Implications for Fusion Reactors
For future reactors that will rely on tungsten walls, such as WEST and ITER, the combination of ELM suppression and advanced wall conditioning could determine whether high-power operation is feasible. The Useful for Government analysis of ITER’s divertor stresses that both transient bursts and steady heat must be managed to avoid melting, and that turbulence can sometimes spread heat in a way that protects surfaces. The Princeton and PPPL results suggest that if ELMs can be suppressed while keeping the plasma hot and dense, operators may be able to run at higher power without relying solely on deeply detached divertor regimes that can degrade core confinement.
A late 2025 Primary peer-reviewed study directly addresses this integration challenge by combining RMP-based edge instability control with divertor detachment, reporting quantitative changes in ELM frequency and size alongside reductions in heat flux and specific density limits. That work shows that edge control and detachment do not have to be competing strategies and can instead be tuned together to reduce both transient and steady-state loads. For ITER-scale devices, where the margin between acceptable heat flux and material damage is narrow, the DIII-D experience points to a path in which AI-guided magnetic control and engineered boundary conditions share the load.
Integrating with Other Techniques
The US breakthrough also sits within a wider portfolio of techniques that are being tested on other machines. On the fully tungsten-walled WEST tokamak, a Primary peer-reviewed study shows that boron powder injection can act as real-time wall conditioning, reducing impurity content and improving operational conditions. That WEST research finds that boron coatings limit tungsten erosion and contamination, which aligns with the DIII-D powder experiments that saw strong reductions in peak parallel heat flux and detachment triggering when boron-based powders were injected into a tungsten-coated divertor.
The late 2025 Contains paper on RMP and detachment also provides quantitative evidence that ELM size can be reduced while maintaining acceptable density, offering a template for how AI-optimized RMPs might be combined with impurity seeding or powder injection. In practice, a future ITER or DEMO-scale device could deploy machine-learning controllers to tune RMPs and plasma shape in real time, while separate actuators meter boron or other powders to keep tungsten sources low. The convergence of these methods suggests that no single fix will protect reactor walls; instead, a coordinated package of AI control, magnetic perturbations and tailored impurities will likely be needed.
What Remains Uncertain and Next Steps
Despite the strong DIII-D results, important questions remain about how far this approach can scale. The Nature Communications Author version of the manuscript frames the technical problem as preventing damaging boundary instabilities with three-dimensional perturbations without triggering new modes, and acknowledges that full deuterium tritium plasmas and more complex geometries could introduce additional constraints. A separate peer-reviewed Primary ELM study uses nonlinear resistive MHD modeling to confirm that magnetic islands can suppress ELMs, providing mechanistic support for RMP-based control but also highlighting the sensitivity of outcomes to details such as island size and rotation.
Other work is still filling in the control landscape. A Primary RMP study and a related ELM paper show that combining edge control with detachment can reduce both transient and steady heat loads, yet ITER-specific tests are not available, leaving some uncertainty about how these regimes will behave at full scale. The DOE has also highlighted in a DOE explainer that AI is expected to play a growing role in extreme-scale simulations and predictive models for heat-load width, which could feed directly into next-generation controllers. As one recent institutional account from DOE-funded engineers working on AI control of the power grid puts it, the broader goal is to use machine learning to manage complex, unstable systems in real time, a description that now fits fusion plasmas as much as it does electricity networks.
Why This Looks Like a Turning Point
The Princeton and PPPL team’s success at DIII-D has already drawn wider attention, including coverage that framed the result as a US team unlocking nuclear fusion plasma control. An Interesting Engineering report on the Nature Communications work stresses that the AI controller can respond on millisecond timescales, a leap beyond traditional optimization cycles that took tens of seconds, and highlights the claim that the approach could translate across devices. That framing matches the Nature Communications-linked institutional account, which connects the suppression of “harmful edge instabilities” directly to reducing the risk of component damage.
At the same time, reinforcement-learning experiments at DIII-D targeting other instabilities, such as tearing modes, show that the same AI toolbox is already being applied to different failure modes. A DOE DIII lab report on real-time control work aimed at maintaining high-power operation while avoiding tearing modes reinforces the sense that fusion research is shifting from static scenario design to dynamic, data-driven steering. If the early promise of these controllers holds up under harsher conditions and in larger machines, they could give operators something fusion has never really had before: a way to keep pushing performance without stepping over the line where the walls start to melt.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
