The Plasma Diagnostic Suite and Real-Time State Estimation for a Compact Spherical-Tokamak Breeder: Magnetics, Thomson, Interferometry, and Bolometry Fused by a Graph-Neural-Network Reconstructor

Abstract

An autonomous controller can act only on the state it can see, and in a compact spherical tokamak the state must be produced faster than the plant evolves, from a diagnostic set that survives a 14 neutron field and keeps working when individual channels drop out. We formalise the Hyperion breeder diagnostic suite — magnetics, Thomson scattering, interferometry, bolometry, and a neutron camera — and its real-time state estimator as a Bayesian inverse problem, write the forward observation model of each diagnostic and the maximum-a-posteriori estimator with its posterior covariance, and cast observability of the internal state as the rank and conditioning of the associated Fisher information. The estimator's learned core is reduced to practice at the metrics that gate it into service. A graph-neural-network reconstructor, which we identify as an amortised posterior-mean estimator over the sensor graph, recovers the plasma state and a quench precursor jointly at an area under the receiver-operating-characteristic curve (AUC) of 0.980 at a reconstruction normalised RMS error of 0.396, and degrades gracefully rather than catastrophically as channels are removed. A multi-modal fusion of a fibre-Bragg strain-rate channel with a magnetization channel reaches 0.992 (against 0.976 and 0.921 alone) and a true-positive rate of 0.857 at a 1% false-positive rate, the corner where a protection trip is decided; that separation buys ≈47 of quench warning against 2.6 for the raw magnetization signal — an 18× lead time that lets the controller act on the trip rather than merely record it. A recursive Kalman/particle filter fuses the estimate across time as well as across modality, driving the state error to the full-suite floor within a few tens of milliseconds and supplying a shadow state that leads the plant. Internal state — ambipolar potential, T_e, n_e and β — is reconstructable from as few as six diagnostics at a mean coefficient of determination R^2≈0.86, confirming observability for the control layer, and a D-optimal sensor-placement criterion selects where to add or move channels to raise it. The suite adds an electron-cyclotron-emission radiometer and a fast-ion tracker on the same backbone, and the estimator is guarded by an epistemic gate: when its posterior variance or an out-of-distribution score crosses threshold, authority is handed back to a deterministic floor, with a three-detector majority vote gating every precursor trip. An in-winding nitrogen-vacancy (NV) diamond magnetometer resolves a conductor-scale quench step at a shot-noise-limited sensitivity of 0.189 pT Hz^-1/2 at 125 bandwidth, for a per-shot signal-to-noise ratio of order 10^9. The estimator holds the plasma state at the breeder design point (I_p=9.66, Q=3.076, P_ fus=85.04, δ=-0.30, centrepost peak field 16.84), which is a distinct magnet from the burner plug coil (26.49); the physical-inversion analyzers are built to the demonstrated estimator pattern on a staged, acceptance-gated schedule. Against published quench-detection, quantum-sensing and equilibrium-reconstruction benchmarks the figures are competitive in a deliberately hard, weak-signal regime.Key results (frozen anchors): L_W = 5.7 ×10^-32; Ar = 1.45 ×10^-33; fracg_B = 2 ×2.Methods & codes: FreeGSNKE, FreeGS, OpenMC, ENDF/B-VIII.0, Sauter.Live verification: 7 gate validator(s) with live-recompute cards (7 reproduced). See the Verification section and data/verification.csv.Related: Paper page · De-risking register · 3D model · Learn more about KronosPart of the 2026 Kronos publication series; independently re-run and stamped in the Kronos de-risking register (DOI 10.5281/zenodo.22645689).All numerical values are frozen design-point anchors; see the register.Public research artifact. No proprietary, financial, or supply-chain information is included.