One machine that's both reactor designs at once

Argon keeps ARC's exhaust cool while the core burns near 1 GW

How do you run a fusion core at 100 million degrees while keeping the exhaust wall from melting — at the same time?

The ARC fusion power plant — a compact design from Commonwealth Fusion Systems and MIT — is designed to produce roughly 800 MW of fusion power in a building-sized footprint. But a nagging engineering question hangs over it: can the core run hot enough to work while the exhaust region stays cool enough not to destroy itself? Think of a wood-burning stove. Inside the firebox, temperatures are extreme. The pipe carrying smoke out needs to stay manageable, or it corrodes. In a fusion reactor the equivalent of that smoke pipe is called the divertor — the region where hot plasma exhaust is channeled away from the main reaction zone. If the divertor gets too hot, it erodes rapidly and contaminates the plasma with wall material. A team led by researchers at MIT ran integrated computer simulations to test a solution: seeding the plasma with small amounts of argon gas. Argon radiates heat away before it reaches the divertor wall, acting like a thermal buffer. The simulations found that with the right argon seeding, divertor temperatures can be held below 2 eV — roughly the temperature of a candle flame — while the core sustains fusion power between 750 and 1000 MW. They also tested neon as an alternative. Neon contaminated the core plasma too aggressively, reducing performance and making it harder to sustain the high-confinement operating mode ARC depends on. Argon wins for now. The catch: this is all simulation. ARC does not exist yet. The models are sophisticated — they couple multiple physics codes simultaneously — but real plasma has a habit of surprising even good models. The next step is experimental validation on existing machines before anyone builds anything.

One coil array that can run as a tokamak or a stellarator

What if you didn't have to choose between the two leading fusion machine designs — because the same coils could do both?

There are two main approaches to magnetic fusion. Tokamaks — the kind ITER is — hold plasma in a donut shape using a combination of external magnets and a current driven through the plasma itself. Stellarators — like Wendelstein 7-X in Germany — use only external coils, twisted into complex shapes, avoiding the need for that internal current. Each design has real advantages. Each has serious problems. Tokamaks are more efficient but need current drive to stay stable. Stellarators don't, but their complex coil shapes are fiendishly expensive to build. A team at Columbia University asked a different question: could you build one machine that can run as either — or something in between? Think of it like a guitar that can be retuned, using the same strings, to produce completely different tonal characters. The Columbia team designed a ring of small, flat magnetic coils — made from high-temperature superconductor tape, which carries current without electrical resistance — positioned around a standard tokamak frame. By adjusting how much current flows through each coil, the same physical machine can generate tokamak plasma shapes, twisted stellarator-like configurations, or hybrids. Their simulations, anchored to the geometry of Columbia's own HBT-EP experiment, show the design achieves a surprisingly wide range of plasma configurations — all with forces on the coils safely within what HTS tape can handle. The catch: this is a purely computational feasibility study. No hardware has been built. The field strength in the simulation — 0.5 Tesla — is far below what a power-producing machine needs. And flexibility comes with trade-offs: optimizing for one configuration generally means compromising another. This is a proof-of-concept, not a blueprint.

An AI rerouted fusion heating beams live when hardware broke mid-experiment

A transmitter failed mid-experiment at DIII-D — and the control AI simply handed its job to the remaining devices without missing a beat.

At the DIII-D tokamak in San Diego, one of the main ways researchers heat the plasma is with gyrotrons — devices that fire high-powered microwaves into the plasma, like a precision microwave oven. The microwaves need to land at very specific locations inside the reactor to heat the right part of the plasma at the right moment. Getting that right while the plasma shape is constantly shifting is genuinely hard. A team at General Atomics built an algorithm called ECHO — ECH Optimization — that handles this in real time. Here is the everyday version: imagine a concert hall with six motorized speaker clusters. You need the sound to hit specific seats precisely. ECHO is the system continuously adjusting each cluster's angle and volume, in milliseconds, even as the audience shifts. The technical trick is a neural network trained to mimic a complex physics simulation — called TORBEAM — that predicts where microwaves will land given the plasma's current state. The original simulation takes too long to run during an experiment. The neural network does the same job in about 75 milliseconds, fast enough for real-time use. A genetic algorithm — an optimizer that works like natural selection, testing thousands of candidate solutions quickly — then finds the best mirror angle and power level for each gyrotron. ECHO was tested in live DIII-D experiments. When one gyrotron broke mid-shot, the algorithm automatically redistributed the heating workload to the remaining devices. The target heating profile was maintained. The catch: ECHO's neural network was trained specifically on DIII-D's geometry. Applying it to another machine means retraining from scratch. The AI knows one kitchen very well. It doesn't know yours.

These three papers aren't about the same machine or the same problem — but they're circling the same underlying shift. The ARC simulations are mapping the operational envelope of a power plant Commonwealth Fusion plans to build in the 2030s: not whether fusion is possible in principle, but whether specific engineering choices hold up under full-system pressure. The Columbia hybrid design asks whether fusion machine architecture needs to be more flexible than today's fixed designs — a hedge against the possibility that neither tokamaks nor stellarators alone will prove ideal. And the DIII-D work shows that real-time AI control of live plasma heating is no longer a concept paper — it ran in an actual experiment and recovered from hardware failure on the fly. Together these point at something specific: the field is moving from 'can we achieve fusion?' toward 'how do we operate a machine reliably, for years, with all its imperfections?' That is a harder and more tractable problem at the same time.

Commonwealth Fusion Systems has indicated engineering milestones for their SPARC device are expected in the next year or two — today's ARC simulation results make those announcements more consequential to follow. On the control side, watch for whether General Atomics publishes more data from ECHO deployments across different plasma scenarios on DIII-D, since live AI-controlled experiments are still rare enough to be meaningful. The open question I'd want answered next: does argon seeding in ARC cause any stability problems on timescales longer than these simulations capture — the models are steady-state, and real plasma isn't.

• Wikipedia: Divertor — what it is and why it matters for fusion exhaust
• Wikipedia: Stellarator — the alternative to the tokamak
• Quanta Magazine: The inside story of the race to build a fusion power plant