Sub-Millisecond Neural Equilibrium Reconstruction Robust to Sensor Dropout: A Virtual-Sensor Mesh for Radiation-Degraded Diagnostics in a Compact Spherical-Tokamak Breeder
Year: 2026
Type: dataset
Abstract: Every control decision in a compact spherical tokamak begins with the equilibrium, and in a nuclear device that equilibrium must be reconstructed faster than the plant, verified against the governing physics, and kept trustworthy when a diagnostic degrades or drops out under a neutron field. We formalise and demonstrate a real-time reconstruction stack for the Hyperion breeder (design point I_p=9.66, Q=3.076, P_ fus=85.04, negative triangularity δ=-0.30, B_0=8, elongation κ=2.0, major radius =1.20, aspect ratio A=2.5, centrepost peak field 16.84). A physics-informed neural surrogate reconstructs the poloidal flux (R,Z) to a root-mean- square error of 0.139% of range against a free-boundary Grad–Shafranov reference — inside the 1% acceptance target and improving a prior 1.79% attempt — at a 2877× speedup (0.87 inference versus a ~SI2.5 Picard solve), with a discretised Grad–Shafranov operator residual of 0.32% on the interior confirming the reconstruction is a solution of the governing equation and not a free-form fit. A graph-neural virtual-sensor mesh assimilates the diagnostic set into a plasma-state estimate and, in the same forward pass, discriminates disruption and quench precursors at an area under the receiver-operating characteristic (ROC) of 0.980, above any single channel; a quadrature fusion of an uncorrelated strain-rate and magnetization pair reaches 0.992 and a true-positive rate of 0.852 at a 1% false-positive rate, independently reproduced under a separate implementation to five significant figures; the same fusion converts that discrimination into operational lead time, buying 47.2 ms of quench warning against 2.6 ms for terminal-voltage detection — a ~mathbf18× margin that is enough to prevent the quench rather than record it. The mesh is architected for graceful degradation: an in-winding nitrogen-vacancy (NV) diamond magnetometer at a shot-noise-limited 0.189 pT Hz^-1/2 (125 bandwidth) supplies a local vector channel that does not share a failure mode with the plasma-facing optics, and a model-fusion backup holds precursor AUC 0.95 with the NV channel degraded by up to 39×, so the protective function survives radiation-induced sensor loss. Every reconstruction additionally carries a distribution-free split-conformal predictive interval with a finite-sample marginal-coverage guarantee (1-α), a deep-ensemble epistemic variance, and a Mahalanobis out-of-distribution score calibrated against a 40,000-sample joint-uncertainty envelope; when the plant state leaves that validated envelope these gates hand authority back to a deterministic free-boundary baseline in a Simplex-style fallback rather than acting on a confident extrapolation. Each estimator is admitted to control authority only after clearing an explicit acceptance inequality. The result is a fast, physics-consistent, dropout-tolerant, uncertainty-aware reconstruction with a demonstrated core and a defined, testable path to full diagnostic coverage. Key results (frozen anchors): k = 0 ,K-1,; fracg_B = 2 ×2; alpha = 90 %. Methods & codes: FreeGSNKE, FreeGS, CGYRO, OpenMC, NIMROD, GENE, DAGMC, Sauter. Live verification: 5 gate validator(s) with live-recompute cards (5 reproduced). See the Verification section and data/verification.csv. Related: Paper page · De-risking register · 3D model · Learn more about Kronos Part 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. Part of the Kronos Fusion Energy 2026 Publication Series — 115 papers on a compact spherical-tokamak breeder (Hyperion) and D–3He tandem-mirror burners (Aegis / MetroVolt). Every numerical value is a frozen design-point anchor, independently re-run and stamped in the register. Zenodo record: 10.5281/zenodo.22645801 De-Risking Register (mothership): 10.5281/zenodo.22645689 Read online: https://www.kronosfusionenergy.com/publications/sub-millisecond-neural-equilibrium-reconstruction-robus.html Live validators: Physics Validation & Simulation © 2026 Kronos Fusion Energy — a family-owned American company. Licensed CC BY 4.0.
Source: Harvard Dataverse
DOI: 10.7910/dvn/lk70by
Authors: Ford, P I, Kulcinski, G L
Language: English
Subfield: Nuclear and High Energy Physics
Field: Physics and Astronomy
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7212062797