Computing for Real-Time Plasma Diagnostics

Turning raw diagnostic signals into physical plasma quantities fast enough to inform control during a discharge.

The task

Diagnostics do not measure plasma quantities directly; they measure light, magnetic fields, particle fluxes, and interferometry signals from which quantities like temperature, density, and current profile must be inferred. Doing that inference fast enough to feed control, within the timescale of the plasma's evolution, is a real-time computing problem layered on top of a physics-inversion problem.

Common diagnostics and what they infer

• Magnetic probes: plasma current, position, and shape
• Interferometry: line-integrated electron density
• Thomson scattering: temperature and density profiles
• Spectroscopy: impurity content and ion temperature

Inversion is the hard part

Most diagnostics give integrals or projections, not local values. Reconstructing local profiles from them is an inverse problem: often ill-posed, sensitive to noise, and needing regularization or a physics prior to give a stable answer. Equilibrium reconstruction, inferring the plasma's magnetic configuration from external measurements, is the classic example and is computationally demanding to do quickly.

def reconstruct(measurements, forward, prior, reg):
    # minimize misfit + regularization (schematic)
    def cost(state):
        return misfit(forward(state), measurements) + reg*prior(state)
    return minimize(cost)

Speed versus fidelity

A full reconstruction may be too slow for real-time control, so a fast approximate version runs in the control loop while a high-fidelity version runs afterward for analysis. Choosing what the controller actually needs, often shape and stability margins rather than every profile detail, keeps the real-time computation tractable.

Kronos framing

For the Hyperion breeder these diagnostic-processing methods are developed and tested against simulation and the digital twin ahead of operation. The machine is not built; construction begins in the second quarter of 2027. Building the diagnostic-computing chain early means the control system is ready when the plasma is.