The Brain-Inspired Controller That Could Keep Fusion Reactors Running

Somewhere inside a future power plant, a centrifuge will spin at 70 revolutions per second, flinging frozen pellets of hydrogen isotopes into a plasma burning at 100 million degrees. Each pellet — a tiny ice crystal of deuterium and tritium — will ablate within one to two milliseconds, dumping a discrete burst of particles into the plasma before vanishing. The density of that plasma must be held inside a narrow window. Too sparse, and the fusion rate collapses. Too dense, and the plasma tears itself apart in an instability that can damage the reactor wall. The entire challenge of keeping a tokamak fusion reactor fuelled is, at its core, a problem of control theory — and it turns out to be a surprisingly hard one.

A new paper from researchers at Eindhoven University of Technology and DIFFER, the Dutch Institute for Fundamental Energy Research, proposes a novel solution (Jansen et al., 2026). Their insight is deceptively elegant: the physics of pellet fuelling looks almost exactly like the physics of how a neuron fires. Build the controller to match the physics, and the math works out cleanly. The result is the first provably stable, formally analyzed neuromorphic controller for plasma density regulation — along with new, rigorous guarantees for the existing industry method it aims to replace.

The Science

The core challenge is what engineers call a hybrid control problem: a system that is partly continuous and partly discrete. The plasma density evolves continuously between injections, governed by smooth differential equations that describe transport, wall interactions, and particle losses. But each pellet injection is a near-instantaneous event — a jump discontinuity in the density trajectory that ordinary control theory was not designed to handle. Most current tokamaks, including the Joint European Torus (JET) in the UK and the ASDEX Upgrade (AUG) in Germany, work around this by pretending pellets are a continuous flow and applying standard proportional-integral-derivative (PID) controllers on top. The discrete pellet firings are then handled separately by a technique called sigma-delta modulation (SDM) — a method borrowed from audio engineering and analog-to-digital conversion, where continuous signals are encoded into rapid sequences of discrete pulses.

SDM works well enough for today's machines. But the researchers argue it will be insufficient for next-generation reactors like ITER (the international experimental reactor under construction in France) and DEMO (the planned demonstration power plant). Those machines operate at higher magnetic fields where gas injection — the other way to add fuel — is blocked by edge transport barriers, making pellet injection essentially the only option. They also have stricter density tolerances and longer required operating times. For these reactors, approximating discrete pellets as a continuous flux is no longer good enough.

To build something better, Jansen et al. (2026) draw on hybrid systems theory — a formal mathematical framework that handles systems combining continuous flows and discrete jumps. They model the plasma density with a first-order differential equation during flow phases and a discrete jump equation at each pellet firing. They then embed two different controllers — their new neuromorphic (NM) design and a formalized version of SDM — inside this hybrid framework and prove stability theorems for both. The physical setup they analyze centers on a centrifuge-type actuator: a rotating disk that can only fire pellets when an ejection port aligns with the plasma entry point, meaning injections are restricted to multiples of the centrifuge period $T_c$.

What They Found

The neuromorphic controller is modeled directly on the leaky integrate-and-fire neuron, a century-old model of how biological neurons work (Lapicque, 1907). An internal accumulator variable $\xi$ — analogous to a neuron's membrane potential — integrates the density error $x=r-n_e$ (the gap between the target density $r$ and the actual electron density n_e$) over time. When $\xi crosses a threshold $\Delta$, the controller fires a pellet and resets $\xi$ to zero. The controller only integrates positive errors — it cannot remove particles from the plasma, only add them — which matches the physical one-sidedness of pellet injection perfectly.

The key difference between NM and SDM is in that reset. When SDM fires, it subtracts the threshold from the accumulator ($\xi^+ = \xi - \Delta$) rather than zeroing it. This seemingly minor change has significant consequences: during the transient phase when the plasma is far from its target density, the SDM accumulator can build up a large surplus — a phenomenon called integrator wind-up — which then causes the controller to keep firing pellets even after the plasma overshoots the target, producing an undershoot in density that persists long after the initial transient is resolved. The NM controller's hard reset to zero prevents this entirely.

The numerical simulations make this concrete. In Figure 3, the SDM controller applied to a plasma with target density $r=5\times 10^{19}$ particles/m³, pellet size $\alpha =1\times 10^{19}$ particles/m³, and confinement time $\tau =0.1$ s shows a pronounced undershoot during the transient — the density dips well below the target before recovering. The NM controller under identical conditions (Figure 2) converges more cleanly, without the undershoot.

The theoretical centerpiece of the paper is Theorem 5.1, which proves that the NM controller achieves practical stability — meaning the density error $x$ converges to and stays within the interval $(-\alpha ,\alpha ]$, i.e., within one pellet-worth of the target density. This requires two conditions to hold simultaneously. First, the centrifuge period must satisfy:

$$T_c\in \left(0,\tau \ln \frac{r}{r-\alpha }\right]$$

This constrains the minimum spin speed of the actuator given the pellet size, target density, and confinement time. Second, the neuron threshold must be chosen within an explicit allowable range derived from those same parameters.

Computed from Theorem 5.1, Jansen et al. (2026). τ = 0.1 s.

illustrates how the maximum allowable centrifuge period — equivalently, the minimum centrifuge speed — depends on the pellet-to-density ratio $\alpha /r$.

