Predict the burst, count the fuel, do the math.

An AI spots dangerous plasma bursts 100 milliseconds before they hit.

A fusion reactor can damage its own walls in milliseconds — here is a neural network learning to feel the sneeze coming.

When a fusion plasma transitions into its best-performing state — what physicists call H-mode, short for high-confinement mode — it builds up pressure at its edge the way water backs up behind a dam. Eventually that pressure releases in a violent burst called an ELM, or Edge Localized Mode. The very first ELM after the transition is typically the worst: it can dump a massive heat spike onto the reactor wall in a fraction of a second. For a future power plant, that kind of repeated battering is exactly what you want to avoid. A team at DIII-D, the tokamak run by General Atomics in San Diego, trained a neural network to spot that burst before it arrives. They fed the model 50 milliseconds of radar-like plasma-edge measurements — a diagnostic called Doppler Backscattering, which bounces microwaves off the plasma to detect turbulence — and asked it: when is the first ELM going to hit? The answer came 100 milliseconds early. In plasma terms, that is a genuinely useful warning window — enough time for a control system to fire a mitigation response. The catch is significant. This paper is a proof of concept, and the available text does not report precision, recall, or false-alarm rates. We do not know how often the model raises a false alarm. In a real reactor that matters enormously: a false alarm wastes fuel and interrupts the discharge; a miss lets a damaging burst through unmitigated. The team adapted a framework called DeepHit, originally designed for medical survival analysis, and demonstrated it works in principle. The next step — thorough statistical validation across many more plasma shots, with clear error metrics — still has to happen before anything like this could run near a real machine.

Thirty-five grams of tritium get stuck in ITER's walls. Baking gets most of it back.

Radioactive fuel worth millions is quietly disappearing into the reactor walls — and the best way to get it back is to run the oven.

Tritium — the rare, radioactive form of hydrogen that will be ITER's fuel — is so scarce and expensive that a future fusion plant will need to track every gram it uses. So here is a number that focuses the mind: according to new computer simulations, after just ten days of fusion pulses, roughly 35 grams of tritium could be sitting trapped inside the reactor walls. That is fuel you have already consumed but cannot use, and a radioactive inventory you are legally and practically obligated to manage. A team building a tool called HISP — Hydrogen Inventory Simulations for Plasma-Facing Components — connected existing plasma simulation outputs directly to a hydrogen transport model to track exactly this. The finding that stands out most: about 80% of that trapped tritium was not in the main metal wall. It was in thin, mixed films that form when eroded wall material re-deposits in the divertor, the exhaust-handling component at the bottom of the machine. These co-deposited boron layers act like a sponge. And the most effective way to wring them out? Baking. Heating the tungsten metal components removed 88% of the tritium trapped there. Other cleaning approaches — flooding the chamber with low-power pulses, or using a gas discharge — were significantly less effective. The honest limit: this is a simulation chain, not a direct measurement. The HISP pipeline hasn't been validated against real experimental tritium retention data, and the plasma inputs it uses involve their own assumptions. Think of it like using a detailed recipe simulation to predict how much salt soaks into your cast-iron pan. The physics model is credible, but until someone measures a real ITER-scale wall after real DT pulses, the numbers carry genuine uncertainty. The directional signal — baking works, boron layers are the main problem — is still useful for ITER's operational planning.

A new formula tells you whether a fusion plant can make money — before it's built.

Fusion has a famous formula for when the physics works — now someone has written the equivalent formula for when the finances work.

In 1957, physicist John Lawson wrote down the minimum conditions a plasma must meet to produce more energy than it consumes. That Lawson criterion became one of fusion's most useful scorecards — a single number that separates experiments that are wasting energy from ones that are not. A new theoretical paper attempts something similar, but for money. The researchers derive what they call Q_econ — an economic gain factor. Like the Lawson criterion, it compresses a complicated reality into a single inequality: if Q_econ is above one, you have at least a chance of running a viable power plant. Below one, you are guaranteed to lose money regardless of how good the plasma is. The formula combines ten normalised design parameters: things like how long a wall component survives before needing replacement, how much energy it can absorb before failing, what price you can sell electricity for, how efficiently you convert fusion heat into grid power, and how much downtime you need for repairs. Think of it like a café's break-even calculation. You don't need to know how big the café is — you need the margin per coffee, the cost per hour, how long the espresso machine runs before its first service, and how long that service takes. The framework is agnostic to fusion technology: magnetic, laser-driven, or anything else. The catch the authors themselves name: Q_econ ≥ 1 is a necessary condition, not a sufficient one. It is the floor, not the ceiling. And most of the ten input parameters are still unknown for a machine that does not yet exist. It is a map drawn before the territory has been explored. But having a map — even a rough one — is still better than walking blind into a trillion-dollar construction decision.

Put these three papers side by side and you see one shared question: what does it take to run a fusion machine, not just to make it fire? The ELM prediction work is about control — keeping the plasma stable enough, long enough, without letting it damage the walls it lives inside. The tritium inventory work is about stewardship — tracking every gram of radioactive fuel, knowing where it goes and how to recover it. The economic viability framework is about accountability — defining the conditions under which all of that effort becomes financially sustainable. None of these are papers about making fusion work scientifically. That part is increasingly assumed. What they are about is the engineering and operational reality of a machine you would actually run for decades. The field is not just trying to ignite plasma anymore. It is trying to figure out how to keep it running, clean up after it, and make the numbers add up. That is a different kind of problem — and the fact that it is getting serious attention is, quietly, a sign of progress.

The most immediate next step on the ELM prediction work is a full statistical validation — precision, recall, false-alarm rate — across a large set of DIII-D shots; watch for a follow-up from the same group. On tritium retention, ITER's first hydrogen phase (no tritium) is targeting operations in the late 2020s, and real wall-retention measurements from that phase will be the first opportunity to check whether simulations like HISP are in the right ballpark. The open question I would most want answered: how sensitive is Q_econ to realistic uncertainty ranges in those ten parameters? A sensitivity analysis from the economic viability team would make that framework genuinely useful for plant designers.

• Wikipedia: Edge Localized Modes — background on the plasma instability the ELM prediction paper targets
• ITER Organization: Tritium — why fusion's fuel is so carefully managed
• Wikipedia: Lawson criterion — the physics benchmark the economic viability paper is modelled on