Taming Plasma Hiccups, Tracking Fuel Atoms, Predicting the Heat

European Teams Found Two Ways to Stop Plasma's Damaging Wall Punches

Your fusion reactor's plasma sneezes thousands of times a second — and every sneeze hammers the wall hard enough to slowly destroy it.

A tokamak holds plasma in a magnetic cage. Every few milliseconds, the edge of that plasma can suddenly burp outward — a pressure spike that blasts heat and particles into the reactor wall. These burps are called Type-I ELMs, short for Edge-Localised Modes. Think of a garden hose with a kink: pressure builds, the kink releases, pressure builds again. Multiplied by thousands of cycles a day, those hammer blows erode the wall faster than it can recover. For a machine the size of ITER — the international fusion project under construction in France — a single large ELM could deposit enough energy on a patch of wall the size of a dinner plate to melt it. A team spanning four EUROfusion machines — ASDEX Upgrade and JET in Europe, TCV in Switzerland, plus MAST-Upgrade — has now shown that two plasma configurations can eliminate Type-I ELMs entirely. The first is called negative triangularity, where the plasma cross-section is shaped like an upside-down D rather than a right-side-up one. The second is quasi-continuous exhaust, where the plasma edge is kept dense enough that heat leaks out in a steady trickle instead of periodic bursts. Swap a pressure-relief valve for a slow, constant drip. The total energy released is similar; the damage is not. Critically, the team showed that quasi-continuous exhaust is compatible with ITER's planned 15-million-ampere operating scenario. Both modes were also demonstrated in deuterium-tritium plasmas at JET — the fuel mix a real power plant would use. The catch: all of this was validated on machines far smaller than ITER. The physics models predict the approach will scale, but a prediction is not a measurement. The real test comes when ITER actually runs.

Tracking Every Atom of Fusion Fuel Through Fifty Simulated Operating Cycles

Fusion's fuel is so rare and so expensive that losing track of even a gram of it is not a rounding error — it is a design failure.

Tritium is a mildly radioactive form of hydrogen. Future fusion reactors will breed it inside the reactor itself and burn it almost as fast as they make it. The problem: tritium doesn't stay where you put it. It seeps into metal walls, gets trapped in tiny defects in the crystal structure, and slowly bleeds back out. Think of it like water soaking into a kitchen sponge — you squeeze most of it out, but some stays behind, and over time the amounts add up. For a power plant, every gram of tritium that goes missing is a regulatory problem, a safety concern, and, since tritium costs roughly $30,000 per gram, a financial one. Researchers developing Tokamak Energy's planned ST-E1 pilot plant used a software tool called TMAP8 to model tritium movement through four key surfaces inside the reactor — the divertor (the exhaust drain), two types of first-wall panels, and the vacuum vessel. They simulated 50 operating cycles, about 22 hours of run time, tracking diffusion, trapping, and recovery at each surface. Because running this full simulation for every possible wall design would take weeks, the team trained a Gaussian process surrogate — essentially a statistical shortcut built from many detailed runs — that can answer design questions in seconds. Why it matters: you cannot get a fusion power plant licensed without proving you know where your tritium is. This framework lets engineers swap wall materials or thicknesses quickly and see the tritium consequences before cutting a single piece of metal. The catch: none of this has been validated against real tritium measurements in an actual operating fusion device. The models are physically grounded and internally consistent — but a simulation that agrees with itself is not yet a simulation that agrees with reality.

A Neural Network That Predicts Fusion's Most Punishing Heat Zone in Seconds

At the bottom of every tokamak is a drain that has to handle temperatures approaching the sun's surface — and engineers need to predict exactly where the heat lands before they build it.

The divertor is the component at the base of a tokamak where exhaust heat and particles from the plasma are channelled out. It is arguably the most thermally stressed surface ever engineered. Before building one, engineers run detailed simulations using a software package called SOLPS-ITER, which models how plasma behaves in the exhaust region. These simulations are accurate, but slow — a single run can take many hours on a supercomputer cluster, making it impractical to explore thousands of possible designs. A team affiliated with EUROfusion built what they call SOLPS-NN: a family of neural networks trained on about 7,200 completed SOLPS simulations. Roughly 30% of their attempted simulations crashed before finishing and had to be discarded — an honest number they report rather than hide. The trained networks learned to take eight input parameters (heat flux, density, machine size, and others) and predict temperature and density patterns across the entire exhaust region — a grid of 5,200 spatial points — essentially instantly. It's like a chess engine that has studied millions of games: it skips the slow calculation and jumps straight to an assessment. The team's key result is that SOLPS-NN correctly predicts when plasma detachment occurs — a beneficial operating state where heat spreads out before hitting the wall rather than slamming into it as a concentrated beam. Getting that transition right is the central question for protecting the divertor in future machines. The catch: the training data uses a lower-fidelity version of the simulations (faster, but less physically detailed). When the team tried to apply transfer learning to adapt the model to ITER's higher-fidelity dataset, it offered no advantage over starting from scratch. For ITER specifically, the work still needs to be redone.

Three stories today, and they share a theme: engineering uncertainty out of fusion, one layer at a time. The EUROfusion ELM work says we can stop the plasma from punching holes in its own cage — and we have the shapes validated across multiple machines to back that up. The Tokamak Energy tritium work says we can track every atom of fuel through fifty operating cycles — and we have the software framework to test design decisions before building anything. The SOLPS-NN work says we can predict where the exhaust heat goes without waiting days on a supercomputer — and we have a neural network that agrees with experiments on the most critical transition. None of these is 'problem solved.' But together they shrink the list of things we genuinely do not know how to do. That list is shorter today than it was last week. That matters.

The ITER machine is in its final assembly phase, and any result confirming ELM-free scenarios are compatible with ITER's baseline current is directly relevant to its first-plasma operations, expected in the late 2020s. On the tritium side, the next meaningful milestone is experimental validation of these transport models against a real device — keep an eye on the results coming from JET's final DT campaign data analysis, which is still being published. The open question I'd most want answered: does negative triangularity hold up when you push to full fusion power, not just ELM-suppression conditions?

• Wikipedia: Edge-localised mode — what ELMs are and why they matter
• Wikipedia: Tritium — the rare hydrogen isotope at the heart of fusion fuel
• Wikipedia: Divertor — the exhaust system of a tokamak