The theorem also yields a formula for the maximum plasma density any given system can sustain:

$$r_{\max }=\frac{e^{T_c/\tau }}{e^{T_c/\tau }-1}\cdot \alpha$$

This is immediately useful for reactor engineers. Given a pellet injector's speed and pellet size, $r_{\max }$ tells you the ceiling on achievable plasma density — before you ever run a single plasma experiment.

The researchers also identify a meaningful role for threshold tuning. At densities well below $r_{\max }$, tuning $\Delta$ shifts where within the stability band $(-\alpha ,\alpha ]$ the error tends to sit: a small threshold pushes the error toward the lower edge (density slightly above target), while a large threshold allows it to drift toward the upper edge (density slightly below target).

From Figure 2 numerical simulation, Jansen et al. (2026).

shows this effect numerically for the two extreme threshold choices validated in the simulations.

For SDM, the paper provides the first formal stability analysis ever published for this method in the fusion context. Two mitigation strategies for integrator wind-up are analyzed: input clipping (capping the signal fed into the integrator) and an adjusted jump map (modifying the reset behavior to prevent accumulation). The adjusted jump map recovers the same stability bounds as the NM controller

, at the cost of slightly more complex implementation.

Why This Changes Things

The significance here is not just that a new controller has been designed. It is that, for the first time, someone has written down the formal conditions under which pellet-based plasma density control is provably stable — and derived those conditions explicitly, in closed form, in terms of quantities an engineer can actually measure and adjust.

This matters enormously for the trajectory of fusion power. Current tokamaks operate with controllers tuned largely by experience and simulation. That approach works when you can iterate — run plasma shots, observe the result, adjust parameters, repeat. But ITER, which aims to produce 500 megawatts of fusion power from 50 megawatts of input heating, will operate in regimes where experiments are expensive and plasma instabilities are dangerous. DEMO, the planned step beyond ITER, must run continuously enough to feed power into a grid. In both cases, the cost of a density excursion — whether too high, risking a disruption, or too low, killing the fusion rate — is not just a failed experiment but potentially significant hardware damage.

Having explicit parameter constraints — precise conditions on centrifuge speed, pellet size, and threshold — transforms controller design from an art into an engineering calculation. The formula for $r_{\max }$ in particular is immediately actionable: it tells designers of ITER's and DEMO's pellet injection systems exactly what actuator specifications are needed to hit a given density target, before the machine is built.

There is also a broader intellectual payoff. The connection between neuromorphic computing and plasma control is not obvious, and finding that it works this cleanly is genuinely surprising. Integrate-and-fire neurons were first modeled by Louis Lapicque in 1907 — seventeen years before the first quantum mechanical description of the atom. The idea that a model of biological neural firing, repurposed a century later, might help control a fusion reactor is the kind of conceptual connection that makes physics worth paying attention to. It also suggests that the growing field of neuromorphic computing — hardware and software designed to mimic neural spike-based computation — has natural applications in any system where discrete events must be controlled against a continuous background: power grid switching, satellite thruster firing, drug delivery systems.

The computational lightness of the NM controller is not incidental. Real-time control in a fusion reactor operates on timescales of milliseconds. The controller must evaluate, decide, and signal the actuator faster than the plasma can change. Heavy optimization algorithms — like model predictive control (MPC), which has also been proposed for this problem — can in principle achieve better performance but demand significant compute at each timestep. The NM controller's logic is essentially a running sum compared against a threshold: it runs on embedded hardware with negligible latency, making it inherently real-time feasible.

What's Next

The paper is candid about its limitations. The plasma model used — a single first-order differential equation with one confinement time constant $\tau$ — is deliberately simplified. Real plasmas are spatially distributed, have multiple particle species, and interact with magnetic field geometry in ways that a scalar density equation cannot capture. The authors note that system identification experiments on the MAST-U tokamak in the UK do support a first-order-plus-delay model for the core density response in the relevant frequency range (7–40 Hz), validating the approach for controller design purposes, but a fuller model would be needed before deployment.

The analysis also assumes no measurement noise and no model uncertainty — idealizations that will need to be relaxed. Extending the stability proofs to include actuator delays (pellet preparation time between shots), sensor noise, and parameter uncertainty are natural next steps identified by the authors. The gas gun actuator case, where pellets can in principle be fired at any time subject to a minimum inter-shot delay rather than only at periodic centrifuge slots, is briefly discussed but not fully analyzed.

Perhaps most importantly, the framework needs to be validated against real plasma data. The numerical simulations in this paper use parameters drawn from existing AUG experiments — the 70 Hz centrifuge, $\alpha =1\times 10^{19}$ particles/m³ pellets, $\tau =0.1$ s confinement time — and the theoretical predictions match the simulations precisely. But a plasma shot on a real machine involves magnetohydrodynamic instabilities, impurity transport, and edge effects that no first-order model fully captures.

Still, the paper represents a meaningful step toward what the fusion community needs: control systems with formal guarantees, not just empirical ones. As ITER approaches its first plasma operations and DEMO moves from concept to design, the engineering questions around plasma fuelling will only intensify. A controller grounded in rigorous mathematics — one that tells you in advance whether it will work, under what conditions, and to what precision — is not a luxury. It is a prerequisite for a machine that is supposed to produce clean energy for a grid.

The brain, it turns out, may have had a head start on the problem